Using AI as leverage, not as the product

There are two ways to build a company with AI in it, and they age very differently.

The first is to make AI the product. The pitch is the capability. People show up because it is new, and the company's position depends on that capability staying scarce.

The second is to find a problem that was already worth solving, and use AI to solve it at a cost or a quality that was previously impossible. The pitch is the outcome. AI is in the engine room, and most customers do not care that it is there.

We build the second kind, and the reasoning is not aesthetic.

The capability keeps commoditizing

If AI is the reason your product exists, your competitive position is a bet that the underlying capability stays hard to get. That bet has lost repeatedly and quickly.

Capabilities that supported entire companies have become a paragraph of documentation and a per token price available to anyone. Not gradually. In months. Every time it happens, a set of businesses whose whole value proposition was access to that capability discover they were a thin layer on someone else's platform, and the platform just shipped their feature.

Building on top of a capability is fine. Building a company whose only differentiation is access to it is a company with a countdown attached, and the countdown is not visible from inside because revenue during the scarce period looks exactly like product market fit.

Where the durable position actually is

The defensible things are the ones that were always slow, and none of them got faster.

Proprietary data that accumulates through use, that a competitor cannot buy or scrape, and that makes the product better in a way users notice.

Workflow entrenchment. Being the place where the work actually happens, holding the history and the process, so that leaving means rebuilding an operation rather than swapping a tool.

Regulatory position. Certifications, compliance posture, audit history. Slow, boring, and precisely why they are a moat.

Distribution. Relationships, integrations, and channels that took years and cannot be shortcut by anyone with a good demo.

Trust in a domain where being wrong is expensive. Health, money, care, anything where a mistake has consequences. Nobody switches to save fifteen percent when the downside is that bad.

Notice what is missing. The model is not on the list. If a product takes weeks to build, it takes a competitor weeks to copy, so a studio operating today has to underwrite for a moat that is not the software, because the software is not one.

What this means for how the studio works

The same distinction applies one level up, to the studio itself, and this is the part that gets confused.

"AI venture studio" describes two different things. One is a studio whose ventures are AI companies. The other is a studio that uses AI as leverage in its own building process, so going from thesis to a working first version costs weeks instead of quarters.

The second is the more durable idea, because it improves the studio's economics regardless of what the ventures turn out to be. It is also the version we run. AI is how the studio operates, and a venture qualifies on whether the underlying problem is durable rather than on whether it is fashionable.

The venture studio model is roughly thirty years old and was, for most of that time, expensive. The reason so few existed is that building the first version of a company cost more than most people could fund repeatedly. That constraint has substantially weakened, which is genuinely good news for the model.

Three things that changed, and one that did not

The cost of a first version collapsed. The studio's scarce resource shifted from engineering capacity to judgment about what to point it at.

The right number of concurrent ventures went up, but less than people expect. Building got cheaper. Distribution did not. Customer acquisition, trust, support, and compliance cost what they always did, and those are where ventures actually stall. Studios that read the cost collapse as permission to run twelve ventures at once mostly discovered they had twelve distribution problems.

Durability got harder to assess. When the build is fast, "we made a good product" stops being a defense, and everything in the moat list above becomes more important rather than less.

What did not change: unit economics still have to work. AI can move the cost of delivering a service, sometimes a lot, and that is real operating leverage. It does not create demand, it does not make retention happen, and it does not remove the requirement that revenue eventually exceed the cost of getting and keeping a customer. A company with beautiful AI-driven margins and no retention is a company that is efficiently losing money.

The test

The question we ask on every venture is simple, and it is uncomfortable in a useful way.

If the AI in this product became free and universally available tomorrow, would the business still be worth something?

If the answer is yes, because of the data, the workflow position, the regulatory standing, the distribution, or the trust, then AI is doing what it should. It is leverage on a durable position.

If the answer is no, we are not looking at a company. We are looking at a feature with good timing, and timing is not an asset you get to keep.

The venture studio playbook covers how this filter fits into the wider model, and why recurring revenue beats a big exit covers the economics that make durability worth insisting on.