Diligencing an AI company from first principles

Most diligence on AI companies measures the wrong thing. It evaluates how good the demo is today, when the question that determines the return is whether the company gets stronger or weaker as the underlying models improve. A great demo built on borrowed capability is a feature waiting to be absorbed. A modest demo built on a compounding advantage is a company waiting to grow into it. Diligence that cannot tell these apart will systematically buy the first and miss the second.

Diligence an AI company from first principles, which means setting aside the polish and asking what, exactly, this business owns that does not get cheaper when intelligence does. Almost everything else follows from that question.

The strengthens-or-weakens test

The single most useful filter is also the simplest. When the next model release lands, does this company get stronger or weaker? Hold a company up to that test honestly and most of the ambiguity resolves.

• A company that gets weaker as models improve is selling a capability gap that is closing. Its advantage has an expiry date.
• A company that is unaffected is at least not in the model's path, but it is probably also not capturing the upside of the shift.
• A company that gets stronger has positioned its core value to compound on top of capability rather than to substitute for it. That is the one to underwrite.

The strengthens-or-weakens test cuts through most of the noise in the category, because it forces you to locate where the company's value actually sits relative to the model. If a better model is a threat, you are looking at a wrapper. If a better model is a tailwind, you are looking at a company that has done the harder work of building above the model layer.

Separating a capability shift from a wrapper

A wrapper takes a capability the model provides and adds a thin layer of interface. It can be a perfectly good early product and a perfectly bad investment, because the thing it depends on is the thing it does not own. To tell a wrapper from a real company, look past the interface to what surrounds it.

• Does the company own a workflow end to end, or does it merely surface the model's answer and hand the user the risk?
• Is there proprietary context, private data, or hard integration that a fast follower cannot assemble in a weekend?
• Does the company take accountability for the outcome, or only responsibility for the suggestion?
• Would the product survive the model provider deciding to offer the same surface natively?

The wrapper question is not about whether a company uses someone else's model. Almost every AI company does, and that is fine. It is about whether the company has built something around that model that compounds and defends, or whether it is capturing a sliver of value the provider could reclaim in its next release.

Data advantage, honestly assessed

Data advantage is the most claimed and least examined moat in the category. Every deck asserts one. Few hold up. The test is whether the data is proprietary, whether it compounds with use, and whether it actually improves the product in a way customers feel. Public data scraped at scale is not an advantage; everyone has it. Generic usage logs are not an advantage; they rarely teach you anything a competitor cannot also learn.

Real data advantage comes from owning an outcome that generates a record no one else has: the corrections a human made, the result that worked or failed, the private context the customer trusted you with because you took responsibility. That exhaust compounds, and it does not get cheaper when the model improves. It gets more valuable. When you assess a data moat, trace it back to its source. If the source is ownership of a real outcome, the moat is probably real. If the source is volume alone, it probably is not.

The right question is never how good the model is. It is what this company owns that survives the model getting better.

Bring the same skepticism to the team that you bring to the product, because the half-life of any specific technical advantage in this market is short. What you are really underwriting is whether this team can keep finding the next advantage as the last one erodes, which is to say you are underwriting their rate of learning. Capability is cheap and getting cheaper. The scarce things are judgment, accountability, and proprietary context, and those are exactly the things a first-principles diligence is built to find.