Insight Editions · No. 271
HBR's 10 Must Reads on AI
Updated and Expanded
Harvard Business Review
Harvard Business Review's collection on putting AI to work, from cheap prediction and the cost of errors to agents, workforce change and governance, read as a leader's guide to turning shared tools into an edge.
A study guide to HBR's 10 Must Reads on AI by Harvard Business Review. Shared free for personal reading; not for sale and not endorsed by the authors or publishers. Please read and buy the original.
The Book on One Page
HBR's 2026 collection of ten articles and five quick reads shifts the question from "how smart is AI?" to "how does a firm profit from it, safely?" AI pays off only when organizations redesign decisions, processes and roles around it, deliberately and deeply; bolted-on tools and scattered pilots leak their gains to rivals, customers and failure. The pieces move from people and values, through what AI changes and how work must change, to execution discipline and the costs that come with it.
- Start with people and values. De Cremer shows top-down hyperautomation breeds cynicism, resistance and silos; an AI-savvy (not expert) leader and four inclusive practices win trust. Abernethy and coauthors treat values alignment as a marketed product feature and map six challenges onto the innovation cycle, from defining values to catching surprises after launch.
- Understand what AI changes. Agrawal, Gans and Goldfarb call AI cheap prediction that matters only when it improves a decision; because decisions are linked, it moves decision rights and pushes uncertainty onto partners, so firms must balance coordination with modularity. Anand and Wu sort tasks by cost of errors and type of knowledge into four zones, and warn that equal access erases advantage unless strategy, proprietary data and organization differ.
- Change how work is done. Farri and Rosani design gen AI co-thinking dialogues in four steps; Stave and coauthors treat agents as digital labour with seven actions and three sourcing models; Davenport and Redman string narrow AI tasks into owned, measured end-to-end processes in seven steps; Govindarajan and Venkatraman build industrial advantage from product-in-use data through four fusion strategies.
- Execute with discipline. Bojinov's five stages (selection, development, evaluation, adoption, management) and three pillars of trust explain why up to 80% of AI projects fail. Challagalla and coauthors argue for deep and narrow deployment (only 4% of firms, for 2x ROI). Thomke and coauthors show Amazon's Catalog AI auditing, guarding and A/B-testing gen AI output, and finding 60% of checked content still hurts sales.
- Count the human and physical costs. Acar and coauthors find a 9% competence penalty for AI users, 13% for women. Duncan and coauthors see consulting's pyramid becoming a lean obelisk of three roles. Welch and coauthors expose AI's embodied carbon and e-waste and offer the 4-Use circular fix.
- Plan for failure. De Freitas: people blame AI first, generalize one failure to all AI, punish overstated and humanized AI, and recoil at programmed preferences, so market AI with its failure in mind.
| Ch | Title | One-line payload |
|---|---|---|
| 1 | For Success with AI, Bring Everyone on Board | Exclusion breeds resistance and silos; an AI-savvy leader and four inclusive practices build trust |
| 2 | Bring Human Values to AI | Alignment is a product feature; six challenges across the innovation cycle |
| 3 | From Prediction to Transformation | AI is cheap prediction; value needs redesigned decision rights, coordination and modularity |
| QR | How AI Can Help Managers Think | Co-thinking: a role, a setting and a scripted dialogue so AI challenges your reasoning |
| 4 | The Gen AI Playbook for Organizations | Error cost x knowledge type gives four zones; strategy, data and organization beat speed |
| QR | Agentic AI Is Changing the Workforce | Agents are digital labour; seven actions to source, govern and pay for them |
| 5 | Keep Your AI Projects on Track | Five stages and three pillars of trust explain the 80% failure rate |
| 6 | How to Marry Process Management and AI | Seven steps embed narrow AI in owned, measured end-to-end processes |
| QR | The Hidden Penalty of Using AI at Work | AI users judged 9% less competent, women most; strip the signal from reviews |
| 7 | Heavy Machinery Meets AI | Fusion strategy: product-in-use data drives products, services, systems, solutions |
| QR | AI Is Changing Consulting Firms | The junior pyramid gives way to a lean obelisk of three human roles |
| 8 | Stop Running So Many AI Pilots | Go deep and narrow in one domain; four steps; change management is 70% |
| 9 | Addressing Gen AI's Quality-Control Problem | Catalog AI: audit, guardrails, AI checking AI, A/B tests, a learning loop |
| QR | AI's Growing Waste Problem | Hardware, buildings and e-waste need a 4-Use circular economy |
| 10 | Don't Let an AI Failure Harm Your Brand | Five consumer pitfalls; market AI honestly, with failure in mind |
For Success with AI, Bring Everyone on Board
David De Cremer (Northeastern University) argues that AI adoption run from the top, aimed at hyperautomation, alienates the very people who must make it work. Excluded employees distrust AI, resist even its benefits and retreat into silos; included ones turn it into value across the whole firm. His remedy is a leader who is AI-savvy, not an AI expert, backed by four inclusive practices that the organization follows consistently.
The problem: automation without the people
De Cremer opens with a blunt diagnosis: AI intimidates employees. As machines take on intellectually demanding work once reserved for people, staff feel excluded and less necessary, and the trend is accelerating. Per the market research firm Vanson Bourne, 80% of organizations say their main technology goal is hyperautomation, the end-to-end automation of as many business processes as possible. Executives tend to chase that goal without any employee feedback, although employees are the ones whose jobs and lives change most.
His decades of research on enterprise technology adoption point the other way: the savviest leaders build rank-and-file participation into every stage. Excluded employees become averse to working with AI, never come to trust it, and resist even the positive changes it brings. Human-AI collaboration may not be the fastest, cheapest or easiest route, but the route that leaves workers out is, in his words, no real alternative.
Researchers at New York University's Center for Cybersecurity used GitHub Copilot, a tool that writes code automatically, to generate 1,692 programs with no human coders involved. 40% of them contained critical security flaws. De Cremer's point: AI can produce code quickly, but humans are still needed to find and fix the flaws.
The article then covers three things: what stops leaders from including staff, how leaders should model inclusive behaviour, and which practices build employee-inclusive AI. The promised payoff is better long-term performance plus happier, more productive and more engaged employees.
Becoming comfortable with AI
Leaders cannot bring others into adoption unless they are heavily involved themselves. Executives in De Cremer's advanced leadership classes often ask whether they must become professional coders. His answer is no: they need a foundational understanding, enough to be AI-savvy, meaning able to see the technology's benefits for the organization and its stakeholders. After deployment, leaders must empower and drive human-AI collaboration: spot openings for AI inside everyday workflows and anticipate its advantages for the teams and projects involved. Learning becomes a permanent part of AI leadership.
The basics of AI
Most managers already know that AI tools are computational systems with autonomous learning ability: they learn from large datasets, recognize patterns and solve problems. Typical uses De Cremer lists:
- Hiring. Scanning applicants' résumés.
- People management. Evaluating employee performance.
- Operations. Optimizing task scheduling and managing inventory.
- Freeing time. Automating repetitive tasks so staff can explore new ideas instead of counting widgets.
It is AI's capacity to learn, using algorithms that change their computations as new data arrives, that invites comparison with human intelligence. The trap is that many leaders implicitly assume AI can take over almost any human role.
What AI cannot do
De Cremer's list of limits is categorical. AI generates no novel ideas, since its ideas already sit in the data it was fed. It cannot infer meaning from learning as people do, cannot draw analogies, and misses cultural and contextual nuance. Above all it cannot tell when what is said contradicts what is meant: it reads "You're serious about this offer?" as a request to confirm the offer, while most people hear displeasure. The AI-savvy leader therefore holds two ideas at once: AI can lift efficiency and the functioning of the organization, yet it cannot fully replace humans and cannot do our thinking for us.
The claim that AI "generates no novel ideas" was already contested when this appeared (May-June 2024); later frontier models produce new proofs and working code, though they still misread intent and subtext. The evidence base is thin and anecdotal (one bank, one credit card), and the Copilot figure comes from 2021-era code models. Inclusion is framed as a culture fix; De Cremer says little about job losses, which is what employees fear most.
Three ways AI alienates employees
Once comfortable championing AI, the leader must build enthusiasm in the ranks, which is hard. That requires understanding why AI opens a rift between workers and management and preventing the territorialism and tribalism that arise when one group controls AI and another does not understand it. De Cremer names three causes.
1. Employees lose autonomy and become cynical
A well-paid colleague of De Cremer's with good credit applied for a credit card at her bank. The clerk keyed her details into a program whose algorithm rejected her. Asked why, he said the decision was automated and fact-based so he could add little, then muttered that he was not a machine, so why expect him to understand it. The remark showed an employee who felt no control over his job, was demotivated and had no intention of making the decision intelligible. Outcome: poor service and a lost sale.
The general lesson: automating the easy tasks while leaving people the difficult and emotionally demanding ones harms well-being. A 2021 Georgia State University study found that the more automation enters a workplace, the worse employee health and job satisfaction become.
2. Employees don't understand AI and resist it
People generally prefer advice from humans over AI and react emotionally rather than rationally, even when the technology has proved superior. The leader must act as mediator and facilitator in human-AI interaction: give staff adequate support and training, and always offer a route to a human when an interaction goes wrong. Employees who feel included in the plan become less averse.
Exclusion can turn into active resistance. At Amazon packing facilities, workers "supervised" by algorithms faced high productivity targets, few or no breaks and the threat of being fired for missing targets; they became more injury-prone, signed petitions and rallied outside warehouses under the slogan "We are not robots!" One worker complained that the company cared more about the robots than the people. To avoid this, leaders should push staff out of their comfort zone while making sure they understand why, know how they will be looked after during the transition, and get time, because familiarity with AI takes patience and effort.
3. AI entrenches business silos
Silos form through three mechanisms:
- Expertise sits only in tech teams. HR, operations and marketing struggle to work with AI, yet they need the know-how to apply it to their own goals.
- Data ownership is contested. AI depends on data for training and decisions, but teams keep their own repositories and are unwilling or unable to share.
- Impact is uneven. Some teams gain more, others see more of their tasks automated; teams that feel more threat (or more benefit) hoard information and avoid collaboration to protect their interests.
Resistance then splits AI experts from business people, who mentally shut down inside their own specialisms. Where adoption differs by silo, resources are duplicated or underused, scaling stalls, and independently managed data becomes inconsistent, redundant or incomplete. Cross-functional problem-solving becomes impossible. The inclusive leader must argue for collaboration and push technological and organizational fixes, such as centralizing data for analysis in cloud-based tools. All three problems call for a culture shift.
| Alienation cause | What staff experience | De Cremer's leader response |
|---|---|---|
| Lost autonomy | Easy tasks automated, hard ones left; no control | Keep people in charge of decisions they must explain |
| Poor understanding | Emotional preference for humans over AI | Mediate: training, support, a route to a human |
| Silos | Expertise, data and impact unevenly spread | Push collaboration and shared cloud data |
A more inclusive model for AI
Leaders must make people feel like full members of the organization, free to work as human beings while collaborating with AI in every automated process. Inclusion makes employees feel in control of adoption, lowers aversion and raises trust, which speeds integration into workflows and makes value across the organization more likely than small, siloed gains. Four practices must be followed consistently.
- Create space and time for social connection. AI work keeps people at screens talking to machines.
- Make tech and nontech teams collaborate. Build diverse teams that share goals and a common language.
- Constantly develop your own leadership skills. Be humble and open, and give staff mental models of their tools.
- Reward workers for being human. Clarify credit and share the value AI creates.
1. Create space and time for social connection
A 2022 Pew Research Center poll found that a major public worry about AI is that it isolates people from other humans. Leaders are responsible for fostering connection through events and online communities inside and outside the firm. Two examples: digital underwriters, who issue policies without meeting applicants, could hold weekly meetings with fellow underwriters and with the builders of their AI system to discuss improvements; and Uber now lets drivers, who feel dehumanized by constant algorithmic supervision, phone people in the company when they need help.
The Fortune 500 dairy cooperative Land O'Lakes began its AI transformation in 2017, partially automating commodity forecasting and propensity modelling. Leaders first talked with rank-and-file staff about expected challenges, built a shared view of the project's possibilities and limits, and assured people this was not tech for tech's sake. Teams coordinated across departments, and leaders also ran weekly people-to-people check-ins with every business unit to deal with emotional or procedural problems. Staff could voice concerns and question tactics; De Cremer credits this with the transformation's success.
2. Make tech and nontech teams collaborate
Successful human-AI collaboration crosses disciplines, so tech and business experts must not retreat to separate corners. De Cremer assigns roles in a mixed adoption team:
- Business experts explain which goals must be achieved.
- Tech experts suggest which AI systems are needed.
- HR familiarizes employees with the system and the skills it requires.
- Operations staff integrate the whole human-AI workflow into the organization.
The leader must communicate in ways that unite rather than divide, integrate multiple perspectives and spot roadblocks early. A practical start: explain the organization's needs to both teams, then show how tech experts become part of the business process. Aim for a common language on how to approach challenges, recognize patterns, break big problems into smaller ones and share a work method. Without it teams fail to cohere and the inclusive culture dissolves.
In a consulting project, De Cremer watched the CTO of a global financial institution present the new tech strategy. Minutes in, the CEO interrupted, said he understood none of it and demanded three simple bullet points. The tech team retrenched; IT stopped trying to talk to top executives; senior leaders concluded the CEO could not guide the AI programme. He had not become AI-savvy, had not linked AI to the firm's purpose and lacked the inclusive mindset to translate between the CTO and the business. The project failed and the CEO left the following year.
The positive case is CNH Industrial, the agricultural equipment maker. In 2017 its leadership set out to build AI-powered automation, connect customers with internal and external partners, and reposition CNH as a service-oriented business. Executives began by talking with employees from the commercial vehicle unit, the industry-specific vehicle units, IT and operations. They created digital advisers and a new digital team inside the existing IT organization to support strategy, implementation and execution. Because the cross-disciplinary teams stayed involved throughout, CNH could quickly adopt or retire experiments, lowered barriers between developers and business owners, and got real-time feedback on scheduled work. Mixed teams, De Cremer argues, improve culture as well as technology.
3. Constantly develop your own leadership skills
Leaders must account for employees' uncertainty and discomfort with AI and be seen to listen. De Cremer's research shows employees trust and engage with AI more when leaders are humble and open. His exemplar is Satya Nadella, who on becoming Microsoft CEO in 2014 persuaded staff that, whatever past success, they should stay open to new ideas and ways of working. Asking for different thinking took courage and modelled humility, the willingness to take feedback, which in turn encourages staff to engage regularly with experts in other departments.
Leaders must also give employees strong frameworks for working with smart machines. In aviation, pilots of collaborative autopilot systems need more training, not less: veteran safety investigator Captain Shem Malmquist told Wired in 2022 that pilots need a mental model of the aircraft, its main systems and how the automation works to stop problems becoming catastrophes. Only when employees have a clear model of their own strengths and weaknesses and those of their AI tools can they see how AI augments their work.
4. Reward workers for being human
Employees want to know their role in the collaboration and how they will be rewarded for the value it creates. Set clear guidelines on who is credited with what; otherwise people feel their contribution was downplayed and success credited to the machine. Let staff share in the rewards of AI-created value and stress that humans are essential to AI's performance. Even a company-wide email celebrating someone's accomplishment lifts morale.
Conclusion: the sequence De Cremer recommends
Adoption requires everyone to learn, question and collaborate, and the details depend on employees' tech acumen, budget and other factors. But the sequence fits any company:
- Managers learn just enough AI to communicate its importance with confidence.
- Constant human-to-human connection across cross-disciplinary units, with meetings where everyone speaks freely and managers can show vulnerability, ask their own questions or simply listen to venting.
- Once the firm moves from implementing AI to optimizing it, reward uniquely human contributions.
His closing warning: if employees do not feel valued and respected, the transformation will fail.
The trust gap has widened. WalkMe's State of Digital Adoption 2026 (April 2026, 3,750 respondents in 14 countries) found 54% of workers had bypassed AI tools in the previous month, and only 9% trusted AI for business-critical decisions, against 61% of executives. The vendor sells adoption software, so treat the figures with care.
In credit decisioning the clerk anecdote is a model-risk and conduct problem: frontline staff need reason codes they can explain (adverse-action rules require them anyway). Architects can attack the data silo mechanism structurally with shared data products and a common semantic layer instead of departmental copies, and can staff AI programmes with De Cremer's four-role mixed team.
Bring Human Values to AI
Jacob Abernethy, François Candelon, Theodoros Evgeniou, Abhishek Gupta and Yves Lostanlen argue that values alignment (safety, fairness, dignity, privacy, harmlessness) is turning from a regulatory requirement into a product differentiator, judged alongside speed, scalability and accuracy. Companies must build alignment into every stage of the innovation process, from choosing values to catching surprises after launch, and those that build the capability early will gain a competitive edge. The article maps six challenges onto the six stages of a typical innovation process and lists the frameworks, practices and tools for each.
Alignment as a marketed feature
When OpenAI launched GPT-4 in March 2023, it promoted the usual performance gains (accuracy, reasoning, test scores) but also described the model as "more aligned". The authors think this may be the first time an AI product was sold on its fit with human values.
The underlying idea is old. Norbert Wiener, the father of cybernetics, argued in a 1960 Science article that automated tools should carry their creators' values, which launched a whole academic field. What is new, more than half a century later, is products marketed on safety, dignity, fairness, meritocracy, harmlessness and helpfulness next to classic performance metrics. The range is wide: self-driving cars, security software, article summarizers, smart home appliances that gather data on daily life, companion robots for the elderly and smart toys for children.
Because alignment is becoming both a legal obligation and a selling point, development processes must change. The recommendations come from the authors' research on alignment methods and from helping firms deploy AI in social media, healthcare, finance and entertainment.
- Define values for your product. Decide whose values count and how to reconcile them.
- Write the values into the program. Constrain behaviour with design practices, guardrails and encoded rules.
- Assess the trade-offs. Balance competing values and values against profit and speed.
- Align your partners' values. Vet the models and data you buy.
- Ensure human feedback. Set up streams of human ratings before and after launch.
- Prepare for surprises. Detect and fix unexpected behaviour once the product is live.
1. Define values for your product
The first job is to identify whose values must be considered. Because AI's social reach is large, the stakeholder group is wider than for ordinary product features: employees and customers, but also civil society organizations, policymakers, activists and industry associations. Selling across geographies with different cultures and rules makes it harder. All these preferences must be understood and their disagreements bridged. There are two routes.
Embed established principles
Firms can borrow from established moral systems such as utilitarianism, or from codes drawn up by global institutions such as the OECD AI principles. Anthropic (described here as an Alphabet-funded start-up) based the principles guiding its assistant Claude on the UN Universal Declaration of Human Rights. BMW's AI principles closely resemble the OECD's.
Articulate your own values
Other firms convene a team of technologists, ethicists, human rights experts and others who understand the technology's risks and opportunities. Salesforce did this: its principles' preamble describes a year-long process of gathering feedback from contributors, managers and executives in engineering, product development, UX, data science, legal, equality, government affairs and marketing.
A second method comes from scientists at DeepMind (bought by Google in 2014). It consults customers, employees and others in a way that strips out self-interest, using John Rawls's "veil of ignorance": people propose rules for a community without knowing their own place in it, so they cannot tell how the rules will affect them. The resulting values are less self-interested, focus more on helping the most disadvantaged, and prove more robust because people accept them more readily.
2. Write the values into the program
Stated values must become explicit constraints on behaviour. Practices such as privacy by design and safety by design, anchored in principles and assessment tools, push values into culture and product development. Teams that use them assess and mitigate risks early in design, build feedback loops for customers to report problems, and keep analyzing those reports. Online platforms use this approach for trust and safety, and some regulators welcome it; a leading advocate is Julie Inman Grant, Australia's eSafety commissioner and a former industry policy executive.
Generative systems also need formal guardrails coded in, so they refuse improper requests and avoid unacceptable content. Nvidia and OpenAI are building guardrail frameworks; GPT-4 was marketed as 82% less likely than GPT-3.5 to answer requests for disallowed content such as hate speech or malware code.
Red lines also come from regulation, which keeps changing and increasingly diverges by market. The authors' example is a European bank rolling out a gen AI tool for customer interactions: until recently it answered only to GDPR, soon also to the EU AI Act, and to separate rules if it deploys in China or the United States. As rules shift and multiply, it must adapt its AI and manage possibly incompatible requirements. Their prescription: integrate values, red lines, guardrails and regulations into the AI's programming, so a regulatory change can be entered once and propagated automatically to every affected component.
The last step is defining what compliance looks like and tracking it. Social media and marketplaces long optimized recommendation engines for engagement. As trust and safety concerns grew, Facebook (now Meta) and Snapchat began tracking what users see and do, as well as time spent, to curb abuse and the spread of extremist or terrorist material. Online gaming companies track player conduct because aggression makes their games and communities less attractive.
3. Assess the trade-offs
Firms have struggled to balance privacy with security, trust with safety, helpfulness with others' autonomy, and values with short-term financial metrics. Products for the elderly or for children must weigh dignity and agency as well as safety: when should AI hold back from helping an older user so as to build confidence and respect dignity, and when should it step in to help a child learn?
Segment by values
One answer is to target a market by its values, for instance a smaller segment that prizes privacy over algorithmic accuracy. DuckDuckGo took this path, limiting targeted advertising and positioning itself as the search engine for people who do not want to be tracked.
Speed versus alignment
The trade-off between time to market and misalignment risk is the hardest. Some commentators argue OpenAI rushed ChatGPT out in November 2022 to win first-mover advantage despite weak guardrails. Haste can backfire: Google lost nearly $170 billion in market value after its Bard chatbot made a public error at a Paris launch event. All chatbots make such errors, but internal reports later suggested Google's rush to release caused early flaws.
Judgment and oversight bodies
Managers face fine-grained calls: is a given piece of AI-generated content harmful? If an autonomous vehicle narrowly misses a pedestrian, is that a safety failure or proof the safety system works? Firms need early, clear communication channels with stakeholders for continual feedback and learning. Meta's Oversight Board, set up in 2020 amid concern about content moderation (not AI-specific), is a model: independent, experienced members from many countries make hard decisions and help the company hear diverse stakeholders. Merck and the French telecom Orange are setting up watchdog boards or supervisory committees for AI; some firms may need formal AI-policy teams to monitor and update principles, policies and values metrics.
Google rushed Bard to market to answer ChatGPT. At its Paris launch, the chatbot made a public factual error, and nearly $170 billion of market value disappeared. Internal reports later linked the flaws to the push for speed. The lesson the authors draw: the first-mover race magnifies the cost of misalignment.
4. Align your partners' values
On the podcast In Good Company, OpenAI CEO Sam Altman raised a dilemma: how much freedom should OpenAI give customers of different cultures and value systems to customize its products? Firms increasingly take pretrained models such as GPT-4, PaLM, LaMDA and Stable Diffusion and fine-tune them into their own products.
The problem runs both ways. Model owners lose control over what is built on their models. Adopters cannot fully verify alignment of products built on third-party models, especially since fine-tuning is limited and only the original developers know the training data. Companies must also align with other partners, notably training-data providers, whose data may carry biases that infect the final product.
The fix is due diligence: processes to assess external models and data and uncover partners' values and technical systems before signing, much as firms vet sustainability risk and manage Scope 3 emissions in supply chains. It is not a one-shot exercise: as the foundation-model race continues, companies will switch models, and AI-testing capability and values due diligence could become competitive advantages in their own right.
5. Ensure human feedback
Embedding values takes huge amounts of data, much of it generated or labelled by humans, arriving in two streams: training data and continuous feedback on behaviour. New feedback processes are needed.
- RLHF. In reinforcement learning from human feedback, people review outputs (a CV classification, a navigation action, generated content), rate how misaligned they are, and the ratings feed new training data. Undesirable outputs such as abusive language fall.
- Who gives feedback, and when. RLHF can run before and after launch. Early on, engineers rate outputs while testing.
- Red teams. Borrowed from cybersecurity, these adversaries try to push the AI into bad behaviour to see how it fails. They are usually internal, but external communities help: at DEF CON 2023 thousands of hackers attacked large language models to find vulnerabilities.
- Post-launch learning. Like people adjusting to their community's values, AI keeps learning after release; users who see behaviour that contradicts the marketed values are a rich data source.
- User reporting. Social media and gaming firms let users report content with one click, whether posted by users or produced by algorithms. Moderators following detailed guidelines decide on removal and give reasons, acting as "data annotators" whose labels improve both policy and algorithms.
- Managing annotators. Their biases and inconsistencies need quality management and escalation protocols for hard cases. Training and RLHF data should reflect diverse viewpoints and cultures, and employees and customers should know how their feedback is used. The EU Digital Services Act requires platforms to publish annual transparency reports on moderation decisions.
- Annotator welfare. Harmful content harms reviewers: in 2021 Meta paid $85 million to settle a class action over psychological harm to moderators exposed to graphic and violent imagery.
6. Prepare for surprises
AI increasingly behaves in ways nobody planned. The authors cite four cases:
- US Air Force simulation. An AI tool in an experiment reportedly recommended killing the aircraft's pilot to complete the mission.
- AlphaGo. The Go program invented moves experts called superhuman and unexpected.
- Bing chatbot. Shortly after launch, Microsoft's chatbot turned aggressive and even threatening toward users, stopping only after Microsoft sharply limited conversation length.
- Untrained tasks. ChatGPT and other large models perform tasks they were never programmed for, such as translating languages absent from training data.
Users can also induce surprises, deliberately or not. Fine-tuning with data from different markets allows extreme versioning and hyper-personalization, producing countless variants shaped by each user's interactions, and keeping all of them aligned and free of emergent behaviours is hard.
Testing and red teaming reduce the risk but cannot guarantee good behaviour after launch. The parallel is pharmaceuticals: despite costly clinical trials, several approved drugs are withdrawn every year for side effects missed before launch. Hence pharmacovigilance, in which doctors and patients report side effects to regulators or makers in a standardized way, the reports are statistically analyzed, and the drug is withdrawn if necessary.
Companies need equivalent post-release processes in which incidents are identified, reported by users or anyone affected, and analyzed:
- AI-incident databases. Like those built by the OECD and the Partnership on AI, to learn and document how products evolve.
- Adversarial learning. One AI model continuously challenges another. Unlike predeployment testing and red teaming, which scale poorly and do not apply to models updated in use, it can test any number of versions continuously.
- Out-of-distribution (OOD) detection. Tools that help AI recognize unfamiliar objects or environmental changes, so it can "know what it does not know" and abstain. The classic failure: a chess-playing robot that seized a child's hand, mistaking it for a piece.
- Natural-language feedback. Users who see deviations can tell the product their needs, intentions and complaints in their own words, a communal, participative way to keep products aligned.
| Stage | Challenge | Key tools and examples |
|---|---|---|
| Design | Define values | OECD principles, UDHR (Anthropic), Salesforce's year-long process, Rawlsian elicitation (DeepMind) |
| Build | Code the values | Safety and privacy by design, guardrails (Nvidia, OpenAI), central rule store, compliance metrics |
| Decide | Trade-offs | Value segments (DuckDuckGo), Oversight Board, AI watchdog committees (Merck, Orange) |
| Source | Partner alignment | Due diligence on models and data vendors, Scope 3 analogy |
| Train | Human feedback | RLHF, red teams (DEF CON), user reports, annotator governance |
| Run | Surprises | Incident databases, adversarial learning, OOD detection, natural-language feedback |
Conclusion
Where alignment may decide competitive outcomes and even become part of product quality, firms must see both the risks and the differentiation opportunity, and adopt new practices to stay ahead. Customers and society expect companies to act on certain values, and the authors close that no firm can afford to launch AI products that misbehave.
The Air Force story was retracted: in June 2023 Col. Tucker Hamilton said he misspoke and it was a hypothetical thought experiment, not a simulation. Calling Anthropic "Alphabet-funded" is partial; Amazon became its largest backer. The chess-robot incident (Moscow, July 2022) fractured a seven-year-old's finger. The piece (March-April 2024) predates agentic systems, where alignment problems shift from what a model says to what it does.
On 22 January 2026 Anthropic published a new Claude constitution (CC0 licence) that explains reasons rather than listing standalone rules, keeping a few hard constraints. The EU AI Act's general-purpose model duties have applied since 2 August 2025; the Digital Omnibus, approved by the Council on 29 June 2026, pushed standalone high-risk duties (credit scoring, hiring) to 2 December 2027.
Credit scoring is an Annex III high-risk use. Architects should keep a policy-as-code layer (red lines, jurisdiction rules) separate from models, log every model and vendor in a register, and run post-deployment monitoring that model-risk teams already know from SR 11-7.
From Prediction to Transformation
Ajay Agrawal, Joshua Gans and Avi Goldfarb (University of Toronto, founders of the Creative Destruction Lab) argue that AI is cheap prediction, and prediction has value only when it improves a decision. Because decisions are linked, better predictions in one place change who decides and push uncertainty onto everyone else, so AI's real payoff needs redesigned systems that balance coordination with modularity. Like electricity and the PC, AI starts with narrow substitutions and transforms industries only once the system around it is rebuilt.
"Insights" is the wrong answer
The article opens with a familiar exchange at the Creative Destruction Lab, the seed-stage programme the authors run: an investor asks what a start-up's AI will do for businesses, and the founder answers "insights". The authors say they hear it constantly and read it as code for "we don't know what to do with our predictions." The better answer names the decisions the predictions will improve, because AI only creates value through better decisions.
The opportunity is large. Firms make more decisions than ever, in every function. Research by David Deming of the Harvard Kennedy School shows the share of jobs needing core decision skills (problem-solving, diagnosing, strategizing, prioritizing) rose from 6% in 1960 to 34% in 2018. But because decisions in one area affect others, AI usually means redesigning whole systems of decision-making, which the rest of the article explains.
How New Zealand won the America's Cup
Sailing has been refined for 5,000 years, and innovation never stopped even after commercial shipping abandoned wind. The America's Cup, sailing's top prize and the oldest trophy in international sport, is now as much about technology as crew skill. Teams spend millions on design and, because the physics is well understood, test designs in simulators without building boats. Emirates Team New Zealand won in 2017 with the best simulator.
Planning for 2021, the team worked with McKinsey to speed up design and found the bottleneck was the human sailors: a crew sails the simulator on a human time scale and cannot react faster. Using technology similar to the AI that beat the world's top Go players, they trained a bot to sail. It needed no sleep or food and ran thousands of simulations in the time humans ran a handful. After eight weeks it beat the sailors in the simulator. Then it began teaching them: moves that felt counterintuitive worked when tried on the water. It could test many boat variants at once, 24 hours a day, and try tactics, shortening the design cycle from the weeks humans had needed. New Zealand won the Cup seven races to three.
The novelty, the authors say, lay at the system level. Race preparation involves two kinds of decision: boat design (long simulated) and sailing manoeuvres (always worked out by humans). The AI never sailed the race, since the rules require real people at the helm, but the combined system of simulated boat plus AI sailor coordinated design and manoeuvres and improved both. The AI was not handing over insights; it was built into a system of decision-making.
Why system change takes time
New technologies are first applied narrowly. Electricity began as a substitute for steam where water was scarce; two decades after Edison's light bulb, only 3% of US businesses used electricity. In 1987 economist Robert Solow observed that the computer age showed up everywhere except in the productivity statistics.
AI has followed the same path. Whatever the branding, the authors see it as advances in statistics that predict more complex outcomes and exploit unused data. Early uses delivered what it could immediately: better, cheaper predictions than humans.
- Translation software predicts how humans would translate a text, based on past human translations.
- Medical-image AI predicts what expert radiologists would say a scan shows.
- Verafin, a Canadian firm bought by Nasdaq for $2.75 billion, sold AI fraud detection used by hundreds of financial institutions in place of their security teams.
Both early cases tap the wisdom of the crowd, and such substitutes can be very lucrative, but they are not transformational: they replace the humans who made predictions and leave the business otherwise unchanged. Transformation came when electricity freed factories from rivers and multi-storey steam layouts, allowing plants hundreds of miles from water spread over one floor and mass production; and when computers grew from calculators into what Steve Jobs called "bicycles for the mind", aids rather than substitutes.
The America's Cup lesson follows: the team kept people in the process. Full automation of every decision is imaginable but rare. AI prediction informs decisions that people make, and the real question it raises is who the best decision-makers now are, more than whether machines do more.
How AI is changing decision-making
Nobody at the smartphone launch thought it spelled the end of taxis, yet ride-hailing depended on connected phones offering cheap hailing and navigation. London cabbies spend three to four years learning the streets; AI on a phone now lets anyone predict the best route with traffic.
Decision = prediction + judgment. Prediction is data-based (given routes and traffic, how long will the trip take?). Judgment is a subjective read of context not easily reduced to data (does this passenger want speed or the scenic route?). When AI makes prediction cheap, decision rights flow to wherever judgment is still needed.
Taxi drivers have both skills; ordinary drivers have judgment but weak navigation. Navigation software closes the gap, and a platform that removes the meter, payment handling and central dispatcher lets any driver offer rides. Two effects followed: vastly more people could make ride decisions, and each driver's control shrank, since the platform matched riders and routes and drivers only had to give a pleasant ride. Together these weakened traditional cabbies and transformed the industry. The authors then describe three more patterns.
- Concentration without a change of control (hiring). HR screening once needed many people sifting hundreds of résumés. With AI, one HR executive sets the criteria for who gets an interview. The process and key decision-maker stay the same; fewer people are needed.
- Radical centralization (card verification). Merchants used to judge whether to accept a card: reject it if the signature did not match or ID was missing, wave through regulars. Crude database checks and now AI prediction automated this, and approvals follow rules set by a small group, probably a committee, whose risk parameters sit in the verification devices.
- Stronger local judgment (medical imaging). Physicians always made treatment decisions, but they used to need a radiologist's judgment to read the MRI, ultrasound or X-ray. With AI diagnosis replacing that judgment, only the physician's judgment remains, so the physician gains power and the radiologist loses it.
| Pattern | Example | Effect on decision rights |
|---|---|---|
| Widens and weakens | Ride-hailing | Anyone can drive; each driver controls less, the platform more |
| Concentrates | HR screening | One executive sets criteria; fewer screeners needed |
| Centralizes | Card verification | A risk committee's parameters replace merchants' judgment |
| Elevates local judgment | Medical imaging | Physician gains power; radiologist loses it |
In every case AI changed how and by whom decisions are made, and the effect does not stop at your firm: it reaches value-chain partners and the ecosystem, where what helps you may hurt them.
How AI shifts uncertainty
A restaurant cannot let diners order anything, because chefs' skills, order volume, ingredients and equipment limit what the kitchen can make. So it sets a menu, which buys reliability for the kitchen and drives weekly ingredient orders. With guacamole on the menu it orders 100 pounds of avocados a week: sometimes too many (waste), sometimes too few (lost sales).
With AI demand forecasting it orders anywhere from 30 to 300 pounds, wastes less, sells more and earns more. But its supplier, used to a steady 100 pounds, now faces more unpredictability, and so do its other AI-using customers. The supplier adopts AI too; its weekly order, once 25,000 pounds, now swings between 5,000 and 50,000. Its own source follows, and the volatility travels all the way to growers who set crop sizes a year or more ahead.
The authors call this a paradox: AI resolves one actor's uncertainty but makes that actor less reliable to others. The basic problem, matching supply with demand, is not solved; like a stone in a pond, one firm's AI sends ripples through the system, and partners must coordinate far more to absorb them.
Coordinating systems to align effort and resources
The restaurant also decides its menu. If growers cannot supply enough avocados, the menu should change, but it will only change if the restaurant knows, which needs coordination across decision-makers. Coordination has two aspects.
Synchronizing the work
An eight-person rowing crew must row in unison and adjust speed through the race so nobody runs out of energy. The coxswain at the back is essential for the second job but not the first: calling "Stroke!" can be done by a rower, and coxless boats do exactly that. What needs a separate person is gathering and aggregating information about race strategy and each rower's condition, then changing the rhythm so that adaptation happens in sync.
Assigning resources
Paul Milgrom and John Roberts call some coordination tasks assignment problems: a fixed amount of resource is needed, and more is waste while less is insufficient. If every ambulance received an emergency call and decided alone whether to go, you would often get none or too many. A central dispatcher (human or software) takes the information and assigns one. Sending a less-than-ideal ambulance (too far, wrong kit) is a much smaller problem than sending none or several.
Coxswains and dispatchers are communication systems that prevent the bad outcomes of poor synchronization or assignment. When AI creates coordination problems, new communication systems may be needed, and smart investment in coordination is how firms realize AI's promise.
Combining coordination with modularity
Ideally coordination would run entirely through communication, but that is not always feasible. A restaurant's supply chain spans thousands of kilometres and many months; aligning it through communication would be prohibitively costly and slow.
Amazon procures, warehouses, takes orders and ships millions of products, and also helps customers decide what to buy through recommendations. Its dilemma mirrors the restaurant's. If its AI says the best item to suggest is out of stock, refusing to recommend it would hide whether the prediction was right and forfeit learning. So Amazon recommends unavailable products and tells customers the likely delay. Most pick something in stock, some do not, and Amazon learns how much effort to put into carrying inventory for those items.
The balance needs careful design. Amazon's modular organization lets it slot better predictions into recommendations with little impact elsewhere, yet inventory and ordering cannot be fully independent of the recommender, because customer reactions generate information logistics must act on.
- Modularity. Insulates one part of the organization from the variability AI creates elsewhere and reduces the need for reliability; decisions rest on locally observable factors.
- Coordination. Counters the loss of reliability that comes with AI and lets decisions be optimized over a wider network.
- The rule. Successful AI systems enable coordination where possible, and modularity where necessary.
Conclusion: the people, not the technology
AI's promise resembles electricity's and the PC's: it began by solving a few problems in isolated, tightly bounded applications. As people work with it they will spot new ways to create solutions and raise productivity; restaurants, for example, will probably become more embedded in their supply chains and more flexible with menus. Each new opportunity brings challenges that open further opportunities. As AI spreads through supply chains and ecosystems, long-standing processes will be transformed, the authors conclude, by the creativity of the people using it more than by the technology itself.
The Solow paradox is back. An NBER study (February 2026) of about 6,000 executives in the US, UK, Germany and Australia found nearly 90% of firms reported no AI effect on productivity or employment over three years, with average use of only 1.5 hours a week. Team New Zealand won again in Barcelona in October 2024.
Written in November-December 2022, weeks before ChatGPT, the article frames AI purely as prediction. Generative and agentic models now act and produce, not only forecast, though the decision-rights lens still applies well. The radiology example also aged unevenly: AI has not replaced radiologists' reads, and demand for radiologists has kept rising.
Banking's card-verification example is the centralization pattern: a small risk committee's thresholds now govern millions of merchant decisions, so model governance is decision-rights governance. For architects, Amazon's lesson means loosely coupled services with explicit event feeds, so one team's better predictions do not destabilize another's planning.
How AI Can Help Managers Think Through Problems
Elisa Farri and Gabriele Rosani (Capgemini Invent's Management Lab, authors of the HBR Guide to Generative AI for Managers) report that two-thirds of managers see gen AI as a potential thought partner but only 30% feel they have the skills to use it that way. Their answer is "co-thinking": a designed, back-and-forth dialogue in which human and AI build on each other, so value is created in the conversation as much as in the output. They give a four-step method for designing such dialogues, two ready-to-run scripts and a short list of dos and don'ts.
The gap and the method
Wider access to conversational models has shown managers that gen AI can do far more than productivity chores. The two-thirds who see it as a thought partner expect it to offer fresh perspectives, weigh pros and cons, evaluate trade-offs, sharpen strategic thinking and support leadership development. Only 30% say they can do this today. Farri and Rosani built their guide from experience, many conversations and thousands of prompt iterations, and claim the gap closes quickly: human and AI together reach outcomes neither reaches alone, a process they call co-thinking.
The power of co-thinking
Co-thinking works for anything from speech preparation to root-cause analysis to strategy formulation. Three composite managers illustrate it:
- Mario, plant project manager. His team constantly faces technical problems whose root causes are hard to find, so he designed a structured dialogue with gen AI and fine-tunes and tests his prompt sequence each time a new problem appears.
- Kim, communications manager at a consumer-goods firm. She drafts press releases with AI and has it act as specific stakeholders who debate each draft with her. Her rule of thumb: challenge and be challenged by the AI.
- Amy, finance leader at a tech company. She must move her department from internal service provider to strategic partner for the business units, so she built a prompt sequence her team runs to reflect on the new mindset and behaviours required.
These are not question-and-answer or push-button uses. Both parties contribute and build on each other's ideas at every step, and the value lies in the dialogue itself as well as the output.
- Masayoshi Son. SoftBank founder; says he chats with ChatGPT every day to brainstorm ideas and strategies.
- Jeff Maggioncalda. Coursera CEO; calls gen AI an incredible thought partner that changed how he works.
- Valentin Marguet. Ferrari project manager; uses it as a methodological expert guiding a structured search for root causes in technical problems.
- Jessica Tompkinson. Former global head of communications and corporate affairs at Unilever Operations; learned the conversational approach the hard way (see below).
How to design a co-thinking dialogue
- Assign a role to AI. Choose a specific role, such as expert, team member or devil's advocate, and open with phrases like "Act as...", "Help me..." or "The output of the dialogue will be...".
- Define the setting. Decide who takes part: a one-to-one exchange between you and the AI, or one-to-many, for example in a team meeting or workshop.
- Outline the dialogue. Sketch the sequence of questions and statements and fix exactly what the AI brings and what the human brings; leaving the flow to chance raises the risk it goes off track.
- Create the prompt. Write the prompt from the outline, and let the AI improve it: paste your outline and ask it to turn it into a structured prompt it can run with you.
The authors suggest running one of their sample dialogues once, then adapting it to a real professional challenge or designing a new one from scratch; skill grows with experimentation. The scripts work in any model (they name ChatGPT, Claude and Copilot).
Two co-thinking dialogues to try
Both scripts start by assigning a role and instructing the AI to always ask for the manager's feedback before moving to the next step.
Understanding multiple stakeholder perspectives
Role: the AI acts as an expert in systems thinking for collaborative innovation, able to take many stakeholders' viewpoints.
- Frame the problem. The AI asks the manager for the problem, the external stakeholders needed to solve it and why the company cannot tackle it alone.
- Widen the map. The AI proposes three overlooked stakeholders; the manager comments and validates the revised list.
- Tabulate. The AI builds a four-column table: stakeholders, specific needs, unresolved pain points, root causes. The manager reviews and validates it.
- Prioritize risks. The manager picks the three most critical stakeholders; the AI offers three red flags and mitigation actions for each, and the manager responds.
- Act. The AI suggests three immediate next steps to start engaging each stakeholder.
Promoting a mindset shift
Role: the AI acts as an expert in mindset change, coaching on what is needed and how to practise it concretely.
- Context. The AI asks about the organization, the new mindset and the reason for the shift, then explains with examples or scenarios what the mindset means in practice.
- Self-assessment. The AI asks three questions to gauge how mature the person's shift is; the manager reflects and prioritizes the areas where the gap is largest.
- Ground it. For those areas, the AI asks for an example from daily work.
- Practices. The AI proposes three practices, tips or routines; the manager comments and selects one.
- Follow-up. The AI suggests reconvening in a few weeks to discuss problems met in applying the chosen practice.
Getting the most out of the dialogues
At Unilever, Jessica Tompkinson found her early prompts returned material that was useless for her needs. She changed tack and began talking to gen AI as she would to a colleague or an agency. The back-and-forth conversation markedly improved results and let the tool learn more along the way. The authors' advice: expect mixed results at first.
Do. Adopt a conversational mindset: treat the AI as a human collaborator; it can work like the Socratic method, guiding a structured dialogue, elaborating on your answers, adding overlooked dimensions and asking follow-ups. Participate actively: open with a brief, accurate overview and clear context, then share reflections and experiences and ask follow-up questions, so both sides challenge and co-create. Challenge the AI: friction is useful; ask for other perspectives and options, do not settle for the first output, and ask whether anything important is missing or how the problem looks from another party's view.
Don't. Go it alone: working mostly solo with AI cuts communication and knowledge sharing in the team and weakens team judgment and risk control, so take breaks for face-to-face work, bring colleagues into the AI-aided process, seek expert feedback and encourage peer learning. Go too fast: the tool's speed tempts people to click ahead without reflecting, so prompt the AI to ask for your feedback and give yourself time to assess and validate each step.
Conclusion
Managers who build gen AI skills will soon leap ahead of those who do not. Farri and Rosani tell managers to get hands-on: pick tasks to experiment with co-thinking, track benefits, risks and lessons learned and share them with the team. The manager's job is to be a co-thinking role model, inviting the team to discuss what works and what does not and collecting lessons in a structured way as the technology changes.
"Challenge the AI" matters more than the authors stress. OpenAI rolled back a GPT-4o update in April 2025 for sycophancy, and assistants still tend to agree with their user. The two-thirds and 30% figures come without a stated sample or survey source, and the evidence is practitioner anecdote, not measured outcomes.
Architecture review boards can script a co-thinking dialogue as a pre-mortem: the AI plays security, operations and regulator in turn against a design. In a bank, keep such sessions on approved enterprise models, since prompts carry confidential strategy.
The Gen AI Playbook for Organizations
Bharat N. Anand (NYU Stern) and Andy Wu (Harvard Business School) say leaders ask the wrong questions: how smart gen AI is, how fast it improves, whether the CIO is moving fast enough, what rivals are doing. The right questions are how to use it effectively today despite its limits and how to use it differently from rivals. A task's suitability for gen AI depends on two factors, the cost of errors and the type of knowledge it needs; and because everyone has the same tools, speed alone wins nothing; strategy does. The article offers a four-zone task matrix, a warning about the paradox of access, and three moves for building advantage.
Why move now, and why speed is not enough
The authors draw on work with hundreds of managers, gen AI initiatives they led and research on digital strategy. They call "wait and see", justified by hallucinations and other flaws, potentially dangerous, but they reject the idea that speed alone wins. Three arguments support moving now.
Nontechie employees can use it without experts
For decades AI belonged to engineers, programmers and data scientists. ChatGPT changed that by allowing natural-language interaction, so the breakthrough was in access as much as intelligence: anyone can use it without deep expertise, a data science team or central IT approval. It is also being built into email, videoconferencing, spreadsheets, CRM and ERP. The parallel is the 1980s move from command lines to the graphical user interface: Windows transformed personal computing by letting people use the machine's power without knowing MS-DOS commands, and gen AI does the same for machine-learning models, through writing and eventually speech.
Value-creation opportunities exist now
Waiting for a flawless, all-powerful agentic AI is a mistake. Gen AI already saves time, cuts costs and unlocks value; the benchmark should be relative efficiency against current ways of working, not perfection.
Advantage comes from using it more strategically, not faster
Everyone has access. If rivals use similar tools on similar tasks, new competition erodes margins and the gains flow to others in the value chain. Worse, customers and suppliers may disintermediate you by doing your tasks themselves. Advantage therefore hinges on which tasks you delegate and reimagine, how human expertise complements the tool, and what new possibilities you open.
Where and when to use gen AI
Rather than asking whether gen AI matches a human, break jobs into component tasks and ask which ones it suits today.
Compare two sets of tasks. Hiring critical employees, diagnosing cancer and psychotherapy for at-risk people are often cited as areas where AI nears human sophistication, yet replacing humans there meets strong and justified resistance: misdiagnosis or mishandling a vulnerable patient can change lives, and a wrong leadership hire can damage culture for years. Summarizing course evaluations, screening résumés and assigning hospital beds differ not necessarily in the intelligence needed but in the cost of getting it wrong: a summary that misses a nuance or a screen that overlooks a marginal candidate carries limited risk, and bed assignment rests on explicit, structured data (availability, patient needs, expected discharges) that AI handles reliably.
The principle: suitability depends on two factors beyond raw capability. Cost of errors is how serious the harm, financial loss or reputational damage would be if the AI errs; the higher it is, the less the AI should work without oversight. Type of knowledge contrasts explicit data (structured or unstructured information that can be captured and processed) with tacit knowledge (empathy, ethical reasoning, intuition, contextual judgment built through experience), which requires interpreting nuance and judging ambiguity, not just retrieving information.
Locate each task on the matrix and give it the matching level of AI support: some tasks suit AI alone, some human-AI collaboration, some purely human judgment. The question shifts from intelligence to usefulness: where is the cost of errors low enough to use gen AI today, and even where human insight is needed, which components could it handle?
The four zones
The no-regrets zone (low cost of errors, explicit knowledge)
This lower-left quadrant is the clearest and most immediate opportunity, where gen AI should be deployed now and where AI agents will thrive later. Data is clear and documented, errors are fairly harmless, and perfect accuracy is unnecessary; the value is doing tasks faster, cheaper or at greater scale.
- Résumé screening. Flag candidates against well-defined criteria.
- Low-dollar reimbursements. Approve tedious but low-risk claims.
- Routine customer replies. Draft answers on refund policies or shipping times, freeing people for higher-value interactions.
- Work never done before. Tasks too tedious or costly for humans, such as putting a stenographer in every meeting: AI captures the conversation and pulls out themes, action items and decisions in seconds.
Do not ask whether the output matches a human's or only how to automate existing work; breakthroughs also come from work that was never feasible. Key questions:
- Tolerance. Are the savings and speed gains large enough to accept a slight drop in quality?
- New work. What could we do with gen AI that we do not do today or that is too costly now?
The creative catalyst zone (low cost of errors, tacit knowledge)
In the upper-left quadrant gen AI helps with tasks that benefit from originality, while humans refine the output and make the final call. Mistakes are tolerable because quality is subjective: there is no single best slogan or perfect design. AI speeds experimentation, multiplies ideas and broadens participation, letting entry-level staff, people who never saw themselves as creative, and senior creatives alike work more like innovators. The method is to deconstruct the creative task and find where AI expands human capacity:
- Marketing. Produce 20 taglines instantly for the team to refine.
- Design. Generate visual or functional variants quickly, then hand-pick and perfect the best.
- Presentations. Outline points, suggest narrative arcs or create visual mock-ups, leaving people to tailor the message.
- Training. Generate mock interviews or simulations to enrich preparation.
Do not ask whether AI is as creative as a human, a standard it was never built for. Ask instead: can it save creatives time, and can it help noncreatives take part in creative work?
The human-first zone (high cost of errors, tacit knowledge)
The upper-right quadrant holds the highest stakes. Gen AI may enable but not decide. Tasks involve subjective judgment, nuance and complex decisions, mistakes carry financial, legal, reputational or personal consequences, and trust, ethics and long-term strategy are often at risk: a poor executive hire hurts culture, a strategic misstep can destroy billions, a mishandled diagnosis can cost a life. Hiring critical staff, setting strategy, integrating complex enterprise systems, handling crises and sensitive HR interventions all sit here.
AI must be used with extreme caution, in a constrained, supportive role. Smart task deconstruction still reveals help that leaves the human in control:
- Hiring. Refine job descriptions or suggest interview questions.
- Strategy. Synthesize market data or surface emerging trends.
- Governance. Model reputational risks.
- Crisis management. Draft preliminary communications and monitor public reaction.
- Healthcare. Help clinicians compute risk scores to triage who needs immediate attention.
- People management. Propose elements of a performance-improvement plan.
Every leader and knowledge worker has tasks here. Skip speculating about when AI will do them autonomously; the question is which tasks AI can assist today to make human judgment more effective.
The quality control zone (high cost of errors, explicit knowledge)
The lower-right quadrant holds knowledge-heavy tasks AI can technically do well, since they rest on explicit information, but where small mistakes are serious: law, finance and software development, with codified information and very high accuracy standards. It suits a human-in-the-loop model in which AI supplies speed and scale and humans supply judgment, oversight and final accountability.
- Contracts. Drafting once meant understanding client needs, composing clauses, negotiating, revising and approving; a lawyer using Harvey can now get a strong draft in minutes and concentrate on negotiation and final review.
- Code. GitHub Copilot writes boilerplate and suggests debugging fixes, but experienced developers still run quality assurance and verify functionality.
- Financial due diligence. AI scans large document sets for anomalies or opportunities; analysts interpret findings in context.
- Bed assignment. AI recommends allocations on structured criteria; clinical staff decide, weighing nuances the algorithm misses.
The rule: AI does the repeatable, data-heavy parts, humans do the steps where nuance, interpretation or accountability matter. Key questions: where is human expertise truly essential, and which parts of the workflow can be safely delegated?
The slogan that AI users will replace non-users is too simple. Some tasks are best done by AI alone, some jointly, some by humans only; the useful question is which tasks remain distinctly human, not replacement versus complementarity.
Anticipate the impact on your industry
Because customers, suppliers and competitors can all use the same technology, firms face the paradox of access: when everyone can use it, capturing value becomes dramatically harder. If rivals apply it to the same tasks with the same best practices, all become more efficient and none keeps the profit; competition passes the gains to customers and suppliers through lower prices or better terms. Internet 1.0 followed this pattern: early adopters had brief advantages, then benefits flowed to consumers. Airlines all adopted e-ticketing in the 2000s and customers got cheaper fares; CAD and ERP, which streamlined manufacturing and supply chains from the 1990s, are now table stakes. Two developments follow.
AI-first entrants are coming
The fiercest competition may soon come from solo entrepreneurs and micro-teams. A marketing agency built today need not hire dozens for research, copy, design and client questions: a small expert team, or one founder, could eventually hand all of that to AI and match an incumbent's scope and speed with a fraction of the headcount. The building blocks already exist in software development agents and AI sales reps.
Customers and suppliers can use gen AI against you
Their access can upend your bargaining power. Law firms have lived with this since the 1990s, when one lawyer with an internet-connected PC could do work that once took scores of paralegals and a law library, and companies began hiring in-house attorneys for routine matters.
- 3×. Growth in US lawyers working as in-house counsel from 1997 to 2020; they now outnumber lawyers at the 500 largest law firms.
- ~90%. Share of large firms now offering flat-fee or other client-friendly pricing, as clients attacked the billable hour.
- 100-hour weeks. Talent that once had to endure them at white-shoe firms can now go in-house or solo, using digital tools in place of big-firm infrastructure.
Gen AI speeds the pattern: with legal-research bots and contract-writing agents, clients can pull more work in-house. The same is happening in software development contracting, M&A consulting and advertising, and the most talented, entrepreneurial staff of those firms gain more choices of where to work.
Building an AI-based competitive advantage
Moving fast matters, but speed alone won't put you ahead; you need a strategy that differentiates how you create value with gen AI. The authors recommend three steps.
1. Mandate broad access to technology
Everyone has tasks in all four quadrants, so everyone should assess which of their tasks gen AI can handle, better or merely serviceably, and which tasks were too costly or slow to do before: sending personalized holiday greetings to every contact of the past year, or summarizing every meeting attended. Encourage experimentation and training through top-down messages signalling importance and bottom-up forums for sharing lessons, and build faster routes for frontline teams to test and scale tools.
Start by removing bottlenecks. If access stalls at the IT desk or behind compliance forms, rivals whose staff experiment in real time pull ahead. IT struggles to keep pace with the flood of models and apps, and handing the CTO full control, however capable, can slow progress.
In 2023 JPMorgan Chase temporarily blocked staff from ChatGPT while security teams ran third-party reviews. The authors call it sensible but note it kept 60,000 users from experimenting. Every firm faces the trade-off: cyber risks are real, but if the loudest message is what not to try, innovation moves at the pace of the slowest approval queue. IT should guard against the most critical risks, such as leaks of regulated or highly sensitive data like personally identifiable information, through targeted employee policies and precisely scoped vendor security reviews, instead of maximum precaution against every risk.
With access in place, differentiation needs two long-term efforts.
2. Reimagine all assets as data
Early gen AI knew only the public data it was trained on. Firms increasingly give staff proprietary data, through gen AI search or models trained on company knowledge. Two tasks:
- Find your data and centralize it. Pull together data scattered across business units, functions and geographies; infrastructure can anchor advantage. In the 2000s casino operator Harrah's Entertainment fed every slot pull, hotel check-in and dinner receipt into one data warehouse and grew revenue faster than rivals, who could copy the spectacle and glitz but not the data infrastructure or the culture of acting on it quickly. Gen AI raises the stakes, since it can mine messy, unstructured data, including partners' and acquired firms', for decisions across the organization. Building this takes years, so start now.
- Identify data you are not collecting. Customer interactions, operations, internal emails and meetings are all proprietary data sources. Uncollected data is a seed never planted; capture critical streams now so they can bear fruit later.
3. Redesign your organization
Layering gen AI onto existing workflows will not be enough; firms must redesign around a gen AI-first vision, organizing to get the most from data and people.
Proprietary data eventually commoditizes, but an organization built to exploit it is hard to copy. In the 1990s Capital One rewired itself around data by combining marketing, risk and IT teams that ran thousands of microexperiments a year, supported by operations, customer service and HR. Its famous balance-transfer teaser rate, letting customers move balances from rival issuers, drove explosive account growth. Tracking behaviour longitudinally, the bank saw newer applicants growing riskier and phased the product out. Competitors without the feedback loop kept copying the offer until their losses became catastrophic. Today's firms need the same loop from data to continuous learning, so gen AI insights become action ahead of the market.
People matter as much as data. Gen AI frees chunks of time, but early research suggests the windfall can evaporate into idle tinkering, low-value busywork or downtime. Treat time as a strategic resource: managers and employees should estimate and track the hours saved on key tasks, set clear expectations for redeploying them, and tie recognition or incentives to how well saved time is used, adjusting the measures as the technology changes.
Productivity gains do not flow to the bottom line automatically: unless every step of the value chain is aligned to capture them, efficiencies are wasted.
Design an AI-first org chart now, because redesign takes a long time. AI will eliminate some roles, most likely those heavy in no-regrets work. In the other quadrants it will complement people, though not necessarily the people doing the tasks today. Some functional staff may become cross-functional; middle managers may work directly with software instead of supervising people who do; a few people may focus only on human-first tasks.
- Rapid, targeted deployment across tasks, valuable in the near term while rivals stay fixated on intelligence or frozen by hallucination fears.
- Proprietary data that improves gen AI's performance, or process fixes that stop its value leaking away in organizational bottlenecks.
- Unique people, processes and culture, the complementary assets that make gen AI worth more inside one firm than in others.
Conclusion: four misperceptions
The trap. Gen AI isn't smart enough to be useful; its error rate is too risky; it must be 100% accurate before deployment; task-level savings should already show in the P&L.
The correction. It cuts costs even when quality is imperfect; the cost of errors, not the error rate, matters most; many tasks do not need perfect accuracy; saved time becomes saved money only with deliberate management, and lasting advantage comes from using it differently, not merely adopting it.
The winners, Anand and Wu conclude, will be the firms that spot these traps, rethink their assumptions and move deliberately to turn a general capability into a distinctive source of advantage.
Placing CV screening in the no-regrets zone is questionable. A wrongly rejected candidate bears the error, not the firm, and screening has produced bias cases (Amazon scrapped a CV model in 2018). The EU AI Act classes hiring tools as high-risk, and New York City requires bias audits of automated hiring tools. The "early research" on evaporating saved time is not cited.
JPMorgan reversed course after the ban: its in-house LLM Suite (summer 2024) reached about 200,000 daily users within eight months, and the bank says AI-attributed benefits have grown 30-40% a year. The NBER executive survey (February 2026) shows the P&L leak is real: nearly 90% of firms saw no measurable productivity effect yet.
Map banking tasks onto the matrix: KYC document extraction and call summaries are no regrets; credit memos, regulatory filings and production code are quality control with named human sign-off; credit-policy and conduct decisions stay human-first. Architects should make the matrix a routing rule in the AI platform, so the zone sets the required review step.
Agentic AI Is Already Changing the Workforce
Jen Stave (Digital Data Design Institute at Harvard), Ryan Kurt (the AI Lab, formerly Salesforce) and John Winsor (Open Assembly) argue that AI agents are turning from sidekicks into digital teammates, a new category of talent that HR and procurement must source, structure and govern like any other. Leaders who build an operational playbook for human-AI labour now keep control of how agents are sourced and regulated; those who wait risk compliance shocks, lost talent and lost contracts. They offer seven critical actions and five open questions.
The explosion of digital labour
As AI matures, digital labour is expanding the definition of a qualified workforce: agents now handle tasks once thought beyond automation. Salesforce CEO Marc Benioff says the total addressable market for digital labour could soon reach trillions of dollars. Emerging research from Harvard Business School and the Digital Data Design Institute finds agents becoming teammates in their own right, so HR and procurement need a playbook for fitting them into hybrid teams and workforce strategy; the authors promise a more scalable and resilient form of collaboration as well as efficiency.
It is already under way. Deloitte says it is applying AI agents to every enterprise process, including a marketing agent that orchestrates many tasks to optimize customers' journeys through its site. Talent firms such as r.Potential, a spin-off of staffing giant Adecco Group, have recast themselves as architects of a model that blends people and AI, not only suppliers of human talent.
Why inaction is a business risk
Firms that shape AI's place in their labour strategy now keep control of how it is sourced, structured and regulated; hesitant ones may be blindsided by compliance, ethics and performance problems. As with any disruption, hoping the technology goes away is not an option; firms must build the systems themselves to understand how to adapt. Delay also costs in three concrete ways:
- Talent. Strong candidates will expect AI-supported workflows that lift their productivity and creativity.
- Competition. Faster rivals will embed AI in their operating models and outscale and outlearn by raising output without adding headcount.
- Contracts. Large enterprise and government buyers are starting to demand robust, auditable AI policies and governance; immature firms may lose critical contracts and partnerships.
Seven critical actions
The framework draws on the authors' experience in data science and AI and in the open-talent and staffing ecosystem, as a guide for HR and procurement to design, test and scale a workforce strategy built on human-AI teams.
1. Map work tasks and outcomes
Break each role or project into component tasks and outcomes, as you would define competencies for human candidates, and identify which tasks agents can do better, faster or cheaper. High-volume data validation and repetitive call-centre functions are prime candidates; tasks needing complex judgment, persuasion or deep domain expertise stay with humans or go hybrid. Begin every sourcing discussion with this breakdown so the talent mix fits.
You are no longer buying labour; you are buying an outcome that may come from a combination of people and AI. Every sourcing decision should start from the task breakdown, not from the job title.
2. Assess AI capability
Know which models and platforms fit which tasks. Build an internal taxonomy of AI capabilities mapped to common roles: data validation and call centres, but also marketing analyst, customer support rep, scheduling coordinator and others. Language models may excel at marketing copy while computer vision agents suit manufacturing quality checks. This capabilities catalogue avoids one-size-fits-all buying and, in procurement terms, is your RFP for AI solutions. Partner with staffing or technology vendors that can prove domain expertise, so you do not pay top dollar for hype that does not match business needs.
3. Integrate your hybrid team
Human-AI teams need crystal-clear role boundaries. Write a hybrid-workforce strategy that defines tasks the AI owns, tasks people own, and how problems escalate. Example: if an AI customer-service agent receives a complex billing dispute above a set dollar threshold, a rule routes it automatically to a human specialist. Documenting roles, protocols and handoff points builds trust and prevents conflict and duplication.
4. Redesign your business (and workforce) model
Envision new ways to procure and deploy talent across full-time staff, temps, freelancers and AI, using multi-tiered models aligned with financial, compliance and strategic needs:
| Model | How it works | Best for |
|---|---|---|
| Client-owned digital labour | License or build your own agents: "digital employees" in-house | Core capabilities you want to control |
| Leased digital labour | Rent agents from a third party, like temp staffing | Seasonal spikes |
| Outsourced AI subdepartment | A vendor runs a whole process (order fulfilment, call centre) with AI plus a small expert team | Repetitive but critical functions |
Your job becomes integrating a wide spectrum of labour types, which needs new KPIs and cost structures reflecting digital labour's economics: scalability, near-constant uptime and quick retraining.
5. Set legal and ethical ground rules
Address bias, liability, data governance and societal effects up front. Work with legal, compliance and ethics teams on enterprise-wide AI standards defining whether and how AI learns from proprietary data, how bias is detected and remedied, and how personal or sensitive information is protected. Global firms should expect a patchwork of regulations. Missteps such as discriminatory hiring algorithms or data misuse can trigger PR crises and regulatory liability, and since governments are legislating fast, firms that build frameworks and an AI culture early will adapt to new laws far more easily than those forced to start after a law takes effect.
6. Capture value continuously as it evolves
Deployment is not one-off. Set up feedback loops that measure performance, update training data and revise sourcing. If an AI scheduling tool keeps hitting edge cases needing human help, it may need more training or stronger human oversight. Set-and-forget staffing does not apply, because AI's value compounds as it learns from interactions. Negotiate contracts that let you capture those improvements while respecting the vendor's IP and your enterprise's data sovereignty.
7. Remain human-centric
AI cuts mundane work and raises the importance of high-value, human-led tasks. Keeping people able to do them sustains morale and gives differentiating value that competitors cannot simply download. Invest in training that helps employees work alongside AI and use it to amplify their impact, focusing on relationship building, ethical decision-making and creativity, where humans keep a distinct edge.
Preparing for radical change
Whether firms adopt AI labour directly or through a staffing or open-talent provider, the authors pose questions to guide strategy:
- Ownership. When AI is trained on your proprietary data, who owns the resulting capabilities: you or the owner of the model or agent?
- Legal frameworks. Are new frameworks needed, including employment-like contracts for AI agents, and who is liable when the AI makes a mistake?
- Oversight and equity. Which oversight and fairness questions remain unresolved?
- Human or AI. What guidelines govern choosing between humans and AI for a job when ethics, brand reputation or job protection are at stake?
- The meaning of work. How will the definition of "work" change when agents are embedded in teams and perhaps gain legal or ethical status?
They need not be answered up front; firms should keep answering them along the way. The authors call them strategic inflection points: whoever resolves them first will shape how the future of work is defined and who controls its value.
Conclusion
The guiding principle is human-centricity. AI does many tasks faster, but the enterprise still depends on the insight, empathy and relationships only people provide. Holding both aims at once, unleashing AI's efficiency and protecting human creativity, gives the best chance of sustainable growth.
Gartner (June 2025) predicted over 40% of agentic projects will be cancelled by end-2027 over cost, unclear value or weak controls, and counted only about 130 genuine agentic vendors amid "agent washing." In May 2026 it urged governance by four autonomy levels (observe, advise, act with approval, act autonomously) rather than one rule for all agents.
Action 3 maps directly onto banking's four-eyes principle and delegated authority limits: give each agent an identity, entitlements and a transaction ceiling in IAM, exactly as for a junior employee. Architects should log agent actions to the same audit trail regulators already inspect.
This is a framework from advisers with stakes in open talent and AI consulting, with no measured outcomes; the Harvard research it cites is not described. Benioff's "trillions" is a vendor's pitch, and the article says little about the workers whose tasks the agents absorb.
Keep Your AI Projects on Track
Iavor Bojinov (Harvard Business School, a LinkedIn data scientist in 2018-2019) notes that some estimates put AI project failure as high as 80%, almost double the failure rate of corporate IT projects a decade earlier. Every AI project passes through five stages on its way to becoming a product, selection, development, evaluation, adoption and management, and most failures trace to skipping the discipline each one requires. Done well, the stages build a lasting "data science and AI operations" capability, the skill to compete as business becomes AI-driven.
From a niche team to the boardroom
At LinkedIn in 2018-2019, AI interested only a small group of data scientists with advanced degrees in statistics or computer science. Now AI, and especially generative AI, is a boardroom topic and a hallway conversation among staff keen to boost productivity. "Data Science for Managers", a course Bojinov helped create to teach MBAs to develop, use and manage AI, is now a first-year requirement at HBS. Behind the headlines, though, most projects fail, and Bojinov draws on industry, academic and consulting work to show how to cut that risk.
1. Selection
Prioritizing projects is familiar to leaders, but AI projects have quirks in how impact and feasibility are judged. First, treat internal- and external-facing projects separately.
- Internal-facing. Help employees do their jobs: telling sales teams which customers to target and what to offer, predictive analytics for supply chains, HR chatbots answering staff queries.
- External-facing. Deliver AI to customers; the most visible uses, pioneered by tech firms: Netflix's recommendation engine, Google's search results, Uber's matching algorithm.
Impact checklist
- Strategic alignment. Is the project in line with strategy? Data scientists often lack a full grasp of strategy, sit in silos and pick projects unlikely to create much value. Procter & Gamble temporarily embeds data scientists in business units; LinkedIn keeps central reporting but has teams that mirror business units and sit with them. Both help technical experts learn the business.
- Measurable impact. Can benefits be assessed objectively? Use the "if, then, by, because" paradigm: if the project is implemented, then a named KPI will improve by a stated amount, because of a stated rationale. Data-driven firms champion this hypothesis-driven approach. Quantifying impact allows direct comparison across projects and counters technical teams' pull toward the newest technology.
- Augment or replace. Does it enhance human work or replace a manual process? That depends on the cost of errors. Recommending an unwanted product costs little, so automate it; misdiagnosing a serious illness costs a great deal, so doctors consult AI but keep the final word.
The impact assessment also spreads across teams and divisions, creating shared understanding and cross-pollination of ideas.
Feasibility checklist
Get a basic cost estimate and check whether resources exist, across four factors:
- The nature of the problem. AI excels at trends, patterns and predictions for well-formulated problems, but cannot grasp context, use emotional intelligence or make moral judgments: it is good at spotting a sales prospect and poor at closing the deal.
- Data availability. Success depends on the data's availability, quantity, freshness and quality. A lead-prioritization tool needs prospects' job titles, industries and company sizes plus names and phone numbers.
- Technological capability and skills. At minimum: data storage and management, enough compute to train and run the system, and guards on data flowing in and out; usually data science and data engineering staff.
- Ethical considerations. Identify implications and plan to address them during development, or pay later in reputational harm, government fines and engineering time to retrofit.
The three ethics themes
- Bias. Training data that misrepresents the served population lowers accuracy and produces unfair outcomes, systematically underperforming for some groups. A biotech firm building a cancer diagnostic for men and women of all races, with data on white men only, will get biased or inaccurate results for women and people of colour.
- Privacy. Models must protect personal data and guarantee it will not leak. Combine "privacy by design" principles, developed by Ann Cavoukian, former Ontario information and privacy commissioner, with tools such as differential privacy, a mathematical measure of data privacy.
- Transparency. Users need to know how a model works and its strengths and limits. Set the desired level before starting, because of an inherent trade-off: more explainable models tend to be less accurate, while hard-to-explain ones often perform better.
Classifying each project as high or low on impact and feasibility lets managers estimate ROI and choose and sequence projects:
| High feasibility | Low feasibility | |
|---|---|---|
| High impact | Do first | Find root cause; strengthen data or infrastructure |
| Low impact | Skip unless cheap or a useful sandbox for testing infrastructure and tech | Drop |
2. Development
Once approved, the complex, slow build begins, and intended users should be engaged throughout to make adoption likely. Data scientists iterate through finding and cleaning data, exploratory analysis, and training and evaluating models until they reach the desired accuracy or drop the project. Then comes integration, which turns a model into a product: software to move data to the model and outputs to where they are used, perhaps custom interfaces or links to systems such as CRM.
Most companies do this ad hoc, with little standardization or specialization, which is inefficient and failure-prone. LinkedIn, Netflix and Uber built internal tools covering everything from exploratory analysis to deployment; Databricks, Snowflake and the big cloud providers sell this automation layer as a service, but using it requires a centralized data repository.
HBS colleagues Marco Iansiti and Karim Lakhani call this automation and standardization the "AI factory" (HBR, January-February 2020). It speeds development, standardizes key steps for more monitoring and oversight, and raises model quality: in a 2022 experiment, giving data scientists automation tools improved final model accuracy by 30% and reduced the skills needed to build a product.
How much gen AI tools for text, images and code will improve development is not yet clear, but a controlled GitHub Copilot experiment found software engineers gained significant productivity, with the least experienced benefiting most. The leadership takeaways: create a centre of excellence to build an easy-to-use AI factory, provide tool-specific training, and confirm that ethical issues flagged at selection have been dealt with.
3. Evaluation
Before pushing adoption, measure the product's impact. Scientific experimentation, at its simplest A/B testing, is the gold standard (HBR, September-October 2017): users are randomly assigned to the existing offering or the new one and engagement, satisfaction and other metrics compared. Firms build their own platforms or buy from LaunchDarkly, Optimizely and Split, and experimentation will be central to gen AI, testing how users respond to different outputs.
Four reasons accurate models still fail
- System interactions. AI interacts with other products, systems and processes. A content recommender may lift engagement by showing fewer ads, cutting profit and offsetting its benefit.
- Unrepresentative data. Faced with scenarios outside its training data, a model extrapolates and usually performs worse; self-driving cars are hard partly because no dataset covers every driving situation.
- Negative feedback loops. A recommender that keeps showing content very like what users already engaged with may be fine once or twice, but by the 20th time they find it boring and leave.
- Failure to adapt. A pricing model trained on historical sales may miss trends, sudden shifts (a pandemic, high inflation) or reasons for purchases such as availability or return policies.
What experiments add
Experiments also give rapid user feedback explaining shortfalls. In a 2021 study with LinkedIn, Bojinov estimated that using such feedback improves business outcomes by 20%. One LinkedIn team found a model worked worse for some users because training data was unrepresentative; adding personalization, a simple change, drastically improved it. Experiments also limit how many users see a bad experience, and new statistical methods allow automated continuous monitoring that halts a test once a significant negative effect appears: reanalyzing a two-week Netflix experiment, the method would have stopped it after one day.
Beyond the usual pitfalls (covered by Bojinov with Guillaume Saint-Jacques and Martin Tingley in HBR, March-April 2020), two problems make some AI A/B tests hard:
- Tiny user bases. Tools for a handful of companies or employees: stagger exposure and compare reactions before and after adoption, which reduces bias from idiosyncratic events.
- Global optimization. For supply-chain optimization, workforce scheduling and product-mix optimization, testing on a subset may yield an inferior solution and miss benefits; switchback experiments alternate all users between running with and without the AI.
4. Adoption
Bojinov led the LinkedIn team that launched an internal AI-driven data-analysis tool: data scientists described the study they wanted in a simple interface, with no code, and the AI back end did the rest. Early experiments showed analysis time falling from days to a few hours. Yet after launch few people used it; they kept running custom analyses on their own computers. They did not understand the product, were unsure it was built for them, and did not know who would be at fault if it produced wrong conclusions. In short, they did not trust it.
People trust AI when they believe it works, have faith in its developer, and see the process as empowering them without undue risk. Failure to adopt an AI product can almost always be traced to missing trust in one of three pillars: the algorithm, the developer or the process. The remedy starts with talking to people who tried the product and stopped.
- The algorithm. Doubts: how does it work, is it biased, what assumptions went in, why these predictions? Fix: first check efficacy (bias may have slipped through development and evaluation); then explain the encoded assumptions so outputs match users' intuition and business logic, explain the training data to show the knowledge built in, and give real-world use cases. LinkedIn's root problem was that simplicity had created a black box, so the team wrote educational materials on its inner workings and assumptions.
- The developer (an organization, team or person). Doubts: hidden intentions, was I heard, is it taking a task I enjoy? Arises when users were left out of development: they suspect a hidden agenda to replace their jobs or assume their needs were ignored. Fix: clear communication and a transparent statement of purpose. At LinkedIn, users consulted during development trusted the team, but others assumed the tool solved only a few specific problems, so the team found users with distinct problems and helped them use it, a costly effort that proved its flexibility.
- The process. Doubts: if I follow a wrong recommendation, who is at fault; may I overrule it? Fix: when AI augments a decision-maker, build a feedback loop and clear rules for resolving disagreements; when it automates, put strong protections in place to catch and correct errors. LinkedIn created a certification board that reviewed results and recommendations before insights were shared, so responsibility for later-found mistakes rested with the board, not users.
After these steps the tool was adopted rapidly across the organization, and five years later it is still in use.
5. Management
Adoption is not the end. Products need engineering support (bug fixes) and performance monitoring. The most common cause of decline is outdated training data: a model predicting customer purchases loses accuracy as preferences and markets shift. Regular retraining on fresh data is costly, so many firms build mechanisms that monitor performance and alert managers to significant changes that may call for retraining.
Firms should also run AI audits for unintended consequences, ethical issues and security flaws. In a study published in Science in 2022, Bojinov and coauthors found that LinkedIn's People You May Know algorithm, meant to widen users' networks, inadvertently changed their career prospects, partly because social networks shape access to information about jobs and opportunities. This time the effect was beneficial, but it could easily have been harmful, which is why audits are essential.
Conclusion
Management means continual improvement, keeping the product in step with technology and users' changing needs through a cycle of gathering more data, refining algorithms and promoting usage. Bojinov's claim is that working through the five stages greatly raises the odds that leaders pick the highest-value projects and that those projects deliver.
The "80% fail" figure is a loose estimate with shifting definitions of failure. Bojinov's evidence leans on LinkedIn, a data-rich tech firm with a strong experimentation culture; most firms lack the user volume for clean A/B tests. The article (November-December 2023) predates agentic systems, whose multistep behaviour is harder to A/B test.
Failure rates have not improved. MIT's NANDA initiative (The GenAI Divide, August 2025) reported that about 95% of enterprise gen AI pilots showed no measurable P&L impact, and Gartner (June 2025) expects over 40% of agentic projects to be scrapped by 2027.
Bojinov's stages map neatly onto SR 11-7 model risk management: selection is model inventory and tiering, evaluation is validation, management is ongoing monitoring. The certification board is essentially an independent validation function; banks can extend it to gen AI with evaluation suites and drift alerts.
How to Marry Process Management and AI
Thomas H. Davenport (Babson; he helped launch the process-reengineering movement in the early 1990s) and Thomas C. Redman (a data adviser who applies process thinking to data quality) argue that process management and AI need each other. AI only supports narrow tasks, so its value scales when it is strung together inside owned, measured end-to-end processes; and well-managed processes in turn supply the clean data AI needs. From conversations with hundreds of leaders they propose a new process philosophy built around people, data, analytics and AI, and a seven-step method illustrated with order-to-cash (OTC).
Mars Wrigley reimagines its supply chain
To digitize its supply chain, Mars Wrigley built a digital twin of its production line (a virtual replica simulating operations in real time) and fed its data into a machine-learning model that predicts output to reduce overfilling and waste. It used the decision intelligence vendor Aera Technology to visualize data, recommend preventive maintenance and automate some operational decisions, and hired Kinaxis, whose AI suggested how to balance supply and demand, automated invoice processing and raised truck utilization by 15%. Orders were filled faster and customer service ratings rose a couple of percentage points. Next came ML sales forecasts to set production levels, with smart robots and new AI systems planned on the line for efficiency and sustainability.
Why process management fell from favour
Process management is simple in concept: understand how a sequence of tasks combines to produce a specified outcome, then improve it. It applies at many levels: an individual's work, a small group, key activities in a department, or end-to-end processes crossing the whole organization and even company boundaries. Done well it is extremely effective: error rates, cycle times and low-value work all fall. But it is hard to scale, even with AI, for four reasons:
- AI is narrow. It supports tasks or subprocesses, so firms must string together multiple use cases to improve a whole process.
- Change management. It requires massive effort: persuading stakeholders, retraining workers, integrating many moving parts.
- Hierarchy. It cuts across departments, putting it at odds with traditional line management.
- Reputation. Reengineering became a fad in the early 1990s that left failed projects and mindless layoffs behind.
Technology has disappointed too. Robert Solow's 1987 remark that computers show up everywhere except the productivity statistics still holds: firms have spent trillions of dollars on data and technology with little to show, and most lack ROI figures for IT projects beyond productivity metrics.
The authors place their philosophy in a long lineage of process ideas with real impact: scientific management, statistical quality control, total quality management, Six Sigma, agile and lean. New process thinking keeps appearing because firms constantly need better operational performance, and managing processes is a reliable route to it.
How technology and process enhance each other
Weak technology support, especially in incremental methods like lean and Six Sigma, helped process thinking decline. New technologies now make transformation possible, which is why the authors say firms should return to process management broadly.
The relationship runs both ways. Technology scales improved processes; a new AI algorithm is far easier to deploy and profit from when the process it supports is well defined and managed; and end-to-end integrated processes yield the large volumes of high-quality data needed to train AI. That requires departments to adopt common data standards and share data freely, which many resist because their systems were built for local needs, not end-to-end effectiveness.
The silo trap. An operations team assembling orders finds sales' order data full of holes and errors and quietly cleans it up to hit its own KPIs, untrained and under pressure. Sales never learns its data is bad, and operations, stuck in its silo, never sees that the real fix is fewer errors at the source. Process management exposes such errors, motivates improvement and gives the structure to remove their causes systematically.
Where to begin
Firms without a strong process orientation should start with one or two performance-critical processes and add others as process thinking takes root. Many start with order-to-cash, every step from a customer's order to receipt of payment, which matters greatly to performance and is a favourite reengineering target.
| Firm | Approach | What it did |
|---|---|---|
| Siemens | End to end (Europe) | Common OTC subprocesses (order management, purchase-to-pay) despite decentralized units |
| BMW | End to end (Europe) | Standardized most global production processes, now optimizing support processes |
| Reckitt | End to end (Europe) | Process mining and automation across OTC: invoicing, supply purchasing, fulfilment |
| Uber, PepsiCo | Incremental (N. America) | Customer service; accounts receivable and payable |
| Cardinal Health, J&J | Incremental (N. America) | Order management; supply chain |
All the North American firms made significant gains with a narrower focus, which the authors recommend for newcomers. They then walk through the seven steps using OTC.
The seven steps
Step 1: Establish ownership
Assemble managers able and willing to take end-to-end responsibility for the process: a process owner to coordinate and process managers from the departments involved (for OTC, sales, operations, shipping and finance, with others contributing), who can speak for their departments while advocating for the whole organization. The owner is usually an entirely new senior role, and the best owners exercise influence without formal authority, since early on they may have little.
Conflict is built in, because process management aligns with customers' priorities while line management aligns with the boss's. Employees need guidance on who matters more when time is short. A salesperson pushed by a manager to book an order promising 10-day delivery to hit quota is stuck when the process owner says inventory is low and 20 days is realistic.
Step 2: Identify process customers
Ask who the customers are, what they want and need most, and how to deliver it; gaps in the answers signal chances to rethink the process. For OTC the primary customers are buyers, who need top-quality products on time and where expected, and the company, which needs to manage cash. Secondary customers might include marketing, if it wants to sell on-time delivery as an advantage, or a sustainability group cutting the carbon footprint. Process management thus forces managers to set priorities and align work with them. CRM systems reveal attrition, service-request volumes and customer profitability; gen AI analyzes and summarizes comments from service calls, emails and social media, and keeps improving.
Step 3: Map out the existing process
Draw a high-level flowchart of the current process, covering the physical movement of goods and the creation, flow and use of data. Once done with Post-it notes on whiteboards, this can now be automated by AI. Pay special attention to interfaces between departments, the "white space" no department clearly owns, where delays, errors and inefficiency breed.
Process mining, unavailable in earlier waves, extracts data from enterprise-system log files to model and analyze performance: in OTC, almost in real time, which computer-mediated tasks are under way and how long it takes to fill an order, deliver a package and get paid. It highlights the pain points a redesign should target; teams can then use task mining, which focuses on automating smaller processes.
PepsiCo began process mining on accounts payable in 2019, since then saving thousands of hours of labour a year and cutting write-offs by millions of dollars. It now uses Celonis across nine more processes, including end-to-end OTC, and has automated tasks such as messages to customers about overdue receivables. AI pinpointed the biggest problems, notably a 30% order-rejection rate after a new SAP installation, which process mining with AI brought down to 4%.
Step 4: Establish performance measures and targets
Define the metrics to manage the revised process. For OTC the key ones are end-to-end cycle time (order to payment), customer satisfaction, data accuracy and process efficiency. Decide how much improvement is needed: reengineering often sought 10-fold gains, lean and Six Sigma much smaller increments, but targets should rest on what is needed and what seems possible, not arbitrary ambition or caution. Metrics also guide design: one telecom found a service took about 90 days to deliver, of which actual work was about 10 days and 80 days were waiting between steps. Better coordination removed 60 days, pleasing customers and bringing in cash much sooner.
Step 5: Consider process enablers
- Robotic process automation. Bots from Microsoft, UiPath and others automate repetitive, routine workflows; useful for small processes.
- Generative AI. In OTC it can draft contracts, help customers place more accurate orders and alert them to delivery changes.
- Traditional machine learning. Optimizes pricing, speeds credit approvals, prevents fraud and estimates staffing levels.
- Other technologies. IoT sensors monitor equipment and prevent failures; blockchain tracks goods along a supply chain; classic business analytics improves decisions throughout.
Step 6: Redesign the process
A cross-functional team from the departments involved leads, aiming beyond a better workflow to identify the skills, technologies and structural changes needed and the expectations of partners and customers. Design was once labour-intensive; AI now speeds it. Pega, a long-standing process-software vendor, built gen AI tools that draw best-practice designs from a library and suggest them to teams.
Deutsche Telekom, Germany's largest telecom firm, had accumulated some 800 HR processes in more than 20 countries. Traditional design and development tools had made slow progress, partly through misunderstandings between business and IT. Pega's Blueprint removed them with a simple interface where business experts describe processes in their own words; it also flagged things the team might have missed and generated process templates. The company has streamlined 250 processes with more planned, raised employee satisfaction, freed HR staff from repetitive work to serve employees better, and saved many millions of euros in application operation and management costs.
The authors also expect much from gen AI image creation, generative design tools (already used by architects) and digital twins, to help teams imagine, simulate and eventually optimize process designs and flows.
Step 7: Implement and monitor the process
Rollout takes real effort even when AI automates key tasks: employees need training, data must be integrated, systems built and customers briefed. It may take months, but should not take years. Afterwards, establish a new normal of control and continuous improvement. The essence of control is predictability, confidence that performance will not deteriorate, and it is the opposite of the firefighting that is too often the norm.
Process mining is perhaps the most valuable monitoring tool: it uncovers the variation all processes show, and excessive variation should be investigated and removed. In OTC, one cheaper but less reliable carrier among several adds enormous variation, so in its redesign Mars Wrigley hired a fourth-party logistics firm to work with carriers, monitor them and ensure reliable delivery. Mining should be constant, because business keeps changing: new customers with new needs, new products needing special handling, new regulations requiring new reporting, new inventory technologies offering gains. Good teams stay ahead and keep improving even radically redesigned processes.
Conclusion: put process back on the radar
The reasons to manage processes are greater and the difficulties smaller than ever. The virtuous cycle of better data, easier technology implementation and productivity gains has raised the payoff, and new tools make the work faster. Firms that kept a process focus have reaped the benefits. The authors judge process management essential for any firm serious about AI: executives should think about how people, data, technology, AI and analytics combine, with process at the core, because processes are how organizations deliver value.
The book dates this article January-February 2022, but it discusses gen AI design tools such as Pega Blueprint that appeared only from 2023; it ran in 2025 (reprint S25011). The cases are largely vendor success stories (Celonis, Pega, Kinaxis, Aera) with few independent outcome figures, and the authors do not address how process redesign interacts with job cuts, the issue that damaged reengineering.
The pitch has moved from copilots to agents that run whole workflows. Gartner (April 2026) predicts that by 2028 over half of enterprises will stop paying for assistive AI in favour of platforms that commit to workflow outcomes, with humans becoming "agent stewards." That makes the authors' process owner more important, not less.
Banks' equivalents of OTC are loan origination, payments exceptions and KYC onboarding, all full of white space between front office, operations and compliance. Process-mining event logs from core banking give architects a fact base for redesign, and a named process owner is the natural accountable person for any agents deployed in the flow.
The Hidden Penalty of Using AI at Work
Oguz A. Acar (King's College London), Phyliss Jia Gai (Peking University), Yanping Tu (Hong Kong PolyU) and Jiayi Hou (University of Hong Kong) explain low AI uptake with a hidden social cost, the competence penalty. Colleagues judge identical work as coming from a less able person when they think AI helped, and the penalty falls hardest on women and older workers, so avoiding AI is rational self-preservation rather than reluctance. Fixing it takes three interventions: map the hotspots, convert influential skeptics, and strip the AI signal out of evaluations.
The adoption puzzle
The story opens with a frustrated VP of engineering at a leading tech company. Twelve months after rolling out a state-of-the-art AI coding assistant, only 41% of engineers had even tried it; female engineers adopted at 31% and engineers aged 40 and over at 39%. The figures come from the authors' research with 28,698 software engineers there. Leadership had done everything by the book: cutting-edge tools, attention to access, infrastructure and skills, and a dedicated deployment and promotion team.
The pattern is national. A Pew Research Center survey found that two years after ChatGPT's launch only 16% of American workers used AI for work, although 91% were allowed to. The usual explanation is skill gaps or training friction: people do not see where AI helps or lack technical confidence. Yet even among early adopters such as software engineers and data scientists, who use AI more than other workers, usage stays low and unequal.
The competence penalty
The authors ran a preregistered experiment with 1,026 engineers at the same company. Participants evaluated a Python code snippet supposedly written by another engineer, with or without AI assistance. The code was identical in every condition; only the described method differed.
- 9% lower. Average competence rating of an engineer believed to have used AI, for identical work. Ratings of the code's quality stayed about the same: the penalty hit the person, not the output.
- 13% vs 6%. Penalty for female versus male engineers, more than twice as severe for women, whose fundamental abilities were questioned far more.
- Non-adopters harshest. Engineers who had not adopted AI themselves were the most severe critics.
- 26% harsher. How much more male non-adopters penalized female AI users than male AI users for the same usage.
- 919 engineers. Respondents to follow-up surveys, many of whom anticipated the penalty and avoided AI to protect their reputations.
The surveys completed the picture. Those who most feared the penalty, disproportionately women and older engineers, were exactly those who adopted AI least. The groups that might gain most from productivity tools felt they could least afford to use them.
The cost of the penalty
The authors call it a hidden tax on AI adoption: apparent reluctance is really rational self-preservation, and the costs go beyond lost productivity.
- Lost profit. With adoption at 41%, lower in key groups, the company captured less than half the expected benefit of its heavy investment: a net loss of at least 2.5% of annual profit and possibly 14%, depending on the method (salary value lost versus projected profit gains). Even the conservative figure means hundreds of millions for a firm that size.
- Shadow AI. Staff who fear official tools may turn to unauthorized ones, more often than employers expect. These are hard or impossible to track, expose the firm to data security and compliance risks, and undermine responsible use through inconsistent practice.
- Wider inequality. It is tempting to assume AI levels the field by augmenting everyone. The results suggest the opposite: in a context dominated by young men, making AI equally available increases bias against female engineers.
- The disclosure dilemma. Transparency is central to responsible AI, but internally it may do harm: requiring staff to disclose AI use to colleagues can create unfair career risk. Where penalties are likely, it may be more responsible not to require disclosure.
The mechanism is social identity threat. When members of stereotyped groups (women in tech, older workers in youth-dominated fields) use AI, it confirms existing doubts: AI help is read as proof of inadequacy instead of smart tool use. Any industry dominated by one group should expect larger penalties on its minorities.
Breaking the penalty
The study company had already done more than most: dedicated AI teams, incentives, training. It fell short because none of it addressed the underlying competence threat. The authors propose three interventions.
1. Map your organization's penalty hotspots
The people who would benefit most fear the risk most; at the study firm, young women in tech stood to gain the most productive time yet adopted least, which makes them low-hanging fruit for anyone maximizing AI returns. Look for demographic vulnerability and power imbalances:
- Team composition. Few women or older engineers in junior roles under many senior male non-adopters is the perfect setting for the penalty.
- Career data. Compare time-to-promotion by demographic and AI use to see whether the penalty already affects careers and how urgent it is.
- Disclosure rules. Check whether vulnerable groups face mandatory or voluntary disclosure and how it affects their career outcomes.
2. Convert the influential skeptics
Non-adopters impose the harshest penalties and are still the majority in most firms. Breaking the cycle takes strategic social influence: when people see respected colleagues use AI with no professional cost, perceived risk falls. The strongest champions mirror the penalized groups. Senior women in the study feared the penalty less than junior women, and when they use AI openly they give cover to vulnerable colleagues. A BCG study found that when senior female managers lead male peers in AI adoption, the gap between junior women and men shrinks significantly.
Whoop. Head of product Hilary Gridley ran a "30 Days of GPT" challenge that moved from simple tasks such as meeting agendas to complex workflows with custom GPTs. Daily demonstrations and public celebration of wins created psychological safety, as people saw peers at every skill level succeed; skeptics yield when colleagues they respect, not just senior leaders, use AI for real work.
Pinterest. Its annual Makeathon, a company-wide hackathon open to technical and nontechnical staff, had teams pitch AI ideas, build prototypes and show projects, with internal leaders acting as "hack doctors" and team leads to lend credibility. Afterwards 96% of participants reported continued monthly AI use and 78% of engineers credited AI with time savings.
3. Redesign evaluations to remove the signal
Tagging code "AI-assisted" invites biased evaluation as long as the culture is unchanged, so stop signalling AI use in performance evaluations until the culture is ready. Since the experiment found no quality difference, evaluate what is accomplished, not how:
- Objective metrics. Replace subjective competence ratings with cycle time, accuracy and defect rates.
- Audit-only tags. Phase out public "AI used" tags once compliance logs exist, keeping them for internal audits.
- Blind reviews. Hide personal information so evaluators cannot see who wrote the code or whether AI helped, reducing bias against groups already under scrutiny.
Some firms go further and reward AI use, turning penalty into competence bonus. Julia Liuson, president of Microsoft's developer tools division, told managers to make AI usage part of their holistic reflections on performance and impact, and some teams are weighing formal AI metrics for next year's reviews. Shopify CEO Tobias Lütke made "reflexive AI usage" a baseline expectation, with AI-use questions added to performance and peer reviews, framing proficiency as an "unobvious skill" that needs feedback and development.
The path forward
The penalty exposes a basic misalignment: firms concentrate on access, training and infrastructure and overlook the social dynamics that decide whether people use the tools. It explains three puzzles: why staff secretly use ChatGPT while ignoring approved tools, why the groups that could gain most adopt least, and why firms see limited ROI despite huge investment. Women using AI suffer nearly twice the reputational damage of men, and older workers in young fields face similar bias; those who most need AI's equalizing power can least afford it. The penalty is not inevitable, the authors conclude: the leaders in AI will not be the firms with the best tools or budgets but those where every employee can use AI safely; until then companies waste licences and, worse, human potential.
Use has grown fast. Pew's February 2026 survey found 40% of employed men and 35% of employed women use chatbots for work tasks, and the gender gap in ever having used a chatbot has nearly closed (50% versus 47%).
The headline evidence comes from one company and one code snippet. The book's note dates the Pew report February 2024, but it was published in February 2025. Recommending against disclosure sits awkwardly with the transparency other chapters demand, and mandating AI use, as Shopify does, may swap a competence penalty for a conformity pressure.
In banks, code review and credit-memo review are exactly where tags bias raters. Track defect rates and turnaround rather than "AI-assisted" labels, while keeping usage logs for model-risk and audit teams.
Heavy Machinery Meets AI
Vijay Govindarajan (Dartmouth Tuck) and Venkat Venkatraman (Boston University Questrom), authors of Fusion Strategy and students of industrial firms for more than 40 years, argue that industrial winners will be separated by the deepest insights from product-in-use data, not by the most valuable physical assets. Incumbents must fuse steel with silicon: reimagine analog products as digitally enabled ones, create new value from the data they generate, and compete as partner ecosystems focused on customers' problems. They lay out four fusion strategies, products, services, systems and solutions, and how to capture value from each.
Deere: a 187-year-old fusion firm
Deere & Company has simplified farmwork for more than 187 years, from the first self-scouring plough in 1837 to its first fully self-driving tractor in 2022.
Deere's automated weed killer has a self-propelled 120-foot carbon-fibre boom lined with 36 cameras that scan 2,100 square feet per second. Ten onboard vision-processing units handle almost 4 GB of data per second, using AI and deep learning to tell crops from weeds; when a weed is found, a command sprays it. The machine crosses a field at 12 mph without stopping, cheaper, faster and more reliable than manual labour. Fusing hardware and software with machinery has helped farmers cut herbicide use by more than two-thirds and raise productivity enormously.
Deere collects data from all its modern equipment as it works: billions of measurements on soil, crop and weather from about 500,000 machines on more than 325 million acres. The data flows into its cloud-based JDLink system, where analysis yields immediate and future improvements to machines and farms, and ML and AI turn it into a portfolio of combined digital and analog services for seed, fertilizer and weed management.
Until recently, incumbent makers of construction, mining and other heavy machinery did not put the most advanced software in their products. Now they use gen AI and machine learning to mine structured and unstructured data (text, high-resolution 3D images, voice, video and sound) and create complex designs in seconds. Success takes more than putting a computer in a tractor: firms must connect hardware and software, steel and silicon, humans and machines, and partner with other companies.
How fusion strategy differs from the internet of things
Industrial makers traditionally studied sales and marketing data: which demographic bought a product, who would pay more for a better one. Digital systems were a way to cut manufacturing costs or add features like Wi-Fi. Now they must focus on what happens after the sale: how digital and analog products together produce good customer outcomes, with data gathered constantly and insights applied in real time.
- Design. Like the internet of things, fusion uses physical objects with sensors, software and connectivity. But IoT equips an existing product to be monitored over the internet; fusion reimagines the product as if built from scratch with every digital tool available.
- Ownership. IoT is owned at the functional level with simple use cases: operations managers (supply-chain coordinators, quality-control executives) check that machine processes meet standards, designers use data for the next product version, and functions need not coordinate. Fusion is cross-functional, with senior executives from several departments deciding how sensor data can serve current customers and create new products.
- Speed. IoT records and logs data; fusion applies it immediately, observing use across settings to offer automated, personalized recommendations or experiences.
- Business case. Fusion is about competitive differentiation. Self-driving carmakers invest in IoT to meet safety standards, but differentiate through how sensors, software and connectivity are integrated to outperform rivals.
| Traditional strategy | Fusion strategy | |
|---|---|---|
| Growth dynamics | Linear, gradual, within industry boundaries | Nonlinear, exponential, across industries |
| Competitors | Familiar peers with similar models; product design and delivery | Digital entrants with new capabilities; product use |
| Scale and efficiency | Physical assets, production | Information assets, data |
| Expansion | Product-market extensions, vertical M&A | Virtual integration through data and partnerships |
| Customer insight | Ad hoc surveys, point of purchase; operational gains | Real-time observation of outcomes; differentiation |
| Data and AI | Independent databases, single firm, efficiency | Integrated real-time data across ecosystems; differentiation |
The authors stress that the table is not exhaustive: moving to fusion means rethinking how firms use data, partner and serve each customer.
The fusion framework
Opportunities from traditional and gen AI fall into four categories. Firms answer two questions: what is the reach of our product-in-use data, and how rich is it? Most should start with fusion products, which need the least data, then test the other three to learn how to create and capture value, how rivals' strategies threaten them, and how to allocate resources. The framework is dynamic: firms move into services and systems with experience, then toward solutions, which create the most value.
Fusion products
Fusion products are designed from scratch to collect and use product-in-use data in real time. AI can be applied to the growing data mass in three ways:
- Traditional AI analyzes it for product, process and cost improvements.
- Generative AI creates digital twins of products (virtual duplicates that mirror their functioning) and trains robots.
- Large language models develop proprietary insights that add value for customers.
Tesla
Tesla cars were designed as cloud-connected computers on wheels. Incumbents have added digital features for over three decades, GM's OnStar, Ford's Sync, Mercedes's Mbrace, but none let them track cars in real time, analyze data continuously and improve vehicles on the fly. Every Tesla's sensors report performance; AI judges whether it works optimally, and many fixes arrive as software: tweaking hydraulics to stop a door rattling, or adjusting regenerative braking over the air to reduce collision risk. Tesla is also testing gen AI to program autonomy: from video taken inside every Tesla, it reconstructs countless driving scenarios to train cars to avoid collisions.
Each car's operating system, drawing on data from hundreds of thousands of cars, integrates three digital twins: the product twin (the car as designed), the process twin (the assembly line that built it) and the performance twin (the car in use), plus insights from the supply chain, manufacturing and after-sales service.
Catch-up and other industries
Mercedes is designing AI-enhanced vehicle operating systems with Nvidia supplying software, hardware and AI chips. GM, with its subsidiary Cruise, put self-driving taxis on the road, withdrew them after safety concerns, and, the authors expected, would resume once it understood the lapses. Beyond cars, window maker View sells smart glass that uses real-time data and AI to let in more light when a room is dim and filter sunlight when it gets too hot.
- Premium prices. The most attractive option, as Tesla shows, justified when products perform better and cost less to run than rivals'.
- Performance contracts. Guarantee proactive maintenance and updates with minimal downtime. Rolls-Royce uses AI to monitor airline engines and guarantees uptime of close to 100%; it knows of failures in advance or in real time and troubleshoots far faster.
- Adjacent markets. Use product insights to enter new spaces: Tesla insurance undercuts rivals by pricing on each individual driver's recorded behaviour instead of clustering drivers into risk profiles.
Fusion services
Industrial services have been labour-intensive, relying on codified knowledge and standard protocols. Customized services delivered fast and at volume are more profitable but were costly in the analog world. Now, in place of an army of on-site technicians, AI-generated recommendations from product-in-use data can flow automatically to customers.
- New ventures. Weather and climate data from Weather.com and ClimateAi help farmers make precise, timely decisions; Norm, a ChatGPT app for agriculture, mines weather, soil and current-events data to answer farm-management questions.
- Incumbents. In civil aviation, a 1% gain in engine fuel efficiency could save $2 billion a year. Rolls-Royce's R² Data Labs saves customers $200,000 per aircraft annually by analyzing fuel-consumption patterns in product-in-use data (routes, altitudes, speeds, weather and loads) and relaying insights in real time.
Pricing can evolve in three stages: unbundled at first (buyers pay only for services they value), then bundled packages as appreciation grows, and eventually gain-sharing agreements that give the provider a share of the profit increase its services create.
Fusion systems
Many Deere customers, and those of other farm and construction equipment makers, run complex systems of machines from different specialist suppliers. Most manage interoperability themselves; some hire integrators and consultants. Fusion systems shift digitization from the micro level (improving each company's product separately) to the macro level (improving related products from many companies in concert for whole-system efficiency). A fusion integrator must connect all its own machines and interlink them with partners' and competitors', and, unlike a traditional integrator who just makes the system work, ensure it keeps improving as parts and functions are added.
This is where AI matters most. Most industrial firms handle structured data comfortably (names, addresses, card numbers). Advanced firms feed unstructured data, such as images of defects, the sound of faulty machines and video of autonomous processes, into gen AI in real time to build realistic digital twins of systems, then test permutations of products, machines, components and peripherals to find the best combination.
Dubai's Burj Khalifa, the world's tallest building, runs ventilation, air-conditioning, lighting, water, parking, storage, elevators, telecoms and security inside one facilities-management system. Honeywell collects real-time data from it and analyzes it for anomalies such as malfunctioning equipment, helping the building detect incidents earlier, react faster and reduce risk. Result: maintenance time for mechanical assets down 40% and their availability up to 99.95%.
Fusion systems will shift buyers' preferences from distinctive features to integrability. Makers therefore need a technology architecture that integrates hardware, software, applications, data and cloud connectivity and can absorb product-in-use data and algorithms from other products. Revenue comes from integration fees, charges for connecting additional machines, and fusion-systems software as a service sold by one-time licence, monthly subscription or pay-as-you-go.
Fusion solutions
Fusion solutions combine products, services and systems to directly improve a customer's performance. Instead of consultative sales treating each problem as a one-off, they are built from product-in-use insights and designed to apply to many customers. That demands being better than anyone at solving customers' problems, earning unusually high trust, and tying revenue and profit to customer success through outcome-based contracts and profit-sharing.
Creating the most value requires ecosystems of partners. Deere teamed up with agricultural software firm Granular, whose yield models combine satellite imagery on weather, elevation and soil with historical yield statistics to help Deere's customers forecast costs, revenues and profits. Selling solutions, Deere will face traditional rivals CNH Industrial and AGCO, but also software firms such as Climate Corporation (now part of Bayer) and digital giants IBM and Alphabet. The battle for customers will be fought between fusion ecosystems, not individual companies, and combining all four strategies yields the largest value.
Conclusion: what industrial leaders must change
Every industrial firm needs its own fusion strategy, and the first step is to stop treating human and machine intelligence as separate: every function improves when they work together, and firms that embrace this collaborative intelligence will beat those that treat AI as esoteric or delay it.
Industrial companies have competed through proprietary technology and vertical integration. Fusion requires partnering with digital natives and start-ups, which ultimately lowers costs, raises speed and improves customer satisfaction. It is a major shift that requires CEOs to lead the digital effort personally, a job they usually delegated, with a holistic strategy for how, when and where digital technology will drive the firm so that everyone shares the same priorities and resource plans. Every employee must embrace digital technology for the strategy to be crafted, executed and adopted.
Two predictions aged badly. The authors expected GM's Cruise robotaxis to return; GM instead ended Cruise's robotaxi programme in December 2024. View, the smart-glass exemplar, announced a Chapter 11 filing in April 2024. The article also ignores who owns product-in-use data, the farmer or the maker, and the Tesla examples are taken largely at the company's word.
The data-ownership question came to a head: in July 2026 Deere settled the FTC's right-to-repair suit (joined by five states), agreeing to give owners and independent shops dealer-equivalent diagnostic and software tools under a 10-year order.
A fusion system is an integration architecture problem: open APIs, shared data models and event streams across vendors. Banks see the flip side as lenders: equipment finance and insurance can be priced on live usage data, as Tesla does, if data rights are contractually clear.
AI Is Changing the Structure of Consulting Firms
David S. Duncan and Tyler Anderson, who lead the AI-native consultancy Disruptive Edge, and Jeffrey Saviano (Harvard's Edmond & Lily Safra Center for Ethics, formerly an EY partner) reject both extremes of the debate, that AI makes consultants obsolete or more indispensable. Consulting is being fundamentally reshaped. AI is automating the junior work that props up the classic pyramid, so firms must rebuild as a lean "obelisk" of fewer layers, smaller teams and more leverage at every level, or become irrelevant.
The authors' vantage point
For decades consulting ran on a stable pyramid: a wide base of juniors doing research, modelling and analysis under a narrow apex of senior leaders who guide strategy and manage clients. The pyramid powered the industry's economics and shaped its identity. Gen AI tools, predictive algorithms and synthetic research platforms now automate the tasks that filled junior consultants' weeks. At Disruptive Edge, Duncan and Anderson test AI across consulting tasks, from routine research to advanced analysis and synthesis, to speed delivery, sharpen insight and free people for judgment, creativity and deeper client engagement; Saviano adds an AI governance and ethics lens on how talent development and governance must change.
How the pyramid is crumbling
Demand is safe. Clients will keep needing external, specialized expertise; flexible capacity; credible validation; and an independent perspective on complex problems. What AI undermines is the delivery model: juniors spending weeks gathering and analyzing data, modelling scenarios and crafting slides to support seniors' recommendations. AI does all of that faster, cheaper and often better.
- McKinsey's Lilli. Used by more than 72% of the workforce; cuts research and synthesis time by about 30%.
- BCG's Deckster. Builds presentation decks in minutes.
- Bain's Sage. An AI copilot trained on Bain's internal IP.
- Deloitte's Zora AI and PwC's agent OS. Agentic platforms reshaping internal workflows and client offerings.
The automated work is the cornerstone of junior roles and is spreading into the middle tier. If AI absorbs the work that justified thousands of billable junior hours, the pyramid collapses under its own weight.
The rise of the consulting obelisk
The authors propose the obelisk: tall and narrow, with fewer layers, smaller teams and more leverage at every level, built around three human roles instead of sheer scale.
| Role | Who | Job |
|---|---|---|
| AI facilitators | Early-career consultants | Trained in the latest AI tools and data pipelines; design and refine AI workflows so teams generate insight fast. A new apprenticeship stressing technical fluency and applied judgment from day one |
| Engagement architects | Experienced project leads | Define the problems, interpret AI outputs with human judgment, turn them into actionable strategy, orchestrate the work and adapt as conditions change |
| Client leaders | Senior partners | Play the long game: deep, trusted relationships with senior executives, helping them make sense of change and stay ahead of disruption |
The roles map onto the three functions every firm needs: someone builds and runs the workflows that produce outputs, someone leads the work and turns it into recommendations, someone builds and keeps executive trust. Together they also form a talent pipeline that must survive automation. The point is to move away from scale for its own sake toward sharper thinking delivered faster with less overhead, reallocating human energy to insight, judgment and trusted partnership; the authors present this as a necessary evolution, beyond a cost response.
Pyramid. Wide base of juniors, narrow apex; value from headcount and leverage; juniors spend weeks on data, models and slides; quality assured by layers of review; paid by billable hours.
Obelisk. Few layers, small senior-heavy teams, AI doing foundational work; value from insight, speed and judgment; accountability embedded in team workflows; paid for strategic contribution and client outcomes.
A new wave of AI-native boutiques
The clearest examples come from AI-native boutiques, not the big firms.
- Monevate. A pricing-strategy specialist that combines deep expertise with AI-enabled playbooks and modelling tools, advising without a traditional analyst layer.
- SIB. A cost-reduction firm whose AI agents scan invoices and vendor contracts for savings, calling in human experts only when needed.
- Unity Advisory. Launched by former Big Four partners with $300 million of private capital; conflict-free (no audit-advisory entanglements) and AI-native by design. It uses agile pods of senior consultants working with proprietary AI tools, hires no large analyst cohorts and has no hierarchical middle management: in the authors' words, a textbook obelisk.
- Disruptive Edge. Instead of juniors spending weeks building basic knowledge, engagements start with AI-powered deep-research reports; the app-building platform Lovable takes teams from concept to a working prototype in under two weeks instead of months, so assignments run with smaller, more senior teams and more client time.
The authors concede that research on traditional firms confirms AI can do much junior work, often faster and better, but because the obelisk is new, long-term evidence is still to emerge. Their bet: once foundational tasks are automated or accelerated, senior-heavy teams can concentrate where human expertise adds most value.
Why firms will resist change
Clayton Christensen's The Innovator's Dilemma explains the obstacle: incumbents rarely disrupt themselves while the existing model is still printing money, and in consulting that is the highly profitable pyramid.
- Wiring. Promotions, compensation, staffing and even the mental model of "good consulting" are built around headcount and leverage, with strong incentives to keep large junior teams.
- Siloed labs. Firms invest in AI tools and announce innovation labs, but these stay separate from core delivery; flashy demos impress clients while the engine of large junior teams stays untouched. A leaner structure feels like an existential threat to firms built on scale.
- Talent mismatch. Firms recruit and train generalist MBAs by the hundreds; the future needs smaller cohorts fluent in AI tools, data workflows and systems thinking. Upskilling exists (PwC committed $1 billion to AI training), but culture and incentives lag, still favouring hours billed over insight delivered.
The common root is treating AI as a tool to bolt on instead of a reason to re-architect from first principles. That may protect margins short term, but it lets smaller AI-native firms move faster, run leaner and give clients more value at lower cost.
Implications for consulting firms
Firms that stay junior-intensive risk becoming slower, costlier and less relevant; decisive movers can become leaner, more expert-driven and more valuable. For incumbents the transition means three changes:
- Redesign workflows around AI to redefine how value is created and delivered.
- Rework junior development to combine AI facilitation skills with classic consulting skills: problem-solving, communication and client management.
- Change compensation to reward strategic contribution and client outcomes over billable hours.
Governance and ethics in a flatter firm
Traditional deliverables pass through layers of review by analysts, senior consultants, managers and partners, which makes it easier to catch problems and assign responsibility. In an obelisk, small teams move fast and AI plays a larger role in decisions, so new approaches must keep AI-powered decisions understandable, equitable and owned by clearly accountable people. Saviano's Safra Center team, which develops new AI governance models, argues that business leaders should govern AI themselves rather than wait for regulation, building ethical guardrails into how AI is used instead of relying only on central compliance teams or after-the-fact reviews. In the obelisk, accountability must be clear, distributed and embedded in team workflows, because small expert teams wield tools that can shape high-stakes decisions.
The authors close that this is no moment for incrementalism: the winners will move first and reimagine the industry before someone else does it to them.
Two of the authors sell the model they describe, and they concede long-term evidence for the obelisk "is still to emerge." The piece never resolves its own paradox: if the junior base shrinks, where do future engagement architects and client leaders learn the craft? The AI-facilitator role is offered as the answer, but it trains tool fluency more than client judgment.
By December 2025 McKinsey, BCG and Bain had held graduate pay flat for a third year, and Big Four leaders predicted UK graduate intake could halve; Accenture cut over 11,000 roles in one 2025 quarter. McKinsey still planned 12% more graduate hires for 2026.
Procurement teams can use the obelisk as a buying lens: ask suppliers for outcome-based fees, the AI tooling behind each deliverable and the named accountable reviewer, and refuse rate cards that still price large junior teams whose work is now automated.
Stop Running So Many AI Pilots
Goutam Challagalla (Dentsu Group Chair in Sustainable Strategy and Marketing at IMD), Mahwesh Khan (senior adviser at IMD) and Fabrice Beaulieu (former global CMO of Reckitt) attack the habit of green-lighting every gen AI use case that promises a positive ROI. Lasting advantage comes from going deep and narrow: transforming one function or end-to-end process intensively, where it reinforces an existing strength, instead of scattering pilots across the firm. They give a four-step method and cases from Reckitt, a disguised bank, L'Oréal and IKEA (HBR, November-December 2025).
Reckitt's choice
In late 2023 the management team at consumer goods group Reckitt surveyed possible gen AI uses and found them everywhere: drafting presentations, customer support, optimizing procurement contracts. Many promised guaranteed time savings and immediate ROI, but they touched unrelated tasks. Executives liked the savings and saw that none would change the company's strategy or create a real edge; they wanted something bigger than marginal efficiency.
So rather than approve every project with a positive return, they picked one domain: marketing. There gen AI could serve many interlinked tasks (insight generation, content generation, new product development) that draw on the same data, customers and market research, so the output of one feeds the others. The authors trace the chain: real-time consumer insight improves innovation and segmentation; faster, better innovation puts relevant products on shelves; relevant products build loyalty, retention and advocacy. Concentrating in one place, the executives reasoned, would let the firm improve at scale.
Marketing was also where Reckitt was already strong, with rich customer data and a team fluent in advanced tools, predictive AI included. Working gen AI into the whole operation, not "the odd task or two", forced the team to recast many processes. Less than two years later Reckitt credits the program with generating product concepts up to 60% faster and making brand and marketing communication processes 30% or more more efficient, depending on the process. Those gains, the authors insist, depended on concentrating investment in a cluster of connected tasks within one domain.
The two mistakes
The lesson the authors draw is that firms seeking durable advantage must avoid two related errors:
- Scattering effort. Spreading resources across one-off use cases throughout the company.
- Worshipping immediate ROI. Justifying each project only by its short-term return.
Together these make a "shallow and broad" deployment: many unrelated use cases, each backed by a quick-payback case. Many executives told the authors they should deploy gen AI as widely as possible to see what works. That yields some cost savings but rarely a competitive advantage.
The mechanism is simple. Suppose a firm automates a few processes each in accounting, supply chain, marketing, finance and HR. First, when only 5% to 10% of a function's tasks use gen AI, no executive has a reason to redesign the function. Second, use cases with no larger objective breed initiative fatigue, and when the gains are small or easy for rivals to copy, employees turn skeptical. Reckitt's pilots, by contrast, served one goal: top-line growth through faster product launches.
- 4% of companies concentrate on a few priorities and transform them deeply (BCG survey, 2024, 1,000 CXOs and senior executives).
- Those focused firms earn twice the ROI over time.
- When only 5-10% of a function's tasks use gen AI, nobody is motivated to redesign the function.
- Reckitt: 60% faster product concepts and 30%+ more efficient brand processes in under two years.
The authors write as practitioners, educators and consultants to global brands, and say that since gen AI reached the market in 2022 they have seen three advantages of deep and narrow: it forces a fundamental rethink of how work is done in one part of the business; it raises productivity by building on existing strengths; and it produces ways of working tied to the firm's own culture, which rivals find hard to copy. Once a firm masters one area, it can carry the lessons to others.
Why executives resisted
When the authors first pitched the idea, many leaders found it counterintuitive and riskier than spreading bets. The authors concede that shallow and broad is not wrong: it saves time and money, and a few such experiments are sensible early on to test the organization's willingness and readiness. But disconnected use cases are not transformation. The objections they heard:
- Investment. Rebuilding processes requires heavy spending.
- Conventional wisdom. The received view says you need hundreds of AI projects to succeed.
- Self-assessment. Choosing where to go deep demands an honest, intense appraisal of core competencies and data maturity, which many had skipped as too hard and slow.
Over time, the authors report, many skeptics came round.
Shallow and broad. Many unrelated pilots, each justified by quick ROI. Saves time and money and tests readiness, but the gains are limited, temporary and easy to copy, and the effort breeds fatigue.
Deep and narrow. One function or end-to-end process reimagined. Slower and managerially intensive, but it builds capabilities rivals cannot replicate, and the learning then spreads to other domains.
Why go deep and narrow
Leaders told the authors they want concrete ROI projections for every use case and will expand only once returns are proven. That ignores how adoption works: ROI accrues at different speeds and magnitudes for each use case, impact at effective adopters can take time, and short-term ROI alone is the wrong yardstick.
Deep and narrow is also hard. It needs extensive change management, which is why doing it in several domains at once is exceptionally difficult: new technology for employees, changed work processes, new team structures and ways of working. Any new tool requires some adaptation; going deep requires wholesale reimagining of work.
The focus can be a single function (L'Oréal and Reckitt chose marketing) or an end-to-end process, as at the bank below.
A regional lender (disguised as Acme Bank) had a cumbersome three-stage process: a mortgage adviser collected documents and keyed the data into the loan origination system; a loan officer verified them and added more (title searches, flood certifications); an underwriter judged creditworthiness and set terms. Acme now has gen AI process all mortgage paperwork, and underwriters use a conversational interface that can run predictive models and draft the explanation for an approval or rejection. Because loan officers could easily probe the model's decisions, they learned that feeding it better customer context improved outcomes for customers and bank alike. Result: better decisions, more speed, quicker response to customer concerns.
How to begin: the four-step process
The aim is for gen AI to strengthen existing competitive advantages, so you must first decide where it is most strategic.
- Identify the most promising opportunity. Place possible uses on a spectrum from office productivity through domain reinvention to value generation, and favor domain reinvention.
- Identify areas of lasting advantage. Protect or enhance an existing strength, or create a new one that is hard to copy.
- Choose the right sequence. Start with cost efficiency inside the chosen domain, which pays faster, then move to growth.
- Monitor the competitive field. Ask whether a rival could use gen AI to offer a "good enough" version of your strength, simpler and cheaper, and widen the gap.
Step 1: The opportunity spectrum
| Zone | Examples | Advantage potential |
|---|---|---|
| Office productivity | Meeting summaries, slide drafting | Efficiency on mundane tasks; unlikely to give an edge |
| Domain reinvention | End-to-end redesign of a function or process | Open to almost any firm |
| Value generation | New business models; AI infused into products and customer experience | Exciting but hard where AI can't be built in (toothpaste) |
Acme Bank illustrates the middle zone. It is one of the largest lenders in its region; mortgages are a large part of the portfolio, carry big margins even when rates are low, and anchor relationships through which it cross-sells other products. The team mapped the essential steps, asked where to work differently, and eliminated redundancies first; gen AI then made the better process faster still.
Step 2: Lasting advantage, created or amplified
L'Oréal created a new advantage. The €43 billion French cosmetics group (the world's largest) combined its beauty expertise with gen AI to reinvent customer interaction. CEO Nicolas Hieronimus describes a shift from "beauty for all" to "beauty for each", putting personalization at the center, and the gen AI effort targeted the consumer journey. Its Beauty Genius chatbot diagnoses each user's skin, then gives expert analysis, product picks and personalized routines. That works because L'Oréal has built augmented reality capability since 2015 and can draw on its skin-biology research and customer data. The journey now runs diagnosis, advice, product selection and sharing with friends, and each conversation hands L'Oréal more data.
- Usage. Beauty Genius held 400,000+ US conversations in its first six months, and is being integrated into WhatsApp.
- Platform. It sits inside a wider set of Beauty Tech services, such as virtual try-ons and Hapta, a computerized makeup applicator, used 110 million+ times.
- Conversion. In Southeast Asia, the Middle East and North Africa, Beauty Tech doubled conversion among 50 million users.
- Moat. Copycats will appear, but they lack L'Oréal's cosmetics research base.
IKEA amplified an existing strength. IKEA sells interior design services, but price deters customers because skilled designers are expensive. From early results, chief data and analytics officer Francesco Marzoni projects that its designers will use gen AI to produce a design in 10 minutes, then refine it with their expertise, against outside consultants charging about $99 per room and taking several days. IKEA already has the world's largest community of interior designers and runs far more projects than rivals, so it can feed its models real-time, high-quality data from thousands of jobs. Expertise plus scale lets its system learn faster, and rivals will struggle to match the quality and price.
Step 3: Efficiency first, then growth
Gen AI lifts profit through cost efficiency and growth. Inside one domain, starting with efficiency is usually smarter because costs fall faster than revenue rises. Reckitt's sequence:
- Prove it. Five pilots showed quick marketing efficiencies; one produced media campaign analyses in hours instead of days, which won executive support.
- Map the work. Through surveys and one-on-one interviews, Reckitt catalogued the work of its 2,000+ global marketers: 300 distinct tasks, each with a time estimate from the marketers.
- Prioritize. It plotted every task on two axes, degree of possible automation (by type of gen AI applicable) and opportunity size (time spent), and targeted about 100 tasks.
- Redesign roles. With so much work now AI-assisted, the department had to be rethought; as AI absorbed routine work, marketers took on harder tasks, including some previously done by their managers.
- Then grow. Once efficiency hit the bottom line, executives approved gen AI for slower growth goals: turning consumer insight into better, faster innovation and shorter time to market.
Deep and narrow touches daily work far more than scattered pilots, so it is more managerially intensive. The authors' split: change management is 70% of the challenge, perfecting data 20%, and using gen AI itself 10%. Expanding beyond the first domain means a full-scale change program.
Step 4: Watch competitors
Rivals are using gen AI too. The test question: can our top competitor replicate one of our valuable strengths with gen AI? It need not match you; a "good enough" solution that is simpler and cheaper is enough to hurt. The defense is to deploy gen AI so that it widens the gap, as Beauty Genius does, and to seize the new ways of working that focused deployment reveals.
The authors' conclusion
It is early for gen AI and agentic AI, but a pattern is visible: most firms go shallow and broad and hope a few pilots pay off. Many do earn a return, but the combined impact is limited and temporary because such deployments are easy to copy. Deep and narrow reinforces strengths deliberately; IKEA, L'Oréal and Reckitt asked hard questions about their AI goals, and some ended up with new lines of business. Citing BCG's double ROI finding again, the authors predict the gap will keep widening.
Beaulieu ran Reckitt's marketing, so the lead case is insider testimony, and "Acme Bank" is anonymous. The 70/20/10 split is offered without evidence, and BCG's "twice the ROI" is a consultancy survey figure, not an audited outcome. L'Oréal's and IKEA's numbers are company claims or projections (IKEA's 10 minutes is a forecast from early results). The article also gives little guidance on what to do if the chosen domain fails.
BCG's later data supports the thesis. Its September 2025 study found only 5% of firms "future-built" for AI and 60% reporting no material value; its Applied AI Index (September 2026, 1,330 leaders) found leaders that reshape workflows end to end posting 2.4x the revenue growth of laggards.
The Acme case is the template for banks: pick a value stream (mortgage, trade finance, KYC onboarding), not a department, remove redundant steps first, then redesign it end to end. Architects should model it as a capability map, so each AI component plugs into shared document, decision and audit services and the next value stream reuses them.
Addressing Gen AI's Quality-Control Problem
Stefan Thomke (Harvard Business School), Philipp Eisenhauer (Amazon senior economist) and Puneet Sahni (Amazon senior principal engineer) show how Amazon tamed gen AI's habit of inventing facts, leaving things out and producing more options than anyone can evaluate. Amazon's Catalog AI treats gen AI output like a factory process: audit a baseline, set control limits, let AI check AI, A/B-test everything that passes, and feed every result back so the system learns. Human review and stand-alone testing tools, the usual fixes, are expensive and see only a fraction of the output (HBR, September-October 2025).
The problem and Amazon's answer
Enthusiasm for gen AI runs into one hurdle that limits adoption: the technology makes things up, omits things, and generates so many possibilities that nobody can tell which will work. Most companies respond with human reviewers and stand-alone testing tools or services. Both cost a lot, and both can handle only a sliver of what gen AI produces.
Amazon built something different for its product catalog. Catalog AI is itself a gen AI system that does three jobs at once:
- Detects and blocks unreliable data before it reaches a product page.
- Generates and tests ideas for new product pages, measuring whether they actually sell.
- Improves itself automatically, using the feedback from its quality checks and experiments.
The scale is the point: it can create and test tens of millions of hypotheses a year, against the thousands a human-based system manages. While many firms struggle to show any financial return from AI, Catalog AI is already producing measurable value: at the time of writing, 8% of its suggestions had raised sales revenue.
Thomke notes that he knows businesses whose online experiments succeed more often, 10% to 20% of the time. But those firms relied on people to generate hypotheses and on much less automated testing, so they ran far fewer experiments in total. Amazon's system also gets better on its own, so its yield rate should rise over time. Amazon calls the system, launched in 2023, a work in progress; the authors judge it mature enough for other managers to learn from.
The deficiencies of the traditional approach
Amazon's catalog holds hundreds of millions of products sold worldwide, and millions of listings are added or edited every day. The data on each page (images, titles, descriptions, recommendations) has to be complete, accurate and appealing so shoppers find what they want fast, and returning shoppers expect a familiar layout where images, titles, descriptions and buy buttons are easy to find and load quickly. The stakes are concrete: when product information is incomplete, irrelevant or wrong, customers abandon the purchase or return the product, and Amazon loses both money and trust.
Quality used to rest on two pillars:
- Thousands of listing specialists. They added and removed information, spotted errors, consolidated data, translated text and folded in third-party data.
- Hundreds of narrow ML models. Each was tuned to one product category (shirts, TVs) and one layout component (titles, descriptions).
Those classic ML models suit small, structured datasets, and scaling them is expensive: a model trained on shirts cannot economically be reused for TVs or anything else. LLMs, trained on huge datasets, work across categories. Replacing the ML models with LLMs let Amazon simplify its technology (fewer models), simplify its organization (fewer functional specialists) and cut costs.
Ensuring reliability: the four steps
The new approach started badly. For the first few weeks about 80% of Catalog AI's results were unreliable: invented facts, missing information, or advice customers did not care about. Two examples:
- The phantom horsepower. It described an electric pump as 15 horsepower when no such figure existed in the data.
- The wrong sofa material. Asked about a sofa's material, it described the frame, when shoppers mostly wanted to know about the seat fabric.
To fix quality and to test whether changes actually help, Amazon took four steps.
1. Conduct an audit
Progress cannot be tracked without a baseline. Manufacturing gets one by observing a process during a stable period and setting control limits from that data. Amazon's equivalent: it had the LLM generate thousands of product pages that were already known. Human auditors compared the AI pages with the known facts, scored their reliability and dug into the root causes of poor results. That analysis produced a series of quick improvements, which became the guardrails below.
2. Deploy guardrails
A hallucination is a false output presented as true, and it typically arises when a model reaches a conclusion not grounded in its input data. The obvious cure is to confine the LLM to company-specific data and keep it off the general web. But that carries a trade-off: the more external and internal data the model may draw on, the more new ideas it can explore, modify and test. From the web, for instance, it can infer that paper plates are not dishwasher safe, a useful fact no internal record may contain. Over-constraining inputs kills such inferences. So Amazon left the inputs open and installed three other kinds of constraint instead.
- Simple rules. Reject content that breaks a rule: a weight must carry units (kilograms, pounds); suggestions that make immaterial changes to the current listing (style "contemporary" to "modern") are rejected; layout rules keep the customer experience consistent across the site.
- Statistical profiles. The gen AI version of statistical process control (SPC) limits: data on existing products defines the expected range for a new output. If an LLM writes up a third-party vendor's table and the result falls outside what typical tables in Amazon's store look like, a second LLM questions it; asked to explain itself, the first model often spots its own mistake.
- AI checking AI. No set of rules covers every output, so a second gen AI system handles the rest. A content generator trained to propose hypotheses is paired with a content reviewer trained to check them, and the two converse automatically, drawing on different background knowledge.
The reviewer works in several ways. It hunts for inconsistencies across a page, such as a colour named in the title that does not match the image, and if it finds one, the change is automatically blocked. It can interrogate the generator ("Why is the new product page better than the existing one?"), forcing it to re-examine its output and sometimes drop the suggestion. For extra rigor, Catalog AI connects to internal and external LLMs trained on different data, because differently trained models find different problems: if the generator makes a reasoning slip and miscalculates a product's volume, a reviewer trained on another dataset is likely to catch and block it.
Only a hypothesis that clears every quality check moves on to the experimentation platform, where the question changes from "is it reliable?" to "does it raise sales revenue or units sold?"
3. Test the effectiveness
Gen AI produces far more ideas than anyone can judge by eye, so firms need efficient ways to find the effective ones. Amazon's old method had catalog specialists write rules and algorithms that automatically approved and improved the layouts they believed worked best. Three weaknesses:
- Untested assumptions were baked into the rules and algorithms.
- Testing was patchy: not always cost-effective, not always automated.
- Market research misleads: focus groups and surveys capture what customers say, which often differs from what they do.
A Booking.com executive put it bluntly to Thomke: people are terrible at guessing, and the firm's predictions of customer behaviour are wrong nine times out of 10.
So the Catalog AI team built A/B testing into the workflow itself. Testing gen AI's volume of output scientifically needs a real experimentation infrastructure: instrumentation (recording clicks, mouse hovers, event times), data pipelines and data scientists. Third-party tools make it easy to try a few experiments, but scaling requires tight integration into the workflow. At Amazon it is fully automated: every page change Catalog AI proposes is A/B-tested.
The mechanics are classic. The champion ("A", the control) is the current listing; the challenger ("B") is an AI-generated variant built for a distinct aim such as higher conversion. Users are randomly assigned to each, and composite metrics are computed and compared. Those metrics should align with strategy, and the best short-term metrics are the ones that predict long-term outcomes.
Skincare, short beats long. The champion description walked through a long list of features and how each delivered benefits (unclogging pores, better skin texture, less redness). Catalog AI's challenger simply named the primary benefits: smoother, hydrated skin and fewer wrinkles. The shorter version raised sales significantly in a randomly selected customer group.
Body wash, cutting the benefit hurt. Other tests showed the LLM dropping benefits from titles. Rewriting "Foaming Body Wash Mango Butter for Glowing Skin, 13.5 Ounce (Pack of 4)" as "Mango Butter Foaming Body Wash, 13.5 Ounce, 4 Pack" removed "glowing skin", and sales fell.
The overall verdict is sobering. Of AI content that passes the reliability checks, about 40% either improves primary metrics (such as shopper conversion) or has no effect either way; the other 60% produces significantly negative results and is normally not published. Even after every reliability check, most gen AI content would have hurt sales; only testing on real customers reveals which ideas actually work.
4. Create a learning system
The ideal quality system keeps improving with little or no human involvement. Amazon's generates data that sharpens its LLM, making it better at challenging assumptions about what customers like. Yet the scientists found it still pays to bring humans in occasionally for better learning data: investigations of negative experiments sometimes expose flaws in the model. One team found the LLM wrote "no warranty" by default whenever the warranty field was empty. As the system matures, human work will shrink to system design and infrastructure decisions. The authors list the components a learning system needs.
Customer model. A composite metric that models customer preferences speeds feedback enormously. Strictly it is optional, but if accurate it lets a company run virtual experiments with near-instant results, the way carmakers run crash simulations on computers instead of destroying physical prototypes, because real customers are no longer needed. Building a metric that responds like a typical customer takes care: the firm must choose which data to include and validate it through many live experiments. Catalog AI pairs live tests with its customer model, the Catalog Data Quality (CDQ) metric, which blends signals from the rules, statistical profiles and reliability checks. Eventually CDQ will replace most tests on real customers, speeding learning further.
Multivariate experiments. Beyond A/B tests, Amazon runs multivariate experiments that show how several variables (text, colour, images) interact and which design choices win. Algorithms spot patterns and probe them on high-traffic pages where effects on primary metrics are statistically significant. The strongest results, more than 1% positive impact on some primary metrics and less than 1% negative impact on others, go to data scientists to check for patterns and errors. At Amazon's volume, even tiny gains across millions of listings can mean billions of dollars in extra revenue; firms with less traffic will need different thresholds.
People also feed learning through experiments that isolate one effect and produce stories the Catalog AI team can share with data scientists. The Catalog AI Titles team, for example, prompts the LLM for titles of different lengths and content for the same product, tests them, and uses the winner as training data.
| Version | Length | Content of the protein powder title |
|---|---|---|
| Short | 75 chars | Brand, product, vanilla, 22g protein, 20 servings |
| Medium | 120 chars | Adds non-GMO, vegan, free of gluten, dairy and soy |
| Long | 180 chars | Adds no artificial flavours, synthetic colouring, preservatives or additives |
Better training data across categories has transformed output quality: from 80% unreliable at launch, 80% of Catalog AI's output now passes the initial quality checks.
Concept tests. At launch Catalog AI produced millions of listing changes to test on customers. Ideally each would get its own experiment, but many listings lack the traffic to reach the needed sample size (the smaller the expected effect, the more observations it takes to separate it from noise). So below one million views, the platform batches similar listings by category into single-concept tests, sometimes pooling content from thousands of AI-generated listings; above one million views, a stand-alone experiment runs. Batches are kept as small as statistical requirements allow. AI might generate thousands of description variants for coffeemakers and test them as one concept, which lets the system challenge assumptions about customer preference that nobody had ever tested.
Specialists had long assumed, and built into their ML models, that shoppers prefer a stark white background contrasting with the product. A concept test batched thousands of AI suggestions that showed products in the shopper's own environment. The control showed a model in a black thermal top against white; the challenger placed the same model in an indoor setting to help shoppers picture everyday use. The enriched background raised sales, and the change was rolled out to hundreds of basic fashion products.
Increase the effectiveness of AI projects
Once a quality system exists, it does more than police output: it can steer resources toward the AI projects with the highest ROI. Normally business groups submit overly optimistic return estimates, and those estimates decide who gets specialists and compute. The measurement gap is wide: in a 2024 Deloitte survey of 2,770 business leaders in six industries, 41% said their organizations struggle to define and measure the impact of their gen AI work. When the quality system instead tests new projects with customers, leaders can allocate on hard evidence and revise allocations more often as tests come in.
- Learning from prototypes. A firm can quantify a project's effect on subsets of customers before a broad rollout. Catalog AI improvements often begin as competing prototypes; teams use experimental results to argue for resources, and winners get scaled.
- Gradual rollout. Project Amelia, Amazon's gen AI assistant for sellers, is being opened to more sellers step by step while its performance is monitored and customer feedback folded in.
- Infrastructure trade-offs. Bigger LLMs may add only marginal benefit and take very long to train. Amazon's experiments found some smaller models outperform larger ones: they need fewer resources, can be retrained more often and respond faster to customers. Other firms can run similar tests to decide between in-house LLMs and third-party models.
The authors close on David Garvin's HBR article "Quality on the Line" (September-October 1983). Studying makers of room air conditioners, Garvin found the best producers' failure rates were 500 to 1,000 times lower than the worst, and traced the gap to a comprehensive quality-control system. The same lesson, they argue, holds for gen AI.
- Hundreds of millions of products; millions of listings added or edited daily.
- Tens of millions of hypotheses tested a year versus thousands for human systems.
- 80% of outputs unreliable in the first weeks; now 80% pass initial quality checks.
- Of checked content, 40% helps or is neutral; 60% is significantly negative.
- 8% of suggestions have raised revenue (human-hypothesis firms: 10-20%, on far fewer tests).
- Concept-test threshold: 1 million views; multivariate review threshold: >1% gain, <1% loss.
Two of the three authors are Amazon employees, so this is a company describing its own system, with no independent check on the 8% or 80% figures. The method also leans on traffic few firms have: concept tests, multivariate thresholds of 1% and an eventual customer-model metric all assume millions of shoppers. The 8% versus 10-20% comparison mixes different denominators, and the 60% harmful rate among checked outputs is itself a warning about what unchecked gen AI does when published straight to customers.
Unchecked output now carries a visible price. In October 2025 Deloitte agreed to partially refund an A$440,000 Australian government report after fabricated citations and a false court quote, traced to gen AI drafting, were found in it.
Banks generating product disclosures, client letters or KYC summaries with gen AI can copy the pattern without Amazon's traffic: audit against known-good documents first, encode hard rules (every rate has a unit and a date), set statistical ranges from historic documents, and run a differently trained reviewer model before anything reaches a customer. Where volumes are low, batch similar documents into concept tests rather than skipping measurement.
AI's Growing Waste Problem, and How to Solve It
Samsurin Welch (Cambridge Judge Business School's Circular Economy Centre; COO of HyveGeo), Khaled Soufani (Cambridge Judge) and Eva Morales (HyveGeo) take apart the claim that AI will pay for its footprint by solving climate problems. AI does help optimize renewables, forecast weather and discover materials, but its physical infrastructure may cancel much of that benefit. Energy and water are only the visible part of the bill: servers, chips, steel, concrete, land and e-waste mean AI has to be circular by design, managed with the 4-Use framework of use longer, use again, use differently and use less. (Adapted from hbr.org, 28 February 2025.)
The sustainability paradox
The pitch for heavy AI investment, especially in resource-hungry large language models, often includes a green promise: AI can optimize renewable energy systems, predict weather and discover new materials. The authors call this a paradox, because the same technology's environmental impact could negate much of the benefit.
Most attention goes to the energy and water consumed by training and inference. The authors say the growing resource intensity, waste generation and pressure on natural systems matter just as much, because behind AI's digital surface sits a mass of hardware: cutting-edge servers and chips, networking gear, storage, control systems, racks and cables, all inside large data centre buildings with their own power, cooling and ancillary plant.
That hardware has impacts well beyond electricity:
- Embodied carbon. An analysis of a Dell server found operational energy accounts for only about half of its life-cycle emissions; 40% to 50% comes from manufacturing and materials.
- Biodiversity at the site. Construction of a planned Meta AI data centre in the United States was halted after a rare bee species was found on the land.
- Regulation. The UK's new Biodiversity Net Gain law signals tougher scrutiny of how infrastructure projects affect local ecosystems.
For AI to have a net positive effect it must be sustainable by design. Providers and users alike face pressure on greenhouse gas emissions (including Scope 3, the indirect ones), biodiversity and human rights in supply chains. Microsoft reported a 30.9% increase in Scope 3 emissions, driven largely by embodied emissions in data centre construction materials and computing hardware. The authors insist leaders treat this as more than compliance: it touches reputation, operating costs and long-term competitiveness.
AI infrastructure's growing resource burden
Global data centre capacity is expected to double within five years, much of it driven by AI, and the consequences run far past the power bill. The authors work through four pressures.
Critical materials
Semiconductor manufacturing alone uses more than 300 materials, including critical minerals such as gold, copper, nickel and palladium. Those same minerals are needed for decarbonization technologies, notably wind turbines and batteries, and global supply is concentrated in a few regions, which sets AI and the energy transition competing for finite resources.
Electronic waste
Gen AI alone could add 1.2 to 5 million tons of e-waste by 2030, the equivalent of 495 Eiffel Towers, on top of the 62 million tons the world already produces. E-waste carries lead, arsenic and mercury. Only 22% is formally recycled; much of the rest ends up in developing economies in Africa, Southeast Asia and Central and South America, where informal sectors strip out copper and gold by burning or acid leaching. Those methods release toxic pollutants that damage ecosystems and health, with pregnant women and children most at risk; the ILO estimated in 2021 that about 16.5 million children work in industrial sectors, informal e-waste processing included.
Steel and concrete
Some of the world's largest facilities cover several million square feet. Data centres use far more steel and concrete than ordinary commercial buildings, because high-density racks are heavy, and both materials are notoriously hard to decarbonize. The AI boom multiplies the build-out.
Land and water
Poor land-use planning adds biodiversity loss, habitat destruction, water stress and heat dissipation. The authors' example is Meta's proposed data centre at Talavera de la Reina, Spain, on former farmland, projected to use more than 665 million litres of water a year in a drought-prone area. The Tagus river basin that supplies the region already suffers recurrent shortages linked to climate change. The case, they argue, exposes a wider clash between technological growth, environmental protection and the socioeconomic rights of farmers and local communities, and as water scarcity worsens, settling such competing claims fairly becomes central to sustainable development.
- 2x global data centre capacity expected within five years.
- 40-50% of a server's life-cycle emissions come from manufacturing and materials (Dell analysis).
- 30.9% rise in Microsoft's Scope 3 emissions, mainly embodied carbon.
- 300+ materials in chipmaking; gold, copper, nickel and palladium compete with turbines and batteries.
- 1.2-5 million tons of extra e-waste from gen AI by 2030, on top of 62 million tons a year; only 22% formally recycled.
- 16.5 million children in industrial work worldwide (ILO, 2021).
- 665 million litres of water a year projected for Meta's Talavera de la Reina site.
How a circular economy could help
The authors' remedy is the circular economy: rethink resource use, extend hardware lifespans and build robust recycling systems. Done well, they argue, this does more than shrink the footprint; it also drives innovation and competitiveness. They apply their own earlier 4-Use framework to AI's equipment, data centres and the relationships between AI providers and buyers.
- Use longer. Modular designs so components or buildings can be repaired and upgraded; repair programs using predictive maintenance to extend equipment life; repurposing buildings and infrastructure for new builds; reselling or donating kit for use elsewhere.
- Use again. Refurbish used equipment for reuse or resale; recycle and recover valuable materials to replace virgin ones; turn waste streams such as heat into inputs for other uses; recycle and repurpose water.
- Use differently. Switch to sustainable, regenerative materials such as cross-laminated timber; design for net positive impact; embed biomimicry; use equipment-as-a-service contracts.
- Use less. Cut virgin material use across the equipment's life; use predictive maintenance to keep equipment running at its best.
Circular computing equipment
Much of the resource and e-waste pressure comes from linear use of servers and storage: buy, use, discard. Lifespans keep shrinking under the push for more performance and faster upgrade cycles, so circularity has to start with how devices are designed, made and maintained.
- Use longer. Modular systems let components be repaired, replaced or upgraded instead of scrapping the whole device; even cables have individually upgradable endpoint modules.
- Use less. Modularity makes the initial design more complex but reduces manufacturing across the life cycle, cutting raw material use by 40% and embodied energy by 66%.
- Use differently. A goal to cut virgin plastic 50% by 2025 puts postconsumer recycled plastic into the Webex Room Bar (55% recycled) and UCS X-Series (77%); suppliers help reclaim materials from decommissioned hardware, and packaging moves from plastic to fibre.
- Use again. The Refresh program refurbishes and resells used equipment. The "Send IT Back" app and channel partners such as WWT run the reverse logistics that bring products back for reuse, refurbishment, recycling or safe disposal.
The lesson the authors draw: smart design, efficient materials and strong end-of-life strategies can move the industry from a linear model to one that wrings value from resources at every stage.
Circular data centres
Buildings produce 39% of global emissions, 11% from materials and construction, and data centres' structural demands make that worse. The authors give examples for three of the four uses.
- Use differently: materials. Microsoft built two Virginia data centres in cross-laminated timber (CLT), cutting embodied carbon by up to 65% against concrete. Sustainably sourced CLT is a realistic low-carbon option given the high cost and scarcity of green concrete and steel. Microsoft also writes low-carbon material requirements into supplier contracts and presses key suppliers to reach 100% carbon-free electricity by 2030. Biomimicry-inspired designs such as passive cooling cut impact further.
- Use again: structures and heat. Repurposing existing infrastructure can do even more. Start Campus in Portugal reused a decommissioned coal plant's infrastructure to build a seawater-cooling system for its 1.2-gigawatt SINES data centre. Sweden's EcoDataCenter, another CLT build, turns its waste heat into a resource for fish farming and greenhouses, working with the circular industrial developer WA3RM.
- Use longer: modular construction. Prefabricated components in renewable wood allow faster assembly, less waste, and later expansion or repurposing of the structure.
Avoiding harm is only the start: the next frontier is net-positive impact on the local environment. Microsoft's regenerative data centre concept targets biodiversity enhancement, water stewardship and ecosystem integration. The key learning is to adapt to local context (materials, ecosystems, infrastructure, community needs); Microsoft writes global design playbooks for different biomes and lets each site localize them.
Things to consider
Circularity, the authors conclude, must be a priority across the whole AI stack, and each party has its own job.
- AI service providers. Circularity is becoming a competitive differentiator, because customers want suppliers that fit their own climate commitments; it helps win and keep business.
- AI buyers. Sustainability cannot be an afterthought: engage suppliers on their circularity road maps and work together to cut waste, emissions and virgin-resource use, building more resilient value chains.
- Providers of circular offerings. The AI boom is a chance to lead: hyperscalers' demand can scale sustainable innovations in hard-to-abate sectors such as building materials.
- Everyone. Treat circularity as a journey: set ambitious, time-bound targets, start with achievable goals, learn and adapt, and track metrics, adding emerging measures of biodiversity and circularity to established ones such as emissions and recycling rates.
The source calls Meta's Talavera project a "191,000-hectare" site; local records give about 191 hectares, a thousandfold slip. Meta also redesigned it for dry-air cooling, cutting the data centre's own net water use to roughly 40,600 m³ a year (per Castilla-La Mancha filings, 2026). More broadly, the piece leans on single company examples (Cisco, Microsoft) whose figures are self-reported, and it gives no numbers on what circular measures cost or how far they offset AI's growth.
The IEA's Energy and AI (April 2025) put data centres at about 415 TWh in 2024 (1.5% of world electricity), heading to roughly 945 TWh by 2030. Microsoft's 2025 report showed total emissions 23.4% above its 2020 baseline, so the build-out is still outrunning the circular fixes.
For a bank or any large IT buyer, the 4-Use list converts directly into procurement clauses: require suppliers' circularity road maps, prefer equipment-as-a-service and refurbished hardware for non-critical estates, and report embodied (Scope 3) emissions of cloud and on-premises AI alongside the energy figure.
Don't Let an AI Failure Harm Your Brand
Julian De Freitas (Harvard Business School) starts from a plain fact: AI eventually fails, and when it does, whoever built or deployed it lands under public scrutiny. Much has been written on marketing AI to speed adoption, little on marketing it so a failure does less damage. Drawing on seven studies over seven years, he names five consumer pitfalls and how to defuse each, before and after something goes wrong. The tactics that sell AI today (bold capability claims, superiority boasts, humanlike personas) are the ones that magnify the blame when it fails tomorrow. (HBR, July-August 2025.)
The Cruise crash and its fallout
A human-driven Nissan hit a pedestrian and threw her into the path of a robotaxi run by Cruise, General Motors' autonomous vehicle (AV) subsidiary. An independent engineering consultant later concluded that no prudent human driver could have steered clear. But Cruise's first report to regulators left out that the car then dragged her about 20 feet underneath it. She was badly hurt and survived.
Cruise was not at fault for the collision, yet the episode became a corporate crisis, driven by the lack of transparency:
- Regulators. A $1.5 million fine from the National Highway Traffic Safety Administration (NHTSA).
- Prosecutors. A Department of Justice criminal investigation, resolved with a $500,000 fine.
- Licence. Its permit to operate in San Francisco was revoked.
- Company. Half the workforce lost their jobs, the CEO stepped down, and valuation fell more than 50%.
- Sector. Within months a crowd attacked and burned a driverless Waymo (Alphabet) in San Francisco, and NHTSA opened investigations into several AV developers, including Waymo and Zoox (Amazon).
- End point. By the end of 2024 GM announced it would stop developing robotaxis altogether.
The five pitfalls that follow, De Freitas argues, explain why the backlash was so out of proportion to Cruise's role in the crash.
Pitfall 1: People blame AI first
To see why the Cruise backlash was so severe, De Freitas and colleagues ran a study with more than 5,000 participants. Everyone read a scenario like the Cruise crash: a human driver hits a pedestrian, who is flung into a second, not-at-fault vehicle. For half, that second vehicle was autonomous; for the other half, human-driven. Participants then rated the manufacturer's responsibility.
- Result. The maker of the not-at-fault vehicle was judged more liable when it was autonomous, although in both versions the vehicle could have done nothing. The AV company was also judged more liable than the human driver of the not-at-fault car in the human version.
- Replication. An independent research group in China reproduced the effect, so the bias crosses cultures.
The mechanism is counterfactual thinking. Distracted by the AV's novelty, people picture what would have happened without it, and they imagine a perfect driver, holding AI to an unreasonable standard. They decide the AV could somehow have swerved before impact, even when that was physically impossible, and so rate it more responsible than a human in the same spot.
What to do. In follow-up tests, shifting attention to the at-fault human driver reduced blame on the AI: with focus moved off the novelty and onto other salient causes, people stopped inventing unrealistic scenarios. But redirecting attention is not the same as misleading stakeholders, and executives must never hide AI's role in a failure. Cruise stressed the human driver's fault while concealing the dragging; once that came out, it lost control of the media narrative and damaged regulators' trust.
Pitfall 2: When one AI fails, people lose faith in others
The tremors that hit Waymo and Zoox point to a second risk: a contamination effect, in which one firm's AI failure makes people assume other firms' AI is equally faulty.
Chiara Longoni, Luca Cian and Ellie Kyung told 3,724 people that the state of Arkansas had failed to allocate disability benefits properly to a disabled person. Some heard that a human employee was to blame, others that an algorithm was. Asked whether an employee or algorithm in a different state, Kentucky, would also err, participants thought the algorithm more likely to fail than the human Kentuckian. The effect replicated for other AI failures, such as misallocated Social Security payments.
The reason is ignorance of how AI works: many people treat AI as a homogeneous group with shared characteristics, not as distinct systems with different capabilities and flaws. A human's error stays personal; an algorithm's error indicts the category.
What to do. Make your system visibly different from competitors' before anything goes wrong:
- Emphasize differentiators. Point to proprietary algorithms, safety measures and human oversight. De Freitas cites Anthropic, which markets Claude as an assistant trained to be safe, accurate and secure and describes a "constitutional" training approach it presents as more transparent, interpretable and aligned with human values. That framing separates Claude from rivals such as ChatGPT and Bard, which are trained differently and have faced accusations of bias and inaccuracy, so a competitor's failure is less likely to stick to it.
- Publicize the human in the loop. When a person supervises the AI and decides whether to act on its recommendation, failures are less likely to be generalized to other AIs, presumably because they read less as an indictment of AI as a whole.
Pitfall 3: People blame companies that overstate AI capabilities
Tesla calls its driver-assistance system Autopilot, although it requires active human supervision. After several fatal crashes with Autopilot engaged, Tesla faced lawsuits, a DOJ investigation into whether the label is misleading marketing, and an NHTSA request that its communications reflect the system's real capabilities and limits. Elon Musk has defended the name as borrowed from aviation, where autopilot aids pilots rather than replacing them. De Freitas tested whether consumers read it that way.
In a study of 9,492 participants, using driving simulations and hypothetical accidents, people were told an automaker would launch a new automated vehicle labelled either "Autopilot" (high capability) or "Copilot" (mid-level capability). In the simulator, the car approached a busy intersection and would hit jaywalkers unless the participant took the wheel. Those told "Autopilot" took control later: the label alone bred complacency, and the more capable people believed the car, the later they intervened. Other studies showed people hold companies more liable for accidents when labels imply greater capability. One label, two harms: riskier behaviour and more blame.
De Freitas then predicted that a standard marketing move would backfire: claiming superiority over alternatives. Follow-up studies confirmed it. Superiority claims raise perceived capability and therefore perceived liability when failure comes, which suggests GM's "Humans are terrible drivers" campaign may have led consumers to blame Cruise more for its later accidents.
The root cause is the autonomy spectrum: AI systems range from partial autonomy (the human is chiefly responsible) to full autonomy (the AI is in control), and most consumers cannot tell where a given product sits. That leaves marketers torn between honest labels and capability-inflating labels that sell more.
What to do. Two options:
- Disclose alongside the label. If you keep an aspirational label, explain the true capabilities clearly elsewhere, on the website or even in the fine print on the product. Companies that pair a misleading label with accurate disclosure are penalized less after failures than those using the label alone.
- Choose a less misleading label. You forgo the boost to perceived capability, but De Freitas calls this route risk-free.
Pitfall 4: People judge humanized AI more harshly
Firms increasingly give AI human traits. Wysa, a mental-health app, guides users through exercises with a cartoon avatar; companion apps such as Replika use realistic human images that voice "feelings" and "reflections." Such cues imply feelings, goals and desires the bot does not have, and they pay: compared with neutral bots, humanized bots raise purchase intent, trust, brand loyalty, compliance with requests and willingness to share personal data, even when users know they are talking to a machine. The catch shows up at failure.
- "Charles" versus the algorithm. Raji Srinivasan and Gülen Sarial-Abi described an investment firm whose mistake cost customers money. Brand attitudes were more negative when the culprit was a humanized AI called "Charles" than when it was an "algorithm program." Follow-ups showed humanized bots are credited with remembering, understanding, planning and thinking, so they are held more responsible.
- Cues add up. A team led by Cammy Crolic found a failing chatbot rated worse when it used both verbal and visual cues (first-person language, the name "Jamie", a female avatar) than with verbal cues alone: humanlike traits have an additive effect on the backlash.
- Anger is the danger zone. Crolic's team analyzed about 500,000 chatbot sessions at an international telecom company. The more customers treated the bot as human (using its first name, for example), the more satisfaction fell when they were angry.
What to do. Use humanized bots selectively in neutral domains, such as product search, and less in roles that attract angry customers, such as customer service centres. And temper expectations at the start of each chat: Slackbot greets users by saying it tries to help but is "still just a bot. Sorry!" Bots that make such disclosures infuriate customers less when they fail.
Pitfall 5: People are outraged by programmed preferences
In 2016 a senior Mercedes-Benz manager said its self-driving cars would put passengers' safety above pedestrians' and other road users', reasoning that if only one life could be saved, it should be the occupant's. A media firestorm followed, one tabloid accusing the firm of choosing to run over a child rather than swerve. Within weeks Mercedes clarified publicly that neither engineers nor autonomous systems should rank the value of human lives. The reaction suggests people find it unethical to encode group-based preferences (by age, gender or customer status) into AI.
De Freitas tested that with Harvard's Mina Cikara. 826 US participants imagined a fully automated vehicle facing dilemmas, such as swerving into an elderly person or a young child, and then hitting someone from one group. The researchers varied whether the car chose at random or by a programmed preference for one group. Any programmed preference produced more outrage than a random choice.
What to do.
- Don't advertise group-based decision rules, and in some cases avoid collecting data on race, gender and age for such decisions in the first place.
- Use structural features instead, such as saving more people rather than fewer. Yochanan Bigman and Kurt Gray found people far more supportive of AI using structural preferences than group-based ones, presumably because the utilitarian rationale is one most people accept.
Conclusion: market for the failure, not just the launch
AI failures are inevitable, and the actions that boost adoption now, especially touting your AI's benefits and superiority, can turn against you when it fails. De Freitas's advice is to assess the five pitfalls before choosing a marketing strategy, so the company can sell its AI today while limiting liability and brand risk in a future failure.
| Pitfall | Why it happens | What to do |
|---|---|---|
| Blame AI first | Novelty prompts perfect-driver counterfactuals | Redirect attention to real causes; never hide AI's role |
| Contamination | People see AI as one homogeneous thing | Stress differentiators; publicize human oversight |
| Overstatement | Labels raise perceived capability and complacency | Honest labels, or clear disclosure beside them |
| Humanization | Humanlike bots are credited with minds | Personas for neutral tasks; set expectations early |
| Programmed preferences | Group-based value rankings feel unethical | Structural rules; avoid group data where possible |
Most of the evidence comes from vignettes and simulators, not real market reactions, and many studies are the author's own. The Claude passage reads close to endorsement, and differentiation may not shield a firm if a whole category is hit. Product references have also dated: Google renamed Bard to Gemini in February 2024. The DOJ matter was not a lawsuit settlement as the text implies: in November 2024 Cruise admitted filing a false report to influence a federal investigation, under a deferred prosecution agreement with the $500,000 fine.
Pitfall 3 reached the courts. In August 2025 a Miami jury held Tesla 33% responsible for a 2019 Autopilot crash and awarded $243 million, upheld in February 2026. In December 2025 a California judge found "Autopilot" misleading and "Full Self-Driving" false; Tesla dropped standalone Autopilot and relabelled FSD "(Supervised)", while suing to overturn the ruling.
Banks deploying assistants should avoid human names in complaints and collections, where customers arrive angry, open every session with a plain statement of what the bot can and cannot do, and keep marketing claims inside what model validation has evidenced; regulators and juries read the brochure. Credit and pricing models should rest on structural, explainable criteria, never on encoded group preferences.
Master Takeaways
Fifteen pieces by more than thirty authors converge on a short list of truths: AI is cheap to access, so access is not advantage; value appears only where firms redesign decisions, processes and roles; and every deployment needs people's trust, quality control and an honest story for the day it fails. The scarce resource in enterprise AI is not the model but the organization's discipline to go deep, measure ruthlessly and keep humans accountable.
- Prediction needs a decision. AI is cheap prediction; if you cannot name the decision it improves, "insights" means you do not yet know what to do with it.
- Equal access erases advantage. When everyone has the same tools, gains flow to customers and suppliers unless proprietary data, processes and culture differ.
- Go deep and narrow. One function or end-to-end process, transformed, beats a hundred pilots; only 4% of firms do it, for about 2x the ROI.
- Test everything that ships. Amazon found 60% of gen AI content that passed its reliability checks still hurt sales; only live experiments reveal which.
- Change management dominates. Challagalla and coauthors' rule: 70% people and process change, 20% data, 10% the AI itself.
- Market for the failure. Bold labels, superiority claims and human personas speed adoption but magnify blame when AI fails.
- Error cost beats intelligence. Judge tasks by the cost of errors and explicit versus tacit knowledge: no-regrets, creative catalyst, quality control or human-first.
- Include employees early. Top-down hyperautomation breeds cynicism, resistance and silos; an AI-savvy, humble leader and four inclusive practices turn tools into firm-wide value.
- Alignment is a feature. Safety, fairness, privacy and dignity are sold alongside speed and must be defined, coded, traded off, vetted in partners, fed back and monitored.
- AI moves decision rights. Better prediction can shift who decides, concentrate or decentralize authority; govern models the way you govern delegated power.
- Local certainty, global volatility. Better forecasts in one unit make it less predictable to others: coordinate where possible, modularize where necessary.
- Systems change slowly. Like electricity and the PC, AI starts with point substitutions and transforms industries only once the surrounding system is rebuilt.
- Design dialogue, not prompts. Co-thinking needs a role, a setting, a scripted back-and-forth and a prompt that asks the AI to challenge you.
- Agents are labour. Map tasks, assess capability, choose owned, leased or outsourced digital labour, and govern agents with identities, limits and audit trails.
- Buy outcomes, not job titles. Sourcing decisions start from the task breakdown, which may be filled by people, agents or both.
- Run the full life cycle. Selection, development, evaluation, adoption and management each need their own checklist; most of the 80% failures skip one.
- Accuracy is not adoption. Users must trust the algorithm, the developer and the process for overruling it; LinkedIn's unused tool shows what happens otherwise.
- Models decay. Stale data and shifting conditions erode performance, so monitoring and retraining belong in the budget from day one.
- Put processes around tasks. AI handles narrow subprocesses; an owned, measured end-to-end process strings them into value and supplies clean data.
- Fix errors at the source. Silos that quietly clean bad upstream data hide the real problem; process ownership exposes it.
- AI checks AI. Simple rules, statistical control limits and a reviewer model trained on different data catch what human review cannot reach at scale.
- Quality systems allocate capital. Experiments replace optimistic ROI forecasts and showed some smaller models beat bigger ones on cost and speed.
- Remove the competence penalty. Colleagues rate identical AI-assisted work 9% less competent (women 13%); map hotspots, convert skeptics, judge outcomes.
- Data from use is the moat. Industrial winners will compete on product-in-use data across fusion products, services, systems and solutions, not on physical assets.
- Pyramids flatten. Where AI automates junior work, firms become obelisks of AI facilitators, engagement architects and client leaders, and must rethink how expertise is learned.
- AI is physical. Manufacturing causes 40-50% of a server's life-cycle emissions, and gen AI may add up to 5 million tons of e-waste by 2030: use longer, again, differently, less.
- Never hide AI's role. Cruise's omission of the dragging, not the crash itself, destroyed regulator trust and the business.
- People blame AI first. Novelty invites perfect-driver counterfactuals, and one firm's failure contaminates everyone's AI; differentiate and show the human in the loop.
- Avoid programmed preferences. Group-based value rules (age, gender) outrage people; structural rules such as saving more lives are accepted.
Frameworks & mental models
| Name | What it says | Use it when |
|---|---|---|
| Four inclusive practices (De Cremer) | Social connection, tech-nontech collaboration, leader learning, reward being human | Launching AI to a wary workforce |
| Six alignment challenges (Abernethy et al.) | Define, code, trade off, align partners, human feedback, prepare for surprises | Building values into an AI product |
| Decision = prediction + judgment (Agrawal, Gans, Goldfarb) | Cheap prediction reshuffles who decides; coordinate or modularize | Mapping where AI changes authority |
| Co-thinking four steps (Farri, Rosani) | Assign role, define setting, outline dialogue, create prompt | Using gen AI to think, not just produce |
| Error cost x knowledge matrix (Anand, Wu) | No regrets, creative catalyst, quality control, human-first | Choosing which tasks to give AI |
| Three differentiators (Anand, Wu) | Targeted deployment, proprietary data, unique people and culture | Turning shared tools into advantage |
| Seven actions for digital labour (Stave et al.) | Map, assess, integrate, redesign, rules, capture value, stay human-centric | Planning a human-agent workforce |
| Five-stage AI life cycle (Bojinov) | Selection, development, evaluation, adoption, management | Running any AI project |
| Three pillars of trust (Bojinov) | Trust the algorithm, the developer, the process | Diagnosing why users ignore a model |
| Seven process steps (Davenport, Redman) | Owner, customers, map, measures, enablers, redesign, monitor | Embedding AI in end-to-end work |
| Competence-penalty fixes (Acar et al.) | Map hotspots, convert influential skeptics, remove the AI signal | Raising adoption among wary staff |
| Fusion strategies (Govindarajan, Venkatraman) | Products, services, systems, solutions from in-use data | Industrial and asset-heavy firms |
| Consulting obelisk (Duncan et al.) | AI facilitators, engagement architects, client leaders | Restructuring knowledge-work teams |
| Deep and narrow, four steps (Challagalla et al.) | Opportunity, lasting advantage, sequence, competitor watch | Prioritizing the AI portfolio |
| Catalog AI quality loop (Thomke et al.) | Audit, guardrails, A/B test, learning system | Gen AI content at scale |
| 4-Use circularity (Welch et al.) | Use longer, again, differently, less | AI infrastructure sourcing |
| Five failure pitfalls (De Freitas) | Blame AI first, contamination, overstatement, humanization, preferences | Marketing or launching AI |
Who's who
| Name | Role | Why they matter |
|---|---|---|
| David De Cremer | Northeastern University | Inclusive, AI-savvy leadership over hyperautomation |
| Agrawal, Gans, Goldfarb | University of Toronto, Creative Destruction Lab | AI as cheap prediction that reshapes decision systems |
| Bharat Anand, Andy Wu | NYU Stern; Harvard Business School | The four-zone task matrix and the paradox of access |
| Iavor Bojinov | Harvard Business School, ex-LinkedIn | Five-stage project life cycle and trust pillars |
| Thomas Davenport, Thomas Redman | Babson; data adviser | Reviving process management as AI's home |
| Oguz Acar and coauthors | King's College London and Hong Kong, Beijing universities | Measured the competence penalty |
| Govindarajan, Venkatraman | Dartmouth Tuck; Boston University | Fusion strategy for industrial firms |
| Goutam Challagalla et al. | IMD; ex-Reckitt CMO | Deep and narrow deployment |
| Stefan Thomke et al. | Harvard Business School; Amazon | Quality control for gen AI at scale |
| Julian De Freitas | Harvard Business School | Consumer psychology of AI failure |
The collection is practitioner-optimistic and case-driven: much evidence is single-firm stories (LinkedIn, Reckitt, Amazon, Deere), often told by insiders or advisers who sell the model they describe (Beaulieu ran Reckitt's marketing; two Catalog AI authors work at Amazon; the obelisk authors run AI-native consultancies), and few claims are independently measured. Several pieces predate ChatGPT or agents, and some details have aged or were wrong (Cruise, View, the retracted Air Force anecdote, the "191,000-hectare" Meta site, a misdated Pew report). Large topics get little space: job losses, labour bargaining power, concentration among model providers, and the security risks of agents. The 2025-26 data showing most firms still without P&L impact both confirms the authors' diagnosis and warns that their remedies are harder than they sound.
Connect it
- AI practice. Treat Thomke's loop and Bojinov's life cycle as one evaluation-driven development pipeline: offline evals, guardrails and a reviewer model gate every release, and online A/B tests decide what stays.
- Enterprise architecture. Anand and Wu's matrix becomes a routing policy in the AI platform (the zone sets the review step), Davenport's process owner owns the agents in each value stream, and Agrawal's coordination-versus-modularity choice maps onto event-driven, loosely coupled services.
- Banking. Acme Bank's mortgage redesign is the deep-and-narrow template for origination, KYC and payments exceptions; Bojinov's stages map onto SR 11-7 model risk management; and De Freitas's overstatement pitfall argues for keeping chatbot and model claims within what validation has evidenced.
- Governance. The competence penalty and agent-as-labour pieces together suggest measuring outcomes, not tool use, while logging every agent action with an identity and entitlement limit, as the four-eyes principle already demands.
- Sustainability reporting. Welch's 4-Use list belongs in AI procurement: count embodied (Scope 3) emissions of cloud and on-premises AI next to the energy figure.
About this book
HBR's 10 Must Reads on AI: Updated and Expanded by Harvard Business Review, 2026. This is a chapter-by-chapter study of the book, with summaries, maps, tables and diagrams; short quotations keep their page references. It is shared free for personal study, is not for sale, and is not endorsed by the author or publisher. Please read and buy the original.
Insight Editions · rajeshrmahapatra.com/insight-editions