On a live deal, the AI-washing question is narrow and expensive. Is the target's AI a defensible asset, or a frontier-model API call wrapped in a marketing layer, and what does the honest answer do to your price and your structure. Finance and legal won't surface it. It needs an independent technical read of the target's AI, and most deal teams have no one in-house who can give one.
Start where the value is supposed to live.
Where a target's AI value actually lives
In almost every AI product on the market, the model is bought. It's an API call to a frontier lab, off the shelf, available to the target's competitors on the same terms. The model is never the moat. If a target's pitch rests on "we use advanced AI," they've told you nothing, because that capability is rentable by anyone with a credit card.
“The model is never the moat.”
The defensible value, when it exists, sits in one specific layer of the surrounding system: the proprietary data, the retrieval logic (how the system decides what information is relevant for a narrow domain), the optimisation layer that improves with use, or the encoding of expert judgement that no vendor sells. Your job in diligence is to find that layer and confirm it's real. The washing runs inside the architecture, not around its perimeter.
So the first question isn't "do they use AI." It's which single layer here holds something a competitor with the same model and the same budget could not reproduce in a quarter. If you can't name that layer after a day in the data room, you are probably looking at a wrapper.
The "why" test, run on the target's own experts
The fastest way to locate the moat, or fail to: get the target's domain expert or lead engineer in a room and ask why the system does something the way it does. Then watch how the answer arrives.
An instant, crisp answer means it's a codifiable fact. If it's codifiable, a vendor has probably already encoded it, and the target's "proprietary" version is replaceable. That's a reprice signal.
The pause is the opposite. When the expert reaches, draws on years of doing the thing, and gives you a judgement rather than a rule, you've found something tacit and accumulated. That hesitation is the closest thing to a real moat you'll see in a management session. It won't reduce cleanly to a rule, which is exactly why no competitor can buy it.
“That hesitation is the closest thing to a real moat you'll see in a management session.”
Note what this means. A target that answers every "why" smoothly and confidently may be the weaker asset. Fluent answers describe commodity. The reaching, the slight struggle to articulate, points at value that resists extraction.
The perception gap inside the target
Ask the CEO what the AI does and why it wins. Then ask the engineers the same thing, separately. A gap between those two answers is the single most reliable washing tell available to you.
When leadership describes a capability the engineers don't recognise, one of two things is true: the deck is ahead of the product, or leadership doesn't understand where their own value lives. Either way, the AI claim driving the valuation isn't anchored to what's been built. The tell is a perception gap, not a capability gap, and it is worth more than any architecture diagram.
Washing or honest immaturity? The line you have to draw
Not every overclaim is fraud. A genuine early-stage product can use AI for real and simply not have proven the economics yet. The line matters because it moves the decision from walk to reprice.
A target that can't fully articulate its moat might be washing, or might be sitting on tacit value it hasn't productised yet. Separate the two with evidence, not vibes. Honest immaturity shows its work: a clear definition of what "good" output looks like, data they can show you, a roadmap that matches the engineers' account. Washing shows polish over substance, a demo that always runs on the same inputs, and a story that gets vaguer the more technical your questions get.
The cleanest discriminator: ask what "good" output means and how they measure it. A team building something real can define good and show you the eval. A team that can't define good has no way to evaluate its own system, which means neither can you, and the AI claim is unfalsifiable. Unfalsifiable is a reprice at best.
What a credible claim looks like at each stage
Calibrate against maturity instead of one bar. At prototype stage, a credible claim is narrow: the model does X on this slice of data, here's the output. At pilot, it's that X holds on real customer data at this volume, with an eval and a clear account of where it breaks. At production, it's that X runs at scale, the unit economics are known, and they can tell you what it costs per inference. A target claiming production-grade transformation while describing prototype-grade evidence is the gap to price in.
Unit economics: who does the value accrue to
If the only AI in the system is a call to a frontier model, the margin story belongs to the vendor, not the target. Value created inside a vendor's roadmap accrues to the vendor. When the model provider raises prices or ships the target's feature natively, the claimed AI moat reprices to zero and the inference line eats the gross margin. Ask for cost per inference at current volume, then at ten times that. If they can't answer, the unit economics are an open question and you're underwriting it blind.
The contrarian read: the asset might be people, structured as software
Most AI-washing diligence hunts for the lie in the product. The more expensive miss is a truth nobody in the deal wants to state. Where a target's AI value is real, it usually lives in tacit expertise held by specific people, knowledge that was never encoded into the system because it resists being written down. That means the defensible thing isn't the software. It's the team. What you are buying might be people, structured as if it were software.
“What you are buying might be people, structured as if it were software.”
This is the diligence version of the avoided conversation. Boards get sold an AI asset; what's actually changing hands is a group of experts whose judgement the system depends on and who can walk twelve months after close. If no one models that, the moat leaves with the people. The concrete move almost no diligence team makes: take the five roles the AI value actually depends on and describe what each one looks like eighteen months post-close. Who's still here, who's retained on paper but disengaged, whose tacit knowledge has been captured and whose hasn't. Say it out loud in the IC memo. The targets where this conversation has an answer are the ones where the value survives the transaction.
Internal AI washing runs the same way. When a target's leadership has overclaimed AI to its own board and investors, the engineers usually know, and they've stopped correcting it. That silence is the same evasion, and it tells you something about both the value and the people you'd inherit.
From finding to decision
Translate the technical read into the only four outputs that matter. A real proprietary layer, defensible against the same-model competitor, with economics that survive inference at scale: invest, at a price that reflects the verified moat and not the deck's. A thin wrapper dressed as a platform: pass, or reprice to a software-margin multiple. Genuine value trapped in unencoded human expertise: reprice and restructure, with the retention, earnouts, and key-person terms that hold the moat in place past close. An acquirer already in love with the deal, treating diligence as an obstacle to clear rather than an input: that's the deal team's tell, not the target's, and it's the one that ends in a write-down.
On the legal side, an inflated AI claim that's material to the valuation is a representation you want in writing, with the indemnity to match. Public AI overclaims also carry live regulatory exposure, so where the target has made specific AI assertions to its own investors or the market, treat the distance between claim and system as a disclosed risk, not a footnote.
When the number turns on whether the AI is real and no one in-house can make that call, an independent technical read is cheap insurance against the most expensive line in the model.
Related insights: The investor’s technical due diligence playbook — the wider read this page sits inside; and Thin wrapper or true AI? — the same real-vs-rented question argued for a general technical audience rather than a live deal.
Proof point: we've built production AI from the ground up -- see an AI product taken from zero to beta. Knowing what real takes from the inside is what lets us tell a genuine system from a marketing layer.