What AI Strategy Means When Models Become Commodities

At 9:00 on a Tuesday, a supplier releases a model that matches the quality of the one your programme spent nine months integrating. It costs one fifth as much. By lunchtime, two competitors have access. By Friday, the difference between their model and yours is smaller than the variation caused by prompting, data quality and human review. Which part of your AI strategy survived the week?

If the answer is only a model name, little survived. If the programme changed how a consequential decision is framed, supplied with evidence, authorised, executed, checked and improved, much more survived. The new model may even increase the value of that work. It can be inserted into an already functioning decision system and tested against evidence the firm has been collecting all along.

The central argument is that commodity-like models do not end AI strategy. They expose whether one existed. When intelligence can be rented by the token, the strategic question changes from “Which model do we possess?” to “Which scarce system of decisions and learning can we build around models that others can also obtain?”

Part IThe commodity test is about substitution

Commodity is a dangerous word because it sounds like a claim about intrinsic technical sameness. Wheat of a given grade is interchangeable in a way that language models plainly are not. Models differ in capability, modality, latency, context handling, tool use, transparency, deployment form, safety behaviour and price. Their performance is jagged across tasks. Stanford's 2026 AI Index reports that top systems are clustered closely on some broad rankings while still failing roughly one third of structured computer-use tasks; the same report notes a 3.3 percentage-point gap between the leading closed and open models in March 2026.¹

Yet a purchase can become commodity-like before the underlying technology is uniform. That happens when several suppliers clear the buyer's minimum performance boundary, prices are observable, capacity is available and switching is feasible within the life of the decision. The strategic fact is not that every model is equal. It is that the buyer no longer captures durable advantage merely by obtaining one of them.

Price and performance trends make this more than a thought exercise. Stanford found that the cost of inference at a GPT-3.5-level MMLU threshold fell from $20 to $0.07 per million tokens between November 2022 and October 2024.² Epoch AI's task-by-task analysis estimated price declines between 9-fold and 900-fold per year at fixed benchmark thresholds, with a median of 50-fold, while warning that the fastest declines may not persist and benchmark contamination can distort the comparison.³ The exact rates will move. The strategic pressure does not require those rates to continue. It requires only repeated improvements in the price, quality or availability of acceptable substitutes.

Value and advantage are different ledgers

A widely available model can create considerable value. It can shorten drafting, improve search, make expertise easier to reach and allow smaller teams to handle more work. The large field study by Brynjolfsson, Li and Raymond found a 15 per cent average increase in issues resolved per hour among 5,172 customer-support agents using a generative assistant. Less experienced workers improved more, and some learning remained during later system outages.⁴ Those are meaningful outcomes.

But an input that every rival can buy tends to become part of the competitive baseline. It may raise the whole industry's productivity while leaving relative positions largely unchanged. The economic value goes to customers through lower prices, to workers through better tools, to model suppliers, or to firms that combine the common input with scarce complements. A strategy must therefore keep two accounts:

The distinction also prevents a cynical error. If a capability is becoming a baseline, the rational response is rarely to ignore it. Refusing email because competitors also have email would not create differentiation. The right move is to acquire the baseline efficiently and avoid confusing the price of admission with the source of victory.

Where the commodity claim fails

Some decisions remain model-constrained. A laboratory may need a specialised molecular model unavailable through a convenient substitute. A device may require a latency, power or privacy profile that only one architecture meets. A jurisdiction may require a deployment form, provenance standard or sovereign control that sharply narrows the set. A frontier model can also create a temporary option on tasks that weaker models cannot perform at all.

These are reasons to make a model-specific investment. They are not reasons to skip the substitution test. State the threshold, the evidence that alternatives fail it, the expected duration of the gap and what must be portable if the gap closes. The strategy then becomes a deliberate purchase of scarce capability rather than a vague belief that the chosen provider is inherently strategic.

Part IIWhen the model diffuses, the rent moves to scarce rights

David Teece's account of technological innovation explains why an inventor does not automatically capture the profit from an invention. When imitation is easy, value can accrue to owners of complementary assets such as manufacturing, distribution or service rather than to the originator of the technical knowledge.⁵ Later work on dynamic capabilities located advantage in distinctive ways of coordinating and combining assets, in firm-specific knowledge positions and in paths that are difficult to reproduce.⁶

AI sharpens this logic because a model is unusually portable at the point of access. An API call can cross an organisational boundary in milliseconds. What does not travel with it is the firm's permission to see a customer's history, its definition of a good decision, the authority to act, the obligation to explain, the process for handling an exception, the ability to observe the eventual outcome and the managerial freedom to redesign the work.

This paper calls those complements decision rights. They are not legal rights alone. They are reliable organisational and technical entitlements that let a system participate in a consequential decision. Six are especially important:

1. Problem rights: the ability to define the decision, target and boundary rather than accept a generic task description.
2. Context rights: lawful, timely access to evidence that changes the answer for this customer, asset, case or moment.
3. Workflow position: presence at the point where attention and action are still available, rather than in a detached demonstration.
4. Action authority: a typed path from recommendation to permitted action, approval, abstention or escalation.
5. Outcome rights: the ability to observe what happened later and link it back to the decision that contributed to it.
6. Redesign rights: the mandate, skills and operating rhythm needed to alter roles, controls, interfaces and incentives when evidence shows a better design.

A small worked case: invoice disputes

This example also shows why “proprietary data” is too blunt. An archive is not automatically an advantage. Data becomes strategically useful when it is entitled, interpretable at the decision moment, linked to action and joined to an outcome. A million unlabelled messages may be less valuable than ten thousand cases with trustworthy state, intervention and postcondition.

An enterprise worked case: credit renewal

This is also where risk management becomes part of strategy rather than a tax on it. NIST frames AI risk management across the design, development, use and evaluation of systems, not as a one-off model certification.⁷ A firm that can map consequences, measure behaviour, constrain action and recover from failure can place models inside higher-value decisions than a firm that cannot. Trustworthy execution expands the economic frontier.

Part IIIBuild a portfolio of learning systems, not a parade of models

Once the source of rent is clearer, the unit of strategy changes. A use case is too small because it often names a feature without the decision, consequence or feedback path. A platform is too broad because shared infrastructure does not identify where value is captured. The useful strategic unit is a learning system around a consequential decision: a repeated choice with an owner, an outcome, evidence, authority and a mechanism for improvement.

This framing explains why general-purpose technologies often show delayed productivity. Brynjolfsson, Rock and Syverson argue that technologies such as AI require complementary intangible investments, including organisational capital and process change, whose cost appears before their gains are fully measured.⁸ Earlier firm-level research found complementarities among information technology, workplace reorganisation and new products or services.⁹ Buying the technical input is an event. Reworking the production system is a path.

The three clocks

An AI portfolio fails when it plans every layer on the same clock. The model market can change in weeks. A workflow may take quarters to redesign because incentives, roles, interfaces and controls must move together. Some outcomes appear only after a full customer, credit, clinical or maintenance cycle. A team can therefore switch models several times before it learns whether its first intervention improved the business decision.

The clocks imply a practical architecture. Model choice should be reversible more often than decision design. Separate model calls from policy, arithmetic, identity, permissions, state and outcome records. Evaluate candidates on the firm's own cases. Retain fallbacks and abstention routes. Record which version contributed to a consequential output. None of this requires pretending that models are identical. It makes their differences testable without allowing one provider's interface to become the definition of the business process.

What to standardise, and what to let vary

Contestability is not achieved by adding a second provider's logo to an architecture slide. It comes from a handful of deliberately stable seams. The first is a decision contract: a versioned statement of the input, allowed evidence, output, abstention conditions and responsible owner. The second is a representative evaluation set drawn from real operating cases, including difficult exceptions. The third is a model adapter that prevents provider-specific prompts, tool calls and error codes from leaking into every workflow. The fourth is a policy boundary outside the model, where permissions, thresholds and mandatory checks remain inspectable. The fifth is an event and outcome schema that preserves what happened after the recommendation. These seams make substitution an engineering exercise rather than an organisational crisis.

Suppose an insurer uses a model to prepare a claims handler's case note. A new model appears with better document vision. In a contestable system, the team can replay a stratified sample of prior claims, compare factual extraction, omission, handling time and downstream correction, and route only suitable document classes to the new model. The authority to settle remains where policy put it. The event record still names the evidence, model version, human intervention and later outcome. A provider change is therefore bounded and observable. In a tightly coupled system, the same change may require rebuilding prompts, interfaces, rules, analytics and training at once. The second organisation may own more code, yet have less strategic freedom.

Portability does carry a cost. Common interfaces can hide a capability that deserves specialised treatment, and maintaining several routes can consume more engineering effort than a small workload warrants. The answer is not universal abstraction. Set a portability threshold by consequence and switching exposure. A low-risk drafting aid may accept a thin adapter and manual export. A high-volume credit, safety or service decision may justify shadow evaluation, dual routing, preserved audit records and a tested fallback. The option should be purchased where being unable to exercise it would materially change the economics or the control position.

A useful negative control is to run the evaluation with the model identity hidden from reviewers. If a team cannot distinguish providers on the outcomes it claims matter, a strategic premium needs a different justification. If one system wins decisively on the firm's hardest cases, keep the advantage but record its boundary and re-test date. Standardise the evidence needed to choose; let the choice vary when the evidence changes.

Four positions, chosen decision by decision

There is no single correct posture for the whole enterprise. The appropriate position depends on whether model-specific capability is scarce and whether the surrounding decision system can differentiate.

A disciplined portfolio has three layers. The commodity core uses adequate low-cost models for reversible, low-consequence work. The specialist edge pays for measurable task advantages where context, latency or modality matters. The frontier option book runs bounded experiments on new capabilities before they are reliable enough for a business case. Each layer has a different funding logic. Confusing them produces either reckless scale or endless experimentation.

A moat audit without mythology

Before calling any layer a source of durable advantage, give a capable rival a fair attempt to reproduce it. Assume the rival can buy the same model and hire competent engineers. Then ask what it would still lack. Can it lawfully obtain equivalent context? Does it hold the relationship that produces the data? Can it insert the system at the same decision point? Is it authorised to take the action? Can it observe the outcome quickly enough to learn? Can it persuade the organisation to alter roles, incentives and controls? Finally, how much time and failed learning would replication consume? A convincing answer identifies a mechanism and a delay. “Our data”, “our people” or “our platform” without that detail is a label, not an economic argument.

This is the hard edge of the resource-based view. A resource contributes to sustained advantage only under demanding conditions such as value, rarity and difficulty of imitation or substitution; possession alone is insufficient.¹² In an AI system, rarity can also be temporary. A private corpus may matter until suppliers train on equivalent public material. A workflow position may matter until regulation mandates portability. An outcome history can compound for years, or merely encode yesterday's policy. Each claim needs an expiry condition as well as an investment case.

The audit also prevents a bad kind of defensibility. Friction created by trapping customers, obscuring decisions or collecting data without legitimate purpose may slow imitation, but it creates legal, reputational and operational liabilities. The strongest complements usually improve service while they improve learning: consented context, clearer accountability, faster correction, reliable provenance and a relationship in which customers have reason to supply better information. Strategy should seek productive specificity, not captivity.

If a rival could reproduce the entire system within the investment's payback period, classify the initiative as valuable efficiency and manage it accordingly. Demand a short route to cash, keep switching costs low and avoid a valuation story based on uniqueness. If replication requires years of outcome observation, embedded permissions and cumulative redesign, fund the learning mechanism explicitly. The distinction is useful even when it disappoints the presentation deck.

Portfolio reviews should ask what became cheaper, what became possible and what the firm learned. A provider leaderboard can inform the first two questions. Only outcome evidence answers the third. The meta-analysis by Vaccaro, Almaatouq and Malone is a useful warning: across 106 experiments, human-AI combinations performed better than humans alone on average but worse than the better of human or AI alone, with decision tasks showing losses and creation tasks showing gains.¹⁰ “Human in the loop” is therefore not a strategic design. It is an untested allocation of work until the roles and baseline are specified.

Part IVA decision-rent instrument for the next investment

The argument needs a decision tool, not a slogan. The instrument below separates annual use value from the share of differentiation likely to survive model diffusion. It does not produce a valuation. It exposes assumptions that teams commonly hide inside a single ROI number.

Start with one repeated decision. Estimate the avoidable loss or available upside, the model's measured improvement over the current baseline and the proportion of decisions that will actually use the system. Then assess five sources of retention: context specificity, workflow fit, outcome coverage, control and recoverability, and portability. Finally, state the expected half-life of the current model advantage. The chart shows how the composition of decision rent changes after that model-specific edge decays.

How to read the result

The addressable value is the decision volume multiplied by the avoidable value per decision. Realised first-year value then applies the measured model uplift, effective adoption and workflow fit. This deliberately leaves many business-case terms outside the calculator, including build cost, operating cost, timing, cannibalisation and uncertainty. Add them in the formal investment case.

The rent bars answer a different question. They allocate a diagnostic score across model uplift, context, workflow, feedback and control. The model component decays according to the stated half-life; the other components persist in this simple version. That persistence is an assumption to challenge. Context can lose value, controls can become obsolete and workflow knowledge can be copied. The tool is useful when it forces those claims into the open.

Make the decomposition a contested meeting

The instrument is most useful when its inputs are argued over by people who own different parts of the outcome. Take a claims operation considering a new triage model. The model team may defend an 18 per cent accuracy improvement. Operations may point out that only 55 per cent of recommendations reach the live queue. Risk may observe that fewer than half of resolved claims have an outcome label reliable enough for learning. Architecture may add that the prompt, retrieval contract and exception routing are tightly coupled to one provider. The disagreement is useful. It locates the work that a single return-on-investment figure would conceal.

Run the meeting in two rounds. In the first, estimate current use value using a range rather than a heroic point forecast. Record decision volume, avoidable loss, measured uplift, adoption and workflow fit, then state which quantities came from an experiment and which came from judgement. In the second, ask who controls each source of retention. Can the firm lawfully reuse the context? Can it observe the eventual outcome? May the workflow owner change the process? Can a failed action be recovered? Could another model be installed without rebuilding the evidence path? An attractive benefit estimate with weak answers in the second round is a valuable capability purchase, not yet a durable strategic position.

Do not resolve uncertainty by averaging every view. Preserve the range and attach an owner to the disputed assumption. A finance lead might accept the value estimate but reject the assumed adoption rate. A data owner might confirm access to case records while denying permission to use them for model improvement. A frontline manager might accept the recommendation yet refuse the proposed handoff because it adds a queue. Each objection changes a different component. The next experiment should target the assumption with the largest effect on the decision, not the one that is easiest for the model team to measure.

A practical review ends with one of four dispositions: buy common capability and keep it replaceable; run a time-bounded experiment to price a scarce model edge; invest in the surrounding learning system; or stop because the outcome cannot be observed or governed. That last result is not strategic pessimism. It prevents a team from calling inaccessible feedback a moat, or treating dependency on a supplier as differentiation. The decomposition does not predict competitive advantage. It establishes whether the organisation has identified the rights and routines from which a defensible advantage could plausibly be built.

Run the negative case

Set all five system-retention inputs near zero while leaving model uplift high; then raise decision volume and adoption. The first-year value can still look attractive, but the 24-month retained score falls sharply as the model edge decays. This is a valid investment when the payback is quick and switching costs remain low. It is not a durable strategic advantage.

Now hold the model uplift constant and raise outcome coverage. Nothing in the immediate model score changes, yet the system becomes capable of discriminating between good and bad interventions over time. That is the original contribution of the instrument: it separates the source of today's gain from the mechanism that may retain tomorrow's gain.