Three projects use the same model. One extracts standard fields into a reversible staging area. One assembles evidence for an accountable expert. One proposes a service for customers the existing economics could not reach. Calling all three “copilots” erases the decision.

Automation, augmentation and invention describe how work, authority and value are recomposed. Automation assigns a bounded step to a configured system. Augmentation changes what a human decision-maker can perceive, compare or execute while accountability remains explicit. Invention creates a value proposition, decision population or operating model that was previously infeasible or uneconomic. The model is an input; the operating contract determines the mode.

The distinction matters because evidence does not transfer automatically. Automation needs codifiable acceptance, representative evaluation, reversibility, monitoring and a lawful delegation boundary. Augmentation needs proof that human judgement actually improves after verification, correction and exception work are counted. Invention needs problem and demand evidence, complementary capabilities, an operating owner, control feasibility and a route to capture the new outcome.

Mode clarity also improves portfolio comparison. Two initiatives can share a technical component while requiring different investment and governance. A retrieval service used inside automation needs reliable machine-readable evidence and an exception route. The same service used for augmentation needs an interface that preserves source context and supports human comparison. Used in invention discovery, it may need rapid proposition change and learning measures. Treating them as one platform benefit obscures which work and outcome each consumer actually buys.

A project can combine modes. Deterministic document classification may be automated, evidence synthesis may augment an underwriter and the combined capability may reach a new service population. Each layer has a separate unit of value and decision right. The portfolio should not choose one aspirational label for the entire initiative.

Nor are the modes a ladder. A safe, high-volume bounded automation can be more valuable than a speculative new product. Augmentation can be the permanent design where context is volatile and comparative judgement matters. Invention can fail despite brilliant generation because customers do not switch or complements are unaffordable. Mode selection begins with the causal structure of work.

01 · Mode definitionsChoose from work structure, not model capability

Automation: transfer a bounded step

Automation is appropriate when the task population and acceptance criteria are sufficiently codifiable, inputs and context remain within a proven boundary, errors are detectable within an acceptable time, consequential failures are reversible or containable, and exceptions can reach an authorised fallback. The system performs the step without routine human approval of every instance. Human work shifts towards boundary design, monitoring, exceptions and recovery.

The value unit is accepted output, reduced delay, higher capacity, consistency or another end-to-end outcome. Labour avoided may contribute, but a spreadsheet of task minutes does not prove feasibility or safe delegation. The automation contract states exactly which decision right moves. A system may extract, reconcile or route without acquiring the right to approve the underlying case.

Acceptance criteria should include abstention and degraded operation. The system must recognise inputs outside its proved population, route them without losing state and preserve a workable fallback when a dependency fails. Monitor not only error rates but entry into the boundary, exception volume, detection latency and recovery. A nominally automated step that requires experts to watch every instance is augmentation with expensive supervision. Its contract and economics should say so.

Augmentation: change a named human decision

Augmentation supplies evidence, comparisons, options or execution support to an accountable person. It is not defined by a chat interface or the presence of a review button. The work must identify the human decision, the information or capability changed, the residual verification burden and the conditions under which the person rejects, corrects or escalates the system's contribution.

The value unit can be decision quality, time to justified judgement, consistency, learning or capacity on cases where comparative judgement matters. Research from bounded settings can show task or worker effects, but it does not establish that every organisation will capture value. Brynjolfsson, Li and Raymond report heterogeneous effects in a customer-support setting.6 Noy and Zhang report controlled evidence on professional writing tasks.7 Local evidence must include the complete human-system workflow.

Invention: make a new proposition feasible

Invention means more than generating new content or adding an interface. It creates a product, service, decision population or operating model that could not previously meet acceptable economics or performance. The new proposition needs evidence of a consequential problem, customer or user demand, complements, authority, delivery feasibility and a capture route. Novel output is possibility evidence, not value evidence.

Acemoglu and Restrepo distinguish task displacement, productivity effects and new task creation in their economic framework.2 Historical work by Autor and co-authors examines the origins and content of new work over time.4 These sources do not classify a particular product idea. They support separating automation of existing tasks from the creation of new work and value structures.

02 · Work-recomposition contractWrite five fields that prevent hidden work and silent authority

The work-recomposition contract begins with what the machine does. Name the task population, inputs, output, tool access, acceptance criteria and boundary. “Assists analysis” is too vague. “Retrieves authorised policy passages and candidate file evidence with provenance for standard renewal cases” can be tested.

Second, record what the human or organisation does differently. Does an analyst stop searching and begin verifying? Does a manager compare options rather than assemble them? Does a service team reach a population it previously could not serve? Include review, correction, exception, monitoring, training and recovery. Work shifted into an invisible queue remains work.

Third, map decision rights. State what moves to the system, what remains with a person and what new authority the proposition creates. Decision rights include access, recommendation, approval, execution, override, risk acceptance, pause and withdrawal. Automation without this field becomes silent delegation; augmentation without it can leave accountability with people who lack genuine control.

Fourth, define outcome and counterfactual. Compare against the real alternative: existing process, deterministic redesign, no service, another product or a human-only route. The measure must match the mode. Fifth, state the escalation, redesign and mode-change triggers. A rise in novelty may move automation to augmentation; stable evidence and exercised controls may support more automation; repeated demand and complement proof may move invention discovery towards controlled delivery.

The contract needs a version and expiry because work drifts. A new policy can turn a codified step into a contextual judgement; a product change can create unfamiliar exceptions; a staffing change can make residual review unserviceable. Record which evidence supports the allocation and which change invalidates it. This gives release governance a direct question: does the proposed change alter machine work, human work, authority, outcome, counterfactual or a migration trigger? If yes, reopen the relevant acceptance.

03 · Temporal tensionAutomation and augmentation reshape each other

Raisch and Krakowski argue that automation and augmentation are interdependent across time and space, and that overemphasis on either can create reinforcing problems.1 This is the strongest challenge to a three-category framework. Work is distributed, capability changes and local automation can alter the human tasks around it. A permanent label will eventually become false.

The response is to treat each category as a governance snapshot. The contract declares the current allocation of work and authority for a bounded population. Augmentation can generate demonstrations, corrections and evaluation cases that later make a stable subset automatable. Automation can release attention for comparative judgement, customer interaction or exception learning. Invention can create new tasks that are then divided across both modes.

The loop must not become an automatic march towards automation. Context can remain volatile, error asymmetry can remain unacceptable and human comparative judgement can remain the value-producing capability. Conversely, keeping a person as ceremonial approver after acceptance has become codifiable may add delay without control. Reassessment should follow evidence and authority, not ideology.

04 · Residual workExpose the augmentation tax and automation mismatch

An assistant produces a polished draft in seconds. Experts now verify every factual claim, reconstruct missing reasoning, correct tone, resolve unsupported citations and accept accountability. If management counts generation time but not this residual work, the apparent augmentation gain is an accounting error. Review can also be harder than original production because the expert must detect plausible but unfamiliar failure.

Measure the whole recomposed workflow: preparation, prompt or request, machine time, verification, correction, exception handling, escalation, documentation, monitoring and recovery. Record distribution, not only average. The system may help routine work and impose a large tail on novel cases. Track who performs each activity and whether the work is recognised, trained, scheduled and authorised.

The augmentation tax becomes a governance problem when accountability stays with experts while control moves elsewhere. A reviewer needs access to source evidence, enough time, a genuine reject path and no incentive to approve by default. If the interface hides uncertainty or the queue makes meaningful review impossible, a nominal human approval does not repair the design.

Workforce distribution should be part of the value evidence. Senior experts may save drafting time while junior staff lose practice, operations teams inherit monitoring and a small group receives the hardest exceptions. Measure learning, workload and decision quality across roles, not only the average user. If the design relies on expertise that the new workflow no longer develops, the organisation is consuming an asset without replenishing it. Training and deliberate case exposure become complementary investments, not optional change management.

Automation mismatch is the converse. A variable, high-stakes exception process is automated because average task time is expensive. Context novelty, asymmetric error and late detection make silent delegation unsafe. Dell'Acqua and colleagues describe a jagged capability frontier in a consulting-task setting, with performance depending on whether tasks fall within the technology's frontier.8 The local boundary still requires representative evidence and a fallback.

Error severity and reversibility often dominate task speed. A reversible formatting correction is unlike an irreversible transfer, declined service or missed safety signal. Time to detect matters: an error corrected before use creates a different exposure from one found months later. Compare the system's error modes with the human baseline, including correlated failures and overreliance.

05 · Decision fieldContext and error shape the safe mode

Task codifiability asks whether inputs, outputs and acceptance criteria can be made explicit for a bounded population. Context volatility asks how quickly relevant facts, rules and relationships change. Comparative judgement asks whether value arises from weighing incomplete, conflicting or tacit evidence. Feedback quality asks whether outcomes arrive soon enough and clearly enough to improve the system.

Automation becomes more defensible where acceptance is codifiable, context is stable, error is contained, detection is timely, feedback is strong and exception routes are real. Augmentation becomes more defensible where comparative judgement matters, context varies and an accountable person can use the system's contribution without carrying unmeasured residual work. Do not deploy where severe error is hard to detect, authority is absent or neither system nor human can control the interaction.

Volume and delay cost affect economics but do not override the safety field. High volume can justify investment in bounded automation. It can also magnify correlated error. Demand elasticity matters because lower service cost may increase use, altering capacity and risk. Complementary workflow, data, skills and authority determine whether a nominal capability can operate.

Agrawal, McHale and Oettl offer a formal treatment of AI enhancing worker productivity without automating tasks through interaction with human capital.3 Autor cautions against mapping task automation mechanically to whole jobs and emphasises complementarities among tasks.5 Workforce design should therefore examine task bundles, learning, discretion, new responsibilities and distribution, not infer headcount from a mode label.

06 · Worked sequenceAutomate a bounded step, augment judgement, then test a new service

A regulated financial institution processes commercial-credit renewals. Deterministic checks and document classification for standard formats are codifiable, reversible before use and covered by representative cases. These steps can be automated into a staging area with monitoring and exception routing. The system gains no credit authority.

Evidence assembly remains an augmentation. The service retrieves authorised sources, extracts candidate values with provenance, identifies inconsistencies and prepares a structured pack. The underwriter compares evidence, resolves context, applies judgement and makes the decision. The contract measures decision-ready time, verified evidence quality, residual review, rework and accepted decision outcomes.

The combined capability may make a new service population feasible: smaller firms whose renewals were previously uneconomic to serve with the required evidence depth. That is an invention hypothesis, not a launch conclusion. The institution tests customer problem, willingness to use the service, acquisition, operational complements, unit economics, risk, authority and capture. The new population receives no consequential decision until the operating boundary passes.

This sequence creates a counter-case because each mode is independently justified. Bounded automation removes stable work. Augmentation improves an accountable decision without hiding verification. Invention discovery asks whether the new service creates and captures a worthwhile outcome. If demand fails, the first two modes may still be valuable. If context stabilises and evidence accumulates, part of the augmentation may later move into automation.

The work-recomposition contract prevents scope creep. A proposal to let the evidence system recommend approval changes machine work and decision rights; it triggers reassessment. A new jurisdiction changes context and authority. A rise in exception load can move a step back to augmentation. Mode migration is an explicit governance event, not a silent software release.

07 · Invention disciplineMove from generated possibility to captured outcome

Invention theatre begins with a striking generated concept and jumps to a business claim. It omits the customer problem, switching behaviour, complementary workflow, data, skills, authority, distribution, risk and capture. A new interface is not a new value proposition if it serves the same population through the same economics and changes no consequential outcome.

Idea-generation research can inform the possibility stage. Meincke and colleagues study large language models in idea generation, with results and limitations tied to that setting.10 Dell'Acqua and colleagues report a preregistered field experiment on product-innovation work and team configuration.9 Idea quantity or quality is not demand, delivery feasibility or captured value.

The invention funnel needs four successive receipts. Demand evidence shows a real problem and behaviour consistent with switching or use. Complement evidence shows the workflow, data, skills and distribution required to deliver. Control evidence shows lawful authority, risk tolerance, monitoring and remedy. Capture evidence shows how the organisation sustains or retains the outcome after full operating cost.

Discovery should preserve option value. Use prototypes, concierge service, synthetic cases and non-consequential tests to learn without pretending that the operating model exists. Define the observation that would stop the idea, narrow the population or change the proposition. An invention test advances only when the next uncertainty can change a decision.

Demand evidence should progress from stated interest to costly behaviour. Interviews can reveal problems and language; prototypes can show comprehension and workflow fit; commitments, switching, repeated use or willingness to bear a real trade-off provide stronger signals. None is sufficient alone. The team also tests whether the organisation can reach the population, serve exceptions, sustain quality and capture value after full cost. An attractive demonstration with no distribution or operating owner remains possibility evidence.

The OECD's evidence synthesis distinguishes task effects, jobs, firm productivity, innovation and entrepreneurship, helping prevent category errors.11 Its work on adoption also describes organisational and process complements.12 The local inference is straightforward: generated novelty must travel through demand, complements, controls and capture before it becomes an invention case.

08 · Executable instrumentMake the present operating contract explicit

The Work Recomposition Test chooses a current mode for one task or decision population. It examines codifiability, volatility, error, reversibility, human judgement, economics, feedback, residual review, new-proposition evidence, complements, decision rights and fallback. It does not ask which mode sounds most ambitious. The output is an operating action with reasons and missing evidence.

Hard boundaries come first. Severe, hard-to-detect or irreversible error combined with missing authority or fallback produces “Do not deploy”. Strong bounded acceptance and recoverability can support automation. Important comparative judgement with manageable review supports augmentation. Demand evidence for a previously infeasible population supports invention discovery. Mixed stable and judgement-intensive layers support a hybrid sequence.

Run the test before design and after representative evidence. The first result determines the discovery contract. The second authorises a bounded operating step or directs redesign. Keep the prior result when the mode changes so reviewers can see what evidence or authority moved. The test never turns a prototype into permission by accumulation of enthusiasm.

09 · Governance cadenceReview work, rights and residual burden together

A mode review begins with the current contract and representative operating evidence. Ask whether the machine still performs the bounded work described, whether human work has shifted, whether the decision right remains accurate, whether the counterfactual is still credible and whether a trigger has fired. Review by case class so a stable majority cannot conceal a dangerous novel tail.

For automation, monitor boundary entry, acceptance, exception, detection time, recovery and correlated failure. For augmentation, monitor decision outcomes, verification and correction, reject behaviour, workload, learning and overreliance. For invention, monitor problem evidence, switching, complement readiness, control, full economics and capture. A common dashboard should not force these unlike units into one score.

Workforce evidence needs distribution. Identify who loses routine practice, who gains exception load, who learns, who bears accountability and who has discretion. Protect time and authority for real review. Redesign roles, staffing, incentives and training as part of the capability. The organisation must not treat newly invisible checking as free labour.

Use a change log that names the trigger and the resulting contract decision. Examples include a stable case slice moving into automation, a new jurisdiction returning a step to augmentation, unmanageable review forcing redesign, or observed demand opening an invention pilot. For every move, archive the prior population, evidence, authority and outcome. This history prevents a gradual expansion in scope from being misrepresented as the original approval and shows whether the promised recomposition actually occurred.

NIST's AI RMF Core emphasises context, intended purpose, human roles, risk, benefits, monitoring and go or no-go decisions.13 It does not choose an operating mode. Its lifecycle practices support the contract: map the real work, measure its effects, manage exceptions and revisit the decision as conditions change. ISO/IEC 42001:2023 supplies management-system context for roles, policy, monitoring and continual improvement, not a value verdict.15

10 · ConclusionDeclare the mode, then earn the right to change it

Automation, augmentation and invention are choices about who or what does work, who holds authority and how value appears. Model capability does not answer those questions. Codifiability, volatility, error, reversibility, judgement, feedback, residual work, demand and complements do.

The work-recomposition contract turns a label into an inspectable design. It states machine work, changed human or organisational work, decision rights, outcome and counterfactual, plus escalation and mode-change triggers. It exposes silent delegation, hidden review and unproved invention claims before they harden into operations.

The categories remain deliberately temporary. Evidence can support a bounded move from augmentation to automation; new exceptions can reverse it. Automation can release attention for higher-value judgement. A combined capability can open invention discovery. None of these movements is automatically progress. Each needs its own outcome, control and owner.

A precise mode is therefore a promise about present operations, not a prediction about technological destiny. Its usefulness comes from making work, authority, residual burden and the next evidence decision visible to everyone affected, including operators, customers, reviewers and accountable executives.

The executive question is not “Which mode is most advanced?” It is “What is the current work-and-authority contract for this population, what evidence justifies it, and what observation will force the mode to change?” That question keeps ambition compatible with accountability.

SourcesReferences

1. Raisch, S. and Krakowski, S. (2021). Artificial Intelligence and Management: The Automation-Augmentation Paradox. Academy of Management Review, 46(1), 192–210.
2. Acemoglu, D. and Restrepo, P. (2018). Artificial Intelligence, Automation and Work. NBER Working Paper 24196.
3. Agrawal, A., McHale, J. and Oettl, A. (2026). Enhancing Worker Productivity Without Automating Tasks. NBER Working Paper 34781.
4. Autor, D. et al. (2022). New Frontiers: The Origins and Content of New Work, 1940–2018. NBER Working Paper 30389.
5. Autor, D. H. (2015). Why Are There Still So Many Jobs? The History and Future of Workplace Automation. Journal of Economic Perspectives, 29(3), 3–30.
6. Brynjolfsson, E., Li, D. and Raymond, L. R. (2023). Generative AI at Work. NBER Working Paper 31161.
7. Noy, S. and Zhang, W. (2023). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science, 381(6654), 187–192.
8. Dell'Acqua, F. et al. (2026). Navigating the Jagged Technological Frontier. Organization Science.
9. Dell'Acqua, F. et al. (2025). The Cybernetic Teammate. NBER Working Paper 33641.
10. Meincke, L. et al. (2024). Using Large Language Models for Idea Generation in Innovation. SSRN.
11. OECD (2025). The Effects of Generative AI on Productivity, Innovation and Entrepreneurship. OECD Artificial Intelligence Papers.
12. OECD (2025). The Adoption of Artificial Intelligence in Firms. OECD Artificial Intelligence Papers.
13. NIST (2023). AI Risk Management Framework Core.
14. NIST (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).
15. ISO (2023). ISO/IEC 42001:2023, Artificial intelligence management systems.