The model is not the moat.

Two years ago, picking a model was a strategy decision. There was a clear leader, a visible gap behind it, and choosing wrong meant your product was noticeably worse than a competitor's.

That gap has been closing steadily. Not because any one lab stopped improving, but because the improvements arrive from several directions at once and diffuse quickly. A capability that was remarkable in one model shows up across the others within a couple of release cycles. Meanwhile the price of the capability you needed last year has fallen by an order of magnitude, and the open weights options are now good enough for a large share of enterprise work.

This is what commoditization looks like from the inside. It does not mean the models are identical or that the differences stop mattering. It means the differences stop being durable, and durability is the only thing that makes a moat a moat.

What this changes for a buyer

If model quality is converging and model price is falling, then any decision you make that is expensive to reverse is a decision you should think about twice.

The expensive-to-reverse decisions are rarely the obvious ones. Nobody gets locked in by the model itself. They get locked in by everything they built around it: prompts tuned to one model's quirks, evaluation pipelinees that only speak one API, retrieval pipelines coupled to one provider's embedding format, and agent frameworks where the orchestration and the model are the same object.

You will not be locked in by the model. You will be locked in by the six things you built to work around it.

The organizations that will move fastest over the next three years are the ones that treated the model as a replaceable component from the start. Not because they predicted which one would win, but because they arranged not to care.

Where the durable advantage actually is

If the model is not the moat, something else has to be. In enterprise work it is consistently these four, in roughly this order.

Access to the data

Your data is the one input a competitor cannot buy. It is also the part nobody scopes properly, because getting it out of the system that owns it, with the right controls and the right freshness, is a genuine engineering project rather than an integration.

The evaluation set

A good evaluation set is a durable asset that improves every time you use it. It tells you whether a change helped, which is the only way to move quickly without breaking things quietly. It also makes the model swappable, because swapping is only frightening when you cannot measure the result.

The workflow around the model

Where the approval gates sit, what happens on low confidence, how a human corrects it, and how that correction feeds back. This is most of the actual product, and none of it is model specific.

The right to operate

Approval from risk, audit, and legal, plus the audit trail that keeps it. This takes months to earn and it transfers across models. It is the least glamorous item here and often the most valuable.

---

The practical test

There is a simple diagnostic we run with clients. Ask your team how long it would take to move a production workflow from its current model to a different one, and to know whether the result got better or worse.

• Under a week and you have built the right way. The model is a component.
• A month and you have coupling you should be paying down now, while the cost is still small.
• Nobody can answer and you do not have an evaluation set, which is the more urgent problem.

Most teams are in the second or third category, and most of them are surprised by that. The coupling accumulated one reasonable decision at a time.

What we do with this

On AI engagements we design for substitution from the first week. Workflows are written against an interface rather than a provider. The evaluation set is built before the first build, not after the first complaint. And the cost model assumes the price will fall, because it will, and a business case that only works at current prices is a business case that gets better rather than worse.

Argo Intelligence is built on the same assumption. It ships with models included so you are not blocked on day one, and it will point at a different model, or your own, without the workflow changing. That is not a feature we are proud of technically. It is just the shape a platform has to be if you believe the model is a component.