You are pre-LOI on an AI-enabled target. Finance and legal are handled. The question no one in the room can answer is whether the AI is real, whether the moat survives contact with a competitor's API budget, and what the answer does to the number you are about to put on paper. That is the question this page is about.
Start here: AI washing is not a marketing problem you catch by reading the deck more skeptically. It is a value-location problem. The only thing that matters about the target's AI is which specific layer of the system encodes something no vendor sells, and whether the premium you are being asked to pay rests on that layer or on a layer anyone can rent. AI washing is what happens when the pitch says the value is everywhere and the architecture says it is nowhere you can't buy.
What AI washing actually means in a deal
The label gets used for everything from fraud to enthusiasm. For diligence, narrow it. AI washing is the target overstating how much of its product, moat, or unit economics depends on proprietary AI rather than a bought capability dressed up as one. In most modern AI systems the model itself is a buy. It is an API call, off the shelf, the same one available to the target's competitors. That is not washing. It is normal. The washing is in the claim that the bought part is the defensible part.
So the diligence question is never "does the target have AI." It will. The question is where the un-buyable knowledge has to live for this business to be worth what they are asking, and whether it actually lives there.
“AI washing is what happens when the pitch says the value is everywhere and the architecture says it is nowhere you can't buy.”
The diagnostics you can run before LOI
Three of these you can run without a data science background. Front-load them, because they tell you fast whether to keep paying for deeper technical work.
The pause. Get into a room with the target's domain expert, the person whose judgement the product supposedly encodes, and ask them why the system does something the way it does. Then watch how the answer arrives. An instant, crisp answer means the logic is an empirical fact, codifiable, and a vendor has almost certainly already encoded it. That is a buy, which means it is not a moat. The pause is the opposite signal. They reach, they draw on years of doing the work, the answer is a judgement that won't reduce to a rule. That hesitation is the moat. If every "why" comes back instantly, the target has nothing here that a competitor can't license next quarter.
The audit handoff. Watch what the target hands you in the data room. If their evidence of AI is a spreadsheet of features with a column of ticks, where each row asserts a capability and the proof is that the row exists, you are looking at the structural tell of washing. A capability list can only ever say "we have these things." It can never say where the value is, because the value question is the one a checklist takes as given. Real proprietary value does not present as a tidy grid. It presents as one or two layers someone struggles to explain quickly.
The perception gap. Ask leadership what the valuable thing is. Then ask the people building and operating it the same question. When they disagree, when the founders point at the model and the engineers point at the data pipeline, or the reverse, you have found something more dangerous than a thin moat. You have found an organisation that cannot see where its own value lives. A target that can't locate its value can't have been deliberately building a moat there, and post-close you inherit a build aimed at a misconception. The tell is a perception gap, not a capability gap, and it is invisible unless you put both groups in front of you separately.
“If every "why" comes back instantly, the target has nothing here that a competitor can't license next quarter.”
Where the value lives, layer by layer
The build-versus-buy line in any AI system runs through the architecture, not around it. Picture the system as layers. At the bottom, the model: bought, commodity. Around it, the plumbing: authentication, transcription, generic summarisation. Nobody's moat is their login screen. The interesting layers are the ones in between. How the target decides what information is relevant for its specific domain, how it structures and labels its data, what it does with the model's output after the fact. The proprietary value, if it exists, lives in exactly one or two of those middle layers, and your job is to find out which.
This is where washing hides, because of what I'd call the seduction trap. The bought layer, the model, gives the fastest satisfaction. The demo works in an afternoon and it feels like doing AI. So engineering attention and budget flow toward the visible, exciting, already-solved layer and starve the unglamorous one where defensible value actually accrues. When you see a target that has poured its build into the model layer and treated data and retrieval as plumbing to be rushed, you are looking at a product that got prettier without getting more valuable. Eighteen months of that produces something polished, undifferentiated, and trivial to replicate.
What AI washing looks like in practice
A common pattern: the "AI engine" is a frontier model API behind a clean interface. The genuinely hard work, the domain-specific data and the logic that decides what's relevant, is thin or absent. The capability is real in the sense that it functions in a demo. It is washed in the sense that the valuation treats a rented capability as an owned one. Run the pause test on it and the experts answer instantly, because there is nothing tacit underneath. Run the perception-gap test and leadership describes a moat the engineers have never heard of.
The moat that resists being written down
The strongest version of a real moat cuts against the current market line that anything describable can be prompted into a model. The deepest proprietary value often resists articulation entirely, and even when you force it into words it still doesn't transfer, because the value never lived in any single rule. It lived in the interaction of many tacit judgements that don't decompose into instructions.
For you this signal cuts both ways. A target whose value genuinely resists extraction has something no acquirer-funded API budget can replicate, which is the best news a technical diligence can deliver. It is also the slowest, hardest thing to verify, and the thing most likely to walk out the door if the people who hold it leave after close. Which leads to the conversation the target has almost certainly avoided.
The conversation the target avoided
Here is the contrarian read, and it is the one that most often changes a price. The thing a leadership team is usually avoiding when it sprays AI across the organisation is the workforce decision: which roles get reorganised, reduced, or redefined. They make it implicitly, by avoidance, hoping it sorts itself out without anyone saying the hard thing out loud. That is cowardice dressed as strategy, and it is operationally destructive, because the people whose roles are most exposed are exactly the people you need engaged to make the AI work, and they can read the evasion.
This is not an ethics aside. It is the ROI. The value of any AI deployment is gated by adoption, and adoption is gated by trust. When a target's leadership rolled out tooling while refusing to say what it meant for jobs, the people who hold the workflow knowledge did the rational thing. They used the tool shallowly, they did not surface the tacit expertise that would make it powerful, because that knowledge is their leverage, and they waited it out. You cannot extract the moat from people the company has been lying to.
“You cannot extract the moat from people the company has been lying to.”
For an acquirer this is a diligence finding with teeth. The moat you are paying for lives in people, and a target that AI-washed externally has almost always evaded internally too. The tell on the inside is the same quiet death: AI quietly reclassified from transformation to cost of doing business, surviving as a line item used for shallow tasks, with a settled internal verdict of "useful, but it didn't change anything." That verdict is nearly impossible to reverse, and you would be buying it.
Turning a finding into a number
A washing finding rarely means walk. It means decide between three moves. If the diagnostics come back clean, the pause is real, leadership and the floor agree on the value, the proprietary layer resists articulation, you can support the AI premium and invest. If the moat is a bought capability wrapped in a UI, you don't necessarily pass, you reprice: strip the premium back toward a conventional software multiple, or move the unverifiable consideration into an earn-out tied to the thing you couldn't confirm. You walk only when the gap between the story and the architecture is wide enough to suggest the rest of the data room deserves the same suspicion.
This is worth a second, independent technical read whenever the price you are contemplating depends on an AI claim that no one on your side of the table can currently verify.
Related insight: The investor’s technical due diligence playbook — what a full technical read covers, beyond the pre-LOI screen.
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.