AI is infrastructure. The moat is everything else.

Most of the pitches we read still open with the AI. Which model, how it was tuned, how good the demo looks. Two years ago that was a reasonable way to open. Today it tells us very little.

The cost of calling a frontier model has fallen roughly tenfold since 2023. Capabilities that once needed a research team are now an API call that a competent engineer can wire up in an afternoon. When everyone can buy the same intelligence at the same price, intelligence stops being what separates one company from another. It becomes infrastructure, like cloud hosting or payments.

So the question we ask is a simple one: what would this product be worth without the AI layer?

If the honest answer is “not much”, the company has a feature. If the answer is “our customers would still be stuck with us”, there may be a business.

What keeps customers

Five things, in our experience, survive the next model release.

Proprietary data. A dataset that grows every time the product is used, and that nobody else can assemble. The model on top can be swapped; the data cannot.

Workflow lock-in. A product so deeply embedded in an operational process that removing it means redesigning how the customer works.

Regulatory positioning. In European finance and employment, compliance decides the purchase. A product built for the AI Act, DORA and GDPR from the first day is ahead of one that bolts them on later, often by many months.

Network effects. In B2B these are rarer than decks suggest, but real where a workflow has several sides and each new participant makes it more useful to the rest.

Embedded expertise. Founders who have spent years inside an industry and turned what they know into product.

And three things that do not count: the choice of model, an impressive demo, and being first in a niche nobody has proven.

What this changes about founders

The commodity AI stack has lowered the technical bar. That moves the scarce skill from building to knowing. Knowing the customer’s operation in concrete terms. Knowing how to find and close the first ten buyers. Being able to iterate with real revenue on the line.

We still want at least one person on the team who has built software professionally. But a team that explains its business mainly through its technology is, more often than not, describing a feature.

Where to look

None of this is an argument against AI companies. Almost everything Steppe backs uses AI heavily. It is an argument about where to look when deciding whether a company will still matter in five years: one layer above the model, at the data, the workflow and the distribution.