Imagine that every bank, retailer and manufacturer receives the same excellent language model tomorrow. It reasons reliably over ordinary business documents, writes competent code and costs almost nothing to query. One company puts a chat window on its website. Another connects the model to customer history, permissions, prices, inventory, service recovery and a ledger of verified outcomes. Both can claim the same underlying intelligence. Only one has altered how consequential decisions are made.

Now make the thought experiment harsher. Six months later a stronger model appears. Both companies can switch to it. The first gains a better chat window. The second reruns its evaluations, routes the new model only to tasks where it passes, preserves its policy controls and continues learning from completed transactions. The difference survives the model change because it never resided principally in the model. It resided in the system that selected evidence, exercised authority, observed results and corrected error.

The central claim is that abundant models relocate competitive differentiation from access to intelligence towards the organisation’s ability to turn intelligence into a distinctive, governed and self-improving decision loop. A model can contribute to advantage, particularly when access is scarce or task performance is unusually specialised. Yet an advantage based solely on a purchasable model has a short half-life. Suppliers improve, open alternatives narrow gaps, employees carry practices between firms and customers learn to expect the feature. Durable returns require complements whose value grows through use and whose assembly is specific to the firm.

This is not a claim that every company needs proprietary data, a foundation model or a grand platform. It is a test of causation. The strategic unit is the smallest system that changes a customer or operating outcome and becomes harder to reproduce as it is used. Sometimes that unit is a carefully governed workflow. Sometimes it is a distribution relationship, a regulated permission, a recovery network or an operating habit. The model is one component. Treating it as the whole advantage confuses technical capability with economic appropriation.

Part IWhat becomes common, and what does not

Abundance is relative. Frontier training remains capital-intensive, leading models do not have identical strengths and access can be restricted by geography, latency, privacy or contract. The relevant strategic question is narrower: can a credible rival procure sufficiently similar capability for the target decision within the period in which returns are expected? If yes, model quality may determine operational performance but cannot, on its own, explain persistent excess returns.

The empirical direction supports taking this possibility seriously. The 2025 Stanford AI Index documented steep inference-cost declines and a narrowing gap between leading closed and open-weight models on selected benchmarks. The 2026 report continues to describe rapid improvement alongside a widening measurement and governance gap. At the workplace level, generative AI has produced material gains in bounded tasks. Brynjolfsson, Li and Raymond found a 15 per cent average increase in issues resolved per hour among 5,172 customer-support agents, with larger gains for less experienced workers. Noy and Zhang found faster, higher-quality performance on mid-level professional writing tasks. These results establish that access can create value. Their broad diffusion also warns that the same value may become the new competitive baseline.

Dell’Acqua and colleagues add an important qualification. In their management-consulting field experiment, AI improved performance on tasks within its capability frontier but could reduce correctness when participants relied on it for a task outside that frontier. The frontier is jagged, not a single line. A generic model can therefore be abundant while reliable task placement remains scarce. Firms differ in whether they know which work to delegate, how to verify it and when to stop automation. That knowledge is a candidate capability, although it becomes strategic only when it affects an important outcome and resists rapid imitation.

Classical strategy helps separate value creation from durability. Barney’s resource-based account asks whether a resource is valuable, rare, difficult to imitate and difficult to substitute. Teece, Pisano and Shuen shift attention from a stock of resources to the ability to sense changes, seize opportunities and reconfigure assets. Under fast model turnover, this dynamic view is especially useful. The advantage is less likely to be one frozen prompt, dataset or vendor choice. It is more likely to be the repeated organisational act of recombining models, knowledge, process, controls and human judgement as conditions change.

A model can be valuable without being rare, and rare without being durable. The strategic burden is to show why the complete decision system will improve through use faster than a rival can procure, copy or substitute it.

Five complements that can carry differentiation

Context means information selected for a particular decision, not an undifferentiated data lake. Customer history, contractual obligations, machine condition, local demand and prior exceptions matter only when they change the available action. Raw volume is a poor proxy. A rival may own less data yet possess cleaner links between conditions, interventions and outcomes. Context becomes defensible when its collection is lawful, useful, difficult to reproduce and tied to a recurring decision.

Authority is the right and technical ability to act. Many assistants can recommend a refund; fewer can calculate entitlement, check fraud controls, issue payment, notify the customer and open a recovery path. Authority converts prediction into consequence. It also creates exposure, so permissions, separation of duties and human escalation are part of the capability rather than administrative drag. NIST’s Generative AI Profile rightly treats risk management across the lifecycle, not output quality alone.

Outcome evidence closes the loop. A recommendation is not an outcome. Firms need to know whether the customer accepted the offer, the machine stayed operational, the claim remained accurate or the credit decision performed over time. Outcome evidence supports evaluation, routing and redesign. It can compound because each completed case makes future task placement and recovery more precise. It can also mislead when selection effects, delayed outcomes or policy changes are ignored.

Distribution and trust determine whether the capability reaches a valuable decision. A technically superior system may lose to an adequate one embedded in the place where customers or employees already act. Distribution is not just a channel. It includes consent, service responsibility, switching cost, reputation and the ability to recover when automation fails. In high-consequence settings, proof of restraint can differentiate as much as breadth of action.

Reconfiguration speed is the capacity to change the combination. Models, prices, regulation and user behaviour move at different rates. A firm that can replace a model, rerun evaluations, revise controls and retrain operators without interrupting service has a durable option. Vendor independence matters only to the extent that it supports this option. Maintaining five unused providers is not resilience; a tested substitution path is.

The imitation clock starts before launch

Managers often estimate imitation from the date a competitor sees the finished product. In digital markets that is too late. Suppliers demonstrate patterns across customers, staff move between firms, consultants codify operating methods and prospective employees can infer the architecture from job descriptions. Rivals may begin assembling substitutes while the original team is still in pilot. The economically useful lead is therefore not calendar time since launch. It is the interval between a firm proving the mechanism and competitors matching the customer outcome at comparable cost and risk.

That interval contains several different delays. A procurement delay covers selecting a model and contracting for access. A context delay covers cleaning histories, linking entities and obtaining permission to use them. An integration delay covers placing the capability inside real work. A legitimacy delay covers frontline trust, risk acceptance and customer consent. A learning delay covers observing enough varied outcomes to improve the policy. Generic capability reduces the first delay sharply. It may also shorten integration through standard tools. It does much less to erase a decade of service relationships or accelerate an outcome that becomes visible only after a year.

The delays are not automatically moats. Slow data preparation can reflect neglect. Lengthy approval can be bureaucratic waste. Customer lock-in can destroy value and invite regulation. A defensible delay is one that arises from creating value or bearing responsibility that a rival cannot instantly reproduce: lawful longitudinal evidence, dense local operations, trusted recovery, specialised know-how or a contract earned through demonstrated performance. Strategy should remove internal friction while preserving value-creating specificity.

Copy time also depends on the level at which equivalence is tested. Reproducing a screen may take days. Reproducing its recommendations may take weeks. Reproducing the effect on churn, uptime or loss may take several operating cycles. A team that compares screenshots will underestimate its advantage; one that treats any implementation difference as unique will exaggerate it. The correct counterfactual asks when a credible rival can cause the same customer-relevant result, including ordinary exceptions and recovery.

This leads to a useful capital rule. Compare the expected time to recover investment with the conservative imitation interval. If a feature costs £2 million and produces £150,000 a month before competitors match it, a nine-month lead returns only £1.35 million before risk and operating cost. It fails even if customers admire it. If the feature also secures service contracts that generate lawful outcome evidence, the calculation can include that option, but the option needs a separate probability and evidence path. Calling every early release “strategic learning” is not a substitute for valuing what is learned.

Part IIHow a common input becomes an uncommon system

The mechanism has four steps. First, a firm chooses a consequential recurring decision, rather than a fashionable interface. Second, it assembles the minimum context and authority needed to improve that decision. Third, it observes a verified result and attributes it cautiously to the intervention. Fourth, it changes the next decision through updated policy, routing, evaluation or human practice. Repetition creates a learning loop. Differentiation appears when the loop improves customer value or cost, when its learning is firm-specific, and when rivals cannot reproduce the complete loop before the return dissipates.

This account excludes mere usage. Millions of prompts do not automatically create a moat. Conversations may lack outcome labels, permissions may prevent reuse and the model vendor may receive as much learning as the customer. Nor does fine-tuning automatically solve the problem. A specialised model can improve consistency, latency or cost, but its strategic value still depends on the decision it changes and the evidence it generates. Technical adaptation and economic differentiation are separate tests.

Worked example at small scale: a specialist repair business

Consider a regional company that repairs commercial refrigeration. It receives 1,200 service calls a month. A generic model can summarise fault descriptions and draft technician notes. Competitors can buy the same feature, so the immediate benefit is shorter administration. Useful, but not differentiating.

The company instead targets first-visit resolution. It links each call to equipment model, installation history, ambient conditions, previous parts, technician actions and whether the unit remained operational for thirty days. The model proposes a diagnostic sequence, but an inventory service checks local parts and a policy engine limits what can be promised. Technicians record what they observed, not merely whether they accepted the suggestion. A sample of completed repairs is audited for misdiagnosis and avoidable replacement.

Suppose first-visit resolution rises from 72 to 80 per cent across a controlled six-month roll-out, while repeat failures do not increase. At 1,200 calls, that is 96 additional jobs resolved on the first visit each month. If an avoidable second visit costs £145 in labour, travel and scheduling, the gross monthly operating value is £13,920 before software, data preparation, audit and change costs. The model creates part of the value, but the harder-to-copy elements are the linked repair history, parts availability, technician behaviour, verification window and dispatch authority.

The strategic move is not to sell an “AI repair assistant”. It is to offer a response-time and uptime guarantee in the equipment segments where the loop is well evidenced. Distribution through existing maintenance contracts turns operating reliability into a customer proposition. The advantage remains bounded. A national competitor may possess more jobs and broader parts coverage. The regional firm wins only where local density, clean outcome evidence and rapid recovery outweigh the rival’s scale.

Worked example at system scale: industrial energy optimisation

Now consider a multinational food manufacturer with 40 plants, thousands of energy-consuming assets and exposure to demand charges, production schedules and food-safety constraints. A general model can explain energy reports at every plant. The feature spreads quickly and reduces analyst time. It does not automatically change the company’s energy position because suggestions have no authority, local constraints differ and avoided consumption is difficult to attribute.

A decision system starts with controllable load. It joins meter intervals, production plans, weather, equipment state, tariffs, maintenance restrictions and sanitation windows. Forecasting models estimate demand; an optimisation service proposes shifts; a language model explains the proposal and handles incomplete operator input. Plant managers retain authority over production risk. For low-risk assets, approved policies can change set-points within a narrow envelope. Every intervention receives a counterfactual baseline, an approval record and a reversal condition.

Assume the group spends £80 million annually on electricity. Analysis identifies £12 million of load as genuinely flexible without affecting throughput or safety. A staged programme reduces the cost of that flexible load by 6 per cent after metering error, rebound and control costs, yielding £720,000 a year. That figure is only a starting estimate. The larger strategic option comes from learning which production states permit flexibility, negotiating tariffs around demonstrated response and embedding energy constraints in future plant design.

The system may become difficult to copy because the advantage spans plant physics, operator trust, approved control envelopes, contractual tariff access and years of intervention evidence. Still, size is not proof. If equipment suppliers provide the same optimisation as part of standard controls, utilities standardise the tariff and competitors possess similar operating data, the return may diffuse. The company should value the programme on verified savings first and treat differentiation as an option contingent on faster learning or privileged market access.

Part IIIImitation, rivalry and the limits of the thesis

Thought experiment: the public learning pool

Suppose regulation requires every company in an industry to contribute de-identified intervention and outcome records to a public pool. Interfaces are open, switching is effortless and customers use a common marketplace. Models train on the shared evidence each night. In this world, proprietary outcome learning is weak. The company with the best decision loop today may teach the whole market tomorrow.

Would differentiation disappear? Not necessarily. It might move to service recovery, local execution, brand, physical assets or the ability to design better experiments. Yet the thought experiment reveals a boundary in the main claim: outcome evidence supports advantage only when the firm can appropriate enough of the learning or translate common learning into a distinct action. Data accumulation is not inherently defensive. Portability rules, shared standards and supplier access can make it widely substitutable.

Negative control: the meeting summariser

Take a company that deploys the same meeting summariser as its peers. Adoption is high, employees save ten minutes per meeting and summaries are competent. The tool does not connect to a consequential decision, does not observe whether commitments are fulfilled and does not alter distribution or trust. Rivals can activate it under the same licence. This is a negative control for strategic differentiation.

The summariser may still have an excellent return. Across 2,000 employees, small time savings can justify the licence. Refusing it could leave the company with a cost disadvantage. The correct label is productive parity. Strategy fails when managers reject parity because it is not unique, or describe parity as a moat because the adoption dashboard looks impressive. Operating necessity and sustained differentiation are both legitimate, but they require different return hurdles and governance.

The strongest rival explanation

The strongest rival is scale at the model and infrastructure layer. Perhaps downstream decision systems matter less than privileged access to frontier capability, compute, distribution and capital. Large model providers can observe usage across many customers, integrate tools into dominant platforms and subsidise adoption. The Federal Trade Commission’s 2025 study of major cloud and AI partnerships documents financial, technical and contractual ties that may shape access, switching and competition. The OECD’s 2026 report on AI markets likewise examines concentration and entry barriers across compute, data, models and downstream services.

This rival can explain substantial value capture. If one provider maintains a decisive capability lead, controls a default work surface and restricts portability, a downstream firm may rent intelligence rather than own differentiation. Its local workflow improvements may be copied into the platform or neutralised by contract. The decision-loop thesis therefore does not imply that value inevitably migrates downstream.

It does, however, make a distinct prediction. Among firms with comparable model access, those that connect intelligence to specific authority and verified outcomes should sustain larger performance differences through model turnover than those relying on prompts and interfaces alone. This is testable. Track cohorts across a provider upgrade. If relative performance resets with each model release and prior outcome learning adds no predictive value, the model-layer rival is stronger. If differences persist and correlate with workflow learning, the complement account gains support.

The explicit failure boundary

The framework fails in four conditions. First, when frontier capability itself is scarce and decisive for the decision, model access may remain the moat. Second, when outcomes are too delayed, rare or confounded to support learning, a feedback loop may create false confidence. Credit loss, medical harm and industrial failure can take years to observe and are shaped by selection. Third, when law or ethics forbids reuse of context or autonomous action, the proposed loop must stop at the permitted boundary. Fourth, when a dominant platform owns the customer surface, outcome data and switching terms, the downstream firm may be unable to appropriate learning.

A fifth practical boundary is organisational. Reconfiguration speed cannot be declared by architects. It depends on procurement, risk review, frontline practice, incentives and recovery. A modular technical design inside a six-month approval cycle may be less adaptable than a simpler system whose owners can change safely each week. Dynamic capability is an observed pattern of effective change, not an architecture diagram.

What evidence is needed before scale

The first evidence layer is technical equivalence. Define a task set that represents ordinary work, rare but material cases and deliberate failure probes. Run the incumbent and at least one credible substitute under the same context, tools, latency and human review. Report the distribution of results, not only the mean. This reveals whether model choice is decisive, whether routing adds value and where a cheaper substitute is adequate. Because the frontier is jagged, aggregate benchmark superiority cannot answer this local question.

The second layer is workflow causality. Randomise or stage access where practical, preserve a contemporaneous comparison and measure the real outcome. If randomisation is impossible, document selection and use matched periods or units with caution. Record manual work displaced into review, exception handling or data repair. Many pilots count faster generation while ignoring slower approval. A full measure includes the human and system effort required to reach a verified, recoverable outcome.

The third layer is appropriation. Map who observes prompts, context, corrections, actions and outcomes under the contract and architecture. Ask which party may reuse each record, who benefits from aggregate learning and what remains when the relationship ends. Export a representative evaluation set and rehearse a replacement. If the firm cannot take its policy evidence, audit trail or customer state to a substitute, locally created value may strengthen the supplier’s position more than its own.

The fourth layer is competitive response. Convene a red team with commercial, operating and technical knowledge. Give it the public product, a realistic budget and twelve months. Require three substitutes: copy the feature, neutralise it through price or distribution, and change the customer criterion so the feature matters less. Then rerun the unit economics. The exercise does not predict what a named rival will do. It prevents the business case from assuming that all other actors remain still.

Scale is justified when these layers agree. The intervention causes a material result, the result survives full workflow cost, the firm retains enough learning to improve and a credible competitive response does not erase the return before payback. Disagreement is informative. Strong task performance with weak appropriation suggests a vendor negotiation or architecture change. Strong operating value with rapid imitation suggests parity funding. Weak causal evidence calls for a better experiment, not a larger rollout.

Part IVThe Differentiation Durability Test

The Differentiation Durability Test is an operational pre-investment instrument. It asks a team to score the target decision after assuming that three credible rivals obtain equivalent model capability. Each dimension uses observable evidence. A score is not proof; it makes assumptions explicit and tells the team which experiment to run.

How to use the result

A commodity feature should face an ordinary automation hurdle. Buy rather than build unless integration economics say otherwise. Measure cost, quality and risk. Preserve exit options. Do not burden it with a narrative about long-term superiority.

A temporary execution lead deserves speed and an explicit expiry date. The team should estimate the imitation window, capture customer learning before it closes and avoid fixed costs that require permanent scarcity. Feature leads are useful when they secure distribution, contracts or outcome evidence that can support a stronger position later.

An operating advantage improves cost, reliability or recovery but may be invisible to customers. Protect the process, keep evidence clean and decide how much of the gain must be passed through. In competitive markets, operating advantages often become survival conditions. Their strategic value can lie in funding experimentation elsewhere.

A candidate compounding position warrants deeper testing, not celebration. Run a portability rehearsal, a rival-response simulation and a negative outcome audit. Separate learning that belongs to the firm from learning captured by the platform. Negotiate data-use, audit, termination and transition terms accordingly. Then test whether the customer-visible difference persists after the model is replaced.

The instrument also disciplines portfolio language. One company can hold all four classes at once. Meeting summaries may be parity, code migration a temporary lead, service recovery an operating advantage and a regulated decision network a candidate compounding position. Forcing every initiative into a single “AI strategy” obscures capital allocation. The right comparison is between the expected durability and cost of each outcome system.

ConclusionBuild the difference that survives the next model

Abundant models do not remove strategy. They make its burden more exacting. When capable intelligence can be procured by many firms, a demonstration no longer explains why one of them should earn superior returns. The answer must lie in a consequential decision, a complement that remains scarce, an outcome loop that improves through use or a trusted position from which action and recovery are possible.

The practical distinction is between owning a capability and owning the conditions under which that capability changes reality. Context without authority produces advice. Authority without verification produces exposure. Verification without appropriation teaches the market. Distribution without recovery spends trust. A durable system joins these elements and can be reconfigured when the model, rule or customer changes.

The changed executive decision is straightforward. Assume model parity before approving a claim of differentiation. Name the customer or operating outcome, trace how evidence changes action, estimate the rival copy time and test what survives provider substitution. If the margin disappears, treat the initiative as parity or a timed execution lead and fund it honestly. If the loop continues to improve, the firm controls enough of the learning and customers experience a difference after imitation, invest in the system. The moat is not intelligence in a box. It is the organisation’s demonstrated ability to make a better decision, verify what happened and make the next one better still.