The shift this analysis tracks
For a few years the working assumption was that the next jump in AI value would arrive with the next, bigger frontier model. Teams who run AI in production have started betting differently. The deployments that pay for themselves tend to be small, narrow models pointed at data the enterprise already owns. The frontier model becomes a general-purpose utility you rent for reasoning, while the workflow that shows up on the P&L is a tightly scoped task running on proprietary records. This piece looks at why that pattern keeps recurring and what a CIO should do about it when funding AI.
A worked example
Take contract renewal review at an insurer. A general assistant can summarise a contract and sound impressive doing it, but no one can say what that saved, because it touches a hundred different documents in a hundred different ways. Now point a narrow model at the firm's own historical renewals, the clauses that caused disputes, and the decisions adjusters actually made. The task is small enough to baseline against last year's manual review, cheap enough that the per-decision cost is obvious, and contained enough that you can watch what it does and catch it when it drifts. One of these produces a number a CFO will accept. The other produces a demo.
Owned data is the advantage and the exposure
The reason a small model can beat a much larger general one on a specific task is the data behind it: internal records, prior decisions, and documents no public model has ever seen. That is where the edge comes from. It is also where the worry comes from, because the moment that data flows into an AI workflow it has to be governed. Who can the model expose it to, what is allowed to leave your own systems, and can you show, later, exactly what was sent where. Teams getting real returns from domain models treat data readiness as the first filter on which use cases are even worth attempting, long before they argue about which model to use.
The fear that kills these projects, and the fix
The buyer worry here is concrete and rarely abstract compliance. It is AI slop reaching a customer, a model leaking something it should not have, and nobody being able to see what the system did after it shipped. A narrow domain model on owned data makes all three easier to handle, because its scope is small enough to reason about, control, and observe. The practical fix is to wrap the use case in a control layer from day one: decide what data it can touch, keep an auditable record of its decisions, and watch its behaviour and cost in production. That is what gets a domain model past the pilot and into real use, fast.
What it means for the portfolio
If the returns sit in narrow, owned-data use cases, the funding strategy follows. Do not build the plan around any single frontier model, because they commoditise; rent the best one available and keep the freedom to swap it. Put the investment into the use cases, the data they depend on, and the control and observability layer that lets you run them safely. Sequence by how quickly each can prove a number. Production teams reach the same conclusion from the cost and governance side that this analysis reaches from the ROI side: own the use case and the controls, rent the model, and ship the governed workflow fast enough that the value lands while the budget conversation is still open.
Frequently asked questions
Are smaller domain-specific AI models better than frontier models?
Not in general. For one well-scoped task on data you own, a domain-specific model is usually cheaper to run, easier to evaluate, and easier to govern, which makes its return straightforward to prove. Frontier models stay the better choice for broad, open-ended reasoning.
Why is owned data central to domain-specific AI ROI?
Proprietary records and past decisions are what let a small model outperform a general one on a specific job. That same data has to be controlled and recorded the moment it enters a workflow, so data readiness becomes the first filter on which use cases are worth attempting.
How should this change AI investment decisions?
Rent the frontier model rather than building around any one of them, and invest in the use cases, their data, and the control and observability layer that runs them safely. Sequence by speed-to-proof so early domain deployments land an attributable number.