The Capability–Value Gap

There is a persuasive argument making the rounds: intelligence is the largest total addressable market humanity has ever encountered. Software historically served a function — Excel transformed analysis, CRM transformed sales, ERP transformed operations. Intelligence is different. It can assist every employee, participate in every workflow, touch every decision. Its market is not a software category. It is a share of global cognitive labour itself.

I think the claim is broadly right. And I think most people are drawing the wrong conclusion from it.

A market that large will not be captured where everyone is looking — at the model. Capability and realised enterprise value are different curves, and they are already beginning to diverge.

The argument in 60 seconds

Model capability keeps rising. But for any given enterprise workflow there is a sufficiency threshold — a point past which more general intelligence stops being the constraint. Beyond it, value depends on context, permissions, reliability, evaluation and integration: the system around the model. Frontier models remain essential for discovering new workflows; specialised, cheaper models increasingly operate the mature ones. The gap between what models can do and what institutions extract from them — the capability–value gap — is closed by the harness, not the next model generation. The model may be rented. The learning system must be owned.

01 · SignalThe iPhone Curve

The move from an ordinary mobile phone to an early smartphone was a step change in what a person could do. Navigation, payments, photography, messaging and the world's information collapsed into one device, and each early generation meaningfully expanded life.

Then the device matured. Newer phones kept improving — faster chips, better cameras, brighter screens. The improvements were real. But for most users, moving between mature generations no longer transformed anything. The device kept getting better; the marginal utility of each improvement got smaller.

AI models may be approaching the same curve. The difference between a model that cannot reliably reason or use tools and one that can is enormous. The difference between two future models that are both highly capable may be decisive at the research frontier — and barely noticeable inside a stable enterprise workflow. A hypothetical GPT-7 will beat GPT-6 on the benchmarks. Whether a KYC review, a claims workflow or a policy assistant receives proportional value from that increment is a different question entirely. Usually, the answer is no.

Fig. 1 — The Capability–Value Gap. Past the sufficiency threshold, raw capability (dashed) keeps climbing while the value an enterprise extracts from the model alone (amber) flattens — the same shape as mature smartphone utility (grey). The teal curve is the only one an institution controls: value keeps rising when the model is combined with context, evaluation and orchestration.

02 · MeaningAfter Sufficiency, the Bottleneck Moves

This is the distinction the current discourse keeps missing:

Model intelligence and realised enterprise value are not the same thing. Once a model is sufficiently capable for a task, the constraint moves out of the model and into the institution.

A model can get better at mathematics, scientific reasoning and open-ended research without improving a bounded business workflow at all. Past sufficiency, the enterprise starts to care about different questions. Does the system retrieve the correct internal information? Does it act with the right permissions? Does it produce a predictable structure, cite its evidence, escalate when it should, operate at an acceptable cost — and improve a measurable outcome? At that point, another increment of general intelligence matters less than better orchestration. The question changes from how intelligent is the model? to how effectively can the institution convert intelligence into reliable action?

Fig. 2 — Where the bottleneck moves. Before sufficiency, the model is the constraint. After it, value is limited by everything to the right of the gate — and none of it ships with the next model generation.

03Consumer AI Rewards Breadth. Enterprise AI Rewards Depth.

Part of the confusion comes from treating consumer and enterprise AI as the same product. They are not.

A consumer wants one assistant that can explain history, plan a holiday, debug code and recommend a recipe. Generality is the feature, and the user alone judges whether the answer was useful.

A bank does not need its KYC agent to understand astrophysics. It needs the agent to command KYC policy, customer records, documentary evidence, exception handling, approval thresholds and escalation routes with extraordinary precision — and the institution must be able to prove the answer was correct, permitted, policy-compliant and reconstructable. Occasional consumer mistakes cost goodwill. Enterprise mistakes create financial and regulatory exposure. The enterprise does not primarily need breadth. It needs depth inside a governed boundary.

Fig. 3 — Breadth versus depth. Consumer AI is wide and shallow by design; the user is the judge. Enterprise AI is narrow, deep and provable — depth inside a governed boundary, where the institution must answer for every action.

04Frontier Models Discover. Specialised Models Industrialise.

None of this makes frontier models unimportant. It gives them a precise job.

At the start of a new use case, the organisation does not yet know the correct workflow, the important exceptions, the required context, the failure modes or the right evaluation criteria. A highly capable general model gives it room to explore ambiguity — this is the discovery phase, and it deserves the strongest practical intelligence available, with every trace, failure and human correction recorded.

But discovery is not the operating model. Once the workflow is known — examples, traces, accepted outputs, private evals, measured outcomes — the organisation may no longer need the broadest possible understanding of the world. It needs a smaller or purpose-built model with exceptional depth in a narrow domain: defined vocabulary, bounded inputs, a known toolset, structured outputs, explicit policies. Less intelligent in a general sense; more valuable in an operational one. Also cheaper, faster, easier to host and easier to govern.

The relevant distinction is not large model versus small model. It is broad intelligence versus task depth — different stages of workflow maturity requiring different forms of intelligence. The frontier model is the best instrument for discovering a workflow. The specialised model becomes the best instrument for operating it a million times.

Fig. 4 — Discovery to industrialisation. Frontier intelligence earns its cost while the workflow is unknown. Once traces and private evals define the task, suitable workloads migrate to specialised models — not because small beats large, but because maturity changes what intelligence is for.

05The Harness Becomes the Moat

Follow this to its conclusion and a deeper truth about enterprise AI comes into view: the model is not the system. The system is the model plus enterprise context, memory, workflow state, tools, identity, permissions, policy controls, human approval, traces, private evals, model routing and the connection to business outcomes. Together these form the harness — the orchestration and learning layer where the enterprise encodes how it actually works.

A foundation-model provider can supply general intelligence. It cannot supply the accumulated judgment of a bank, a hospital or a logistics company. That institutional knowledge lives in how the enterprise selects context, defines acceptable behaviour, measures quality and learns from its own mistakes. This is the same conclusion I keep arriving at from every direction: the value of a deployed AI system converges to the quality of the layer beneath the model, not the model within it.

The model may be rented. The learning system must be owned.

Fig. 5 — The harness stack. The model sits in a dashed, swappable slot. Everything above it — context, orchestration, evaluation, outcomes — is where institutional judgment accumulates. That is the moat, because it is the only layer a competitor cannot rent.

06The CFO Moment

What forces the transition? Not architecture reviews. Finance.

Enterprises will happily pay for the best model while experimentation is cheap relative to salaries — and falling token prices will not reduce total spend; cheap intelligence invites more consumption of intelligence, more agents, more workflows. But rising consumption eventually meets financial discipline, and the CFO asks the questions the demo never had to answer:

The CFO Moment

What revenue did this create?
What cost did this remove?
What risk did this reduce?
Why is the most expensive model doing this task?
Is token consumption growing faster than enterprise value?

That moment accelerates model routing and workload segmentation. High-consequence, ambiguous and frontier tasks will keep justifying the strongest models. Stable, repetitive, well-evaluated tasks will migrate to cheaper or specialised ones. Enterprises will stop optimising for the highest intelligence everywhere and start optimising for the correct intelligence at each point in the workflow — capability per unit of outcome, not capability per benchmark.

07 · ActionBeyond the Model

So yes: intelligence may be the largest TAM ever. But the largest market does not produce a single winner; it produces a hierarchy. Frontier labs create ever-broader intelligence. Compute and energy providers carry it. Orchestration layers decide which model, context and tool to use. Evaluation systems decide whether the result is acceptable. And the institutions that build proprietary learning loops decide whether any of it created value.

The first phase of AI was organised around model capability. The next phase will be organised around model sufficiency. Once a model clears the bar for a workflow, the competitive questions change: who has the deepest context, the strongest evals, the safest tool integration, the smartest routing, the tightest link from AI activity to business outcomes — and who improves the system after every run.

Start Monday

Map your workflows on the curve. For each AI use case, ask: are we still discovering, or is this task stable and measurable? Discovery earns frontier spend; stability earns a routing review.
Build the evals before the migration. A private evaluation set is the bridge from frontier to specialised. No evals, no evidence — no migration.
Audit what you own. If the model vanished tomorrow, what remains? Context, traces, evals and routing logic are assets. If the answer is “a vendor contract,” you are renting the moat too.

The frontier will keep advancing, and new model generations will remain scientifically and strategically important. But for most of the enterprise, the decisive question will no longer be whether the next model is smarter. It will be whether the organisation has learned how to put intelligence to work.

Intelligence may be the largest market ever created. The model is only the entrance to it.

Field guide · free

The Curious Mind’s Guide to Agentic AI — the mechanism underneath this essay: the harness, the orchestration layer, and the trust discipline, rebuilt from first principles in seven illustrated parts.

One brief, when there is something worth your time — on agentic AI, governance, and the operating model for systems institutions can actually trust.