AI due diligence: how to tell real AI from a thin wrapper.

The hardest part of AI due diligence is that the demo always works. Whether the company has built something genuinely defensible or wrapped a thin prompt around a frontier model someone else trained, the demo on the screen looks roughly the same: a clean interface, an impressive output, a confident founder. The surface tells you almost nothing. Diligence is the discipline of getting underneath it — and in AI, what is underneath ranges from a deep, compounding technical moat to a feature that any competent team could rebuild in a fortnight.

I have run and reviewed this kind of diligence across a long career on both sides of the table — as an operator building products, as an investor including as CIO at the Nielsen Innovate Fund, and now through GenovateAI's AI Due-Diligence Memo, where the deliverable is a code-level, independent verdict on an AI asset. The patterns that distinguish real depth from a thin wrapper are consistent enough that they can be written down. This piece does that.

The single question that organises all AI diligence: if the frontier model providers shipped this capability natively tomorrow, what would be left of this business?

Eight layers, not one demo.

A serious AI diligence does not assess “the AI.” It assesses eight distinct layers, each of which can be strong or hollow independently of the others:

Patterns that signal a thin wrapper.

No single flag is fatal — plenty of good companies trip one. The signal is in the cluster. When several of these appear together, the probability that you are looking at a thin wrapper dressed as a defensible AI business rises sharply.

• ▲The demo is the only artifact. There is a polished demo and a pitch, but no evaluation harness, no error analysis, no honest account of where the system fails. A team that has built something real can tell you precisely where it breaks. A team that cannot has not looked.
• ▲“Proprietary model” that is a system prompt. The IP, on inspection, is a prompt template and some glue code over a frontier API. That can still be a real business — but it is not a model moat, and a deck that calls it one is signalling either confusion or spin.
• ▲A data moat made of other people’s data. The “unique dataset” turns out to be publicly scrapable, licensed non-exclusively, or simply the customers’ own inputs that the customers could take elsewhere. No compounding, no exclusivity, no moat.
• ▲Accuracy claims with no denominator. “95% accurate” on what test set, against what baseline, measured how? Headline metrics with no methodology behind them are marketing, and the absence of a real evaluation set is itself the finding.
• ▲Inference economics that only work at the demo. The unit cost to serve a real user at scale has never been calculated, or is quietly negative. A business that has not modelled its token costs at volume does not yet know whether it has a business.
• ▲A roadmap that is one model release from obsolescence. The core value proposition is a capability the frontier providers are visibly moving toward shipping natively. If the answer to “what survives if the providers do this themselves?” is “not much,” the moat is borrowed time.
• ▲Founders who deflect code-level questions. Technical diligence requests are met with NDAs, “it’s complicated,” or redirection to the demo. Real teams are usually relieved to talk to someone who understands the stack. Persistent deflection is a tell.
• ▲No evaluation or monitoring in production. The system ships outputs but has no harness measuring quality over time and no drift detection. This is the difference between a model that compounds and one that silently degrades — and its absence is a maturity flag.

Thin wrapper is not the same as bad bet.

An important nuance, because it is where careless diligence goes wrong in the other direction: “thin wrapper” is a description of technical depth, not a verdict on the investment. Some of the better businesses in this cycle are thin wrappers in the technical sense — their defensibility lives in distribution, a specific workflow, a regulated niche, switching costs, or brand, not in proprietary models. The error is not investing in a wrapper. The error is investing in a wrapper while believing you are buying a model moat, and pricing the deal accordingly.

So the job of AI diligence is not to issue a pass/fail on technical depth. It is to locate the defensibility honestly — to say plainly “the AI here is shallow, but the distribution and switching costs are real, and that is what you are paying for” or “the AI is genuinely deep, but the unit economics do not yet close.” A good memo gives the investor the real shape of the asset, with the red flags named and the moat located where it actually is. That is the deliverable: not a demo reaction, but a defensible, code-level verdict an investment committee can act on.

The frontier models will keep getting better, faster, and cheaper, and that movement will keep dissolving the moats of companies whose only advantage was access to a capability that is becoming a commodity. The discipline that protects capital in that environment is not a sharper demo sense. It is the willingness to get underneath the demo, ask the eight-layer questions, read the red flags as a cluster, and locate the defensibility where it actually lives — or report, honestly, that there isn’t any.