By the time a target calls itself "AI-enabled," someone has already decided the AI is real and is looking to you to confirm it. That is the position to resist. The walk-away decision on AI washing does not turn on whether the target lied. Most don't. It turns on a narrower question you can actually answer inside a data room: when you strip out the bought model and the human-in-the-loop dressing, is there a layer left that no competitor could rent by the same Tuesday? If there isn't, you are underwriting a wrapper at a moat's multiple, and the move is to walk or reprice to what the wrapper is worth.
What AI washing means when it's your money on the line
AI washing is the distance between what a target's marketing says its AI does and what its architecture actually does. In a consumer context that gap is a nuisance. In a live deal it is a valuation error you are about to buy. The version that matters to an acquirer is narrow: a target presenting bought, commodity capability as proprietary, defensible technology, and pricing the equity accordingly.
The model itself is almost always a buy. An API call, off the shelf, available to anyone. That is not washing by itself, and it is not interesting. The washing is the claim that the API call is the moat.
The pause: the fastest test in the data room
Get past the CEO and the head of product to the person who actually does the work, and ask them why the system works the way it does. Then watch how the answer arrives.
An instant, crisp answer means it is an empirical fact. It is codifiable, which means a vendor has probably already encoded it, which means there is no moat there. The pause is the opposite signal. They reach, they draw on years of doing the thing, the answer is a judgement call rather than a rule. That hesitation is the only moat worth paying a premium for, because the thing that is hard to articulate is hard precisely because it is tacit and accumulated, which is exactly what no vendor has.
Now invert it for washing. Ask how the AI works and get a fluent, confident, instantly quotable answer that never bottoms out in anything specific to this company, and you are likely looking at a wrapper. Real proprietary value is hard to put into words. Fluency about the model and silence about the data is a tell.
“Fluency about the model and silence about the data is a tell.”
Run the build-buy line through the architecture, not around it
The common mistake is asking "is the AI real" about the whole system. It is the wrong unit. Ask it layer by layer. The model is bought. The question is which specific layer of the surrounding system encodes something no vendor has: retrieval, how relevance is defined for this domain, how metadata is structured, the data itself and the right to use it. In a defensible target the un-buyable knowledge sits in one of those layers and you can point to it. In a washed one, every layer is rentable and the company has drawn a moat around the perimeter where none exists inside.
So do not accept a system-level claim. Make the target show you the layer. If they cannot name where the proprietary value lives, the build budget went into the part that felt like progress, the model, and starved the part that would have been worth owning.
The demo is the surface, not the value
A polished demo is the cheapest thing to build in an AI company and the easiest to stage. Leaders steer by the part they can see, the interface, because the value layer is illegible to them, and washing exploits exactly that reflex. The red flags in a demo or an RFP response are concrete: reluctance to name the model, reliance on an aggregate accuracy number with no methodology behind it, no production latency data, and an inability to reproduce the same result twice on fresh inputs you supply. An example of AI washing in the wild is a target whose "AI" is a generic model plus a person quietly correcting outputs behind the screen. It is labelled automation and staffed by humans, and it falls over the moment you put your own inputs through it live.
Fraud, overstatement, and what each does to your posture
Most AI washing is not deliberate fraud. It is a founder who half believes their own deck. That distinction changes your posture more than your price. Deliberate misrepresentation is a walk, and depending on what was filed and to whom, a legal matter with disclosure exposure attached. Well-intentioned overstatement is usually a reprice: the technology is real but worth less than claimed, and you can structure around it with consideration tied to the capability actually shipping at the promised level rather than the demo level. The work in diligence is to place the claim on that spectrum, because the same surface finding sends you to two different decisions.
What to request, and what absence signals
Ask for the documentation that a company genuinely doing the work already has: model and system documentation, the evaluation methodology behind any accuracy claim, data provenance and the rights to use that data, and inference cost per transaction at current volume and at ten times current volume. Absence is itself the finding. No eval methodology means the accuracy number is decoration. No data rights means the moat may not survive a single counterparty changing its terms. And if the margin only works because inference is being subsidised or the volume is tiny, the moat question is moot, because the unit economics do not survive the scale your model assumes.
The conversation the target is actually avoiding
Here is the read most diligence misses. When a target sprayed AI across the organisation as productivity tooling and counts seats deployed as transformation, it made a decision about jobs by avoidance, hoping it would sort itself out without anyone saying the hard thing out loud. That avoidance is not an HR footnote in your diligence. It is a direct read on whether the AI value you are underwriting exists at all.
The value of any AI deployment is gated by adoption, and adoption is gated by trust. The people who hold the tacit workflow knowledge that would make the system genuinely powerful are the same people whose roles are most exposed by it, and they are not stupid. When leadership rolls out tooling while refusing to say what it means for their jobs, they do the rational thing: they protect themselves, use the tool shallowly, and do not surface the knowledge that is their leverage. So the "AI transformation" on the target's slide is shelfware with good PR, and the deep workflow value the model was supposed to capture was never extracted, because you cannot extract a moat from people the target is lying to.
“You cannot extract a moat from people the target is lying to.”
There is a concrete test for this, and almost no management team will have done it. Ask the target to model five real roles eighteen months out with the AI in the workflow. Who is coordinating, who is redeployed, who is genuinely redundant, said out loud. If leadership cannot or will not, the transformation is theatre and your integration risk is real, because the value depends on people who were never told the truth and so never trusted the thing.
When you're the one who's already been sold
The most expensive AI washing in a deal is not always the target's. Watch your own side. If you find yourself reaching for reasons the thin layer is fine every time diligence adds a complication, you have stopped evaluating and started confirming a decision someone made emotionally before the data room opened. The tell is that the technical read keeps getting treated as an obstacle to the deal rather than an input to it. When that happens, the diligence is already lost, and the answer is to walk.
“The most expensive AI washing in a deal is not always the target's.”
Briefing the investment committee
Do not bring the committee a capability spreadsheet with ticks in the boxes. That artefact can only ever validate the premise that the target is the right one. Bring three findings instead, each already translated into the decision. Where the proprietary layer is, or that there isn't one. Whether the unit economics survive inference at scale. Whether the workforce story is honest enough for the value to land. Then state the call plainly: invest, reprice, or walk.
Walk when there is no un-buyable layer and the multiple assumes one, when the economics only work below scale, or when your own team has shifted from evaluating to confirming. Reprice when the technology is real but oversold, and structure the consideration around the capability actually shipping. Proceed when the pause is real and the data is theirs to use.
This is worth an independent second opinion whenever the equity check rests on an AI moat that no one on your own side can locate in the architecture without taking the target's word for it.
Related insight: What happens after you discover the AI isn’t real — the decision framework once a finding lands, repricing through walk.
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.