How to tell a real AI company from a wrapper

Every company raising money today is an AI company, at least in the pitch. For an investor, that makes the usual diligence questions insufficient. Traction and growth can both be real and still rest on a foundation that the next foundation-model release washes away. The question that actually matters is harder to answer from a data room: is this a durable company, or a wrapper?

A wrapper is a product that adds a thin layer of interface or prompting over a third-party model, without proprietary data, workflow depth, or a system a competitor could not rebuild in a weekend. Some wrappers are good businesses for a while. Few are good investments at a venture valuation, because the thing they sell is one release away from being free. Here are five questions that separate the two, and what a good answer looks like.

1. What survives the next model?

Imagine the leading foundation model gets twice as capable next year, which is a safe bet. Does this product get stronger or redundant?

Durable companies ride that wave. Their product improves as the underlying models improve, because they have built something that captures the new capability and applies it to a problem they own. Wrappers get swallowed by the wave, because the capability they resold is now native to the platform. Ask the founder directly what a much stronger base model does to their business. The ones with a real answer have thought about it constantly. The ones without one go quiet.

2. Where is the proprietary advantage?

Find the asset a competitor cannot copy quickly. It might be proprietary data that improves the product in a loop others cannot replicate. It might be distribution, a position in a workflow that is expensive to displace. It might be system depth, years of engineering that a prompt cannot reproduce.

If the only asset is a clever prompt or a well-designed interface, there is no moat, because both are visible and copyable. The strongest answers combine a proprietary data loop with deep workflow integration, so the product gets better with use in a way a new entrant cannot shortcut.

3. What is the dependency risk?

Most AI companies depend on one model provider. That is fine until it is not. Ask how exposed the business is to a single provider’s pricing, terms, and roadmap. A company whose margins vanish if the provider raises prices, or whose product breaks if the provider changes a policy, has a single point of failure sitting outside its control.

The good answer is not necessarily provider independence, which is expensive. It is awareness and optionality: they know their exposure, they have tested alternatives, and switching would be painful but survivable. The bad answer is a company that has never considered the question.

4. Do the unit economics hold at scale?

Many AI products are demos that cannot afford their own success. At low volume, inference costs are a rounding error and everything looks healthy. At scale, those costs can eat the entire margin. Model the unit economics at ten and a hundred times current volume. If the margin survives, the business is real. If it inverts, the growth story and the profit story are in conflict, and growth will lose.

This is one of the most common places a promising AI company breaks, and one of the least examined in standard diligence.

5. Can this team build the next version?

The current product is rarely the real asset. The platforms move fast, and today’s advantage commoditizes. What matters is whether this team can keep building faster than the ground shifts under them. That is a judgment about people, not slides: their technical depth, their speed of iteration, their honesty about what is working. A strong team with a mediocre current product is often a better bet than a weak team with a good one, because the product will change and the team is what compounds.

Reading the score

A company that answers all five questions well is riding the model wave: it gets stronger as the technology improves, owns something others cannot copy, manages its dependencies, makes money at scale, and has a team that will build whatever comes next. A company that answers none is standing in front of the wave, and no amount of current traction changes where it is standing.

Most companies sit in between, and that is the point of the exercise. The five answers do not just produce a yes or no. They tell you exactly where the risk lives, which is what you need to price the deal and to know what to watch after it closes.

This is the Wrapper Test, and running it well takes someone who can read the architecture as fluently as the market. If you have a deal that needs this kind of look, that is what our AI diligence engagement is for.