Real AI or a wrapper: how to tell during diligence

The AI wrapper vs real AI question comes up in almost every deal we look at, and it is usually asked the wrong way. Investors want a binary: proprietary technology, good; wrapper around someone else’s model, bad. The honest answer is that nearly every AI product is a wrapper in some technical sense, including several very good businesses. The real diligence question is different: does the price assume technology that is not there?

A wrapper can be an excellent company. It can own a workflow, a distribution channel, and a data loop that compounds. What it cannot honestly claim is a technology moat, and when the valuation is built on one, that gap is the finding.

The claim problem is old and well documented

Inflating the AI label predates the current wave. When MMC Ventures reviewed 2,830 European startups classified as AI companies for its State of AI 2019 report, it found no evidence of material AI in roughly 40 percent of them. The incentive has only grown since: the label attracted funding then, and it supports multiples now. Nothing about foundation models changed the economics of claiming more than you built. Diligence exists because the incentive exists.

What separates AI wrapper vs real AI in practice

Forget "do they train their own models", which almost nobody should do anyway. In diligence we look for engineering depth around the model, because that is where real products differ from thin layers:

• Evaluation: a real system has an eval set built from production failures, scores tracked over time, and releases held back by regressions. A thin layer has vibes and a demo script.
• Error architecture: retries, fallbacks, confidence thresholds, human review queues. The thin layer passes the model output straight through and hopes.
• Data flywheel: does usage produce labeled data that measurably improves the product, or does "our data moat" mean a database of logs nobody uses
• Model independence: can they swap providers in days, and have they ever done it, or is the product one deprecation away from a rewrite (provider dependency is its own diligence topic)
• Prompt-to-product ratio: how much of the codebase is orchestration, validation, and integration versus one long prompt

None of this requires proprietary ML. All of it requires real engineering, and it takes a reviewer about two days in the repos and logs to see which one they are looking at. The evaluation question is the fastest single test: in our experience, teams with a living eval set almost never turn out to be thin layers, because nobody builds evaluation infrastructure around a product that is one prompt. When the claims are quantitative, accuracy numbers, benchmark wins, we verify the performance claims directly rather than debating labels.

Defensibility lives outside the model

If the model layer is rented, defensibility has to live elsewhere, and it can: in workflow depth (the product is welded into how work happens, and ripping it out costs more than the subscription), in proprietary data feedback loops that competitors cannot replay, in distribution, in switching costs. We go deeper on this in evaluating an AI moat when the models are everyone’s.

The pattern to distrust is defensibility claims that restate the model’s capabilities. "Our AI understands legal documents" is not a moat; the same API call is available to every competitor for a few cents. The moat, if there is one, is everything the company built around that call.

Price it as what it is

So the diligence output is not "wrapper: yes or no". It is a repricing. A thin layer with strong distribution and a real workflow position is a good software business with software multiples and a margin structure that needs watching, since a rented model layer makes unit economics sensitive to someone else’s pricing decisions. A company with genuine evaluation infrastructure, data loops, and model independence deserves more benefit of the doubt on the technology premium. And a pitch that claims proprietary AI while the repo shows one prompt and no evals is not a bad business either, necessarily. It is a mispriced one, and now you know.

We run this exact assessment inside our technical and AI due diligence engagements: two to four days in the code, the logs, and the eval sets, ending in a plain statement of what is real, what is rented, and what the difference should mean for the price.