You are buying a company that calls its product AI. Finance and legal are covered. The open question is whether the AI is real, whether it defends anything against a competitor, whether the unit economics survive inference costs, and what the honest answer does to the number on the term sheet. That question is technical, and it is the one no one in the room is yet equipped to answer.
So start where the answer is cheapest to get.
The single move that prices the claim
Before you read the data room, get in a room with the people who built the thing and ask one of them why the system works the way it does. Then watch how the answer arrives.
If it comes back instantly and crisply, the logic is an empirical fact. It is codifiable, which means an off-the-shelf model or vendor has probably already encoded it, which means the target's "AI" is a thin layer over something anyone can rent. That is the signal to reprice toward a conventional software multiple. You are not buying a moat. You are buying an integration with a marketing department attached.
If the expert pauses, reaches, draws on years of doing the work, and hands you a judgement rather than a rule, that hesitation is the asset. The thing that resists clean articulation is tacit and accumulated, which is exactly what no competitor and no model can reproduce on demand.
“You are not buying a moat. You are buying an integration with a marketing department attached.”
The pause is a physical, in-the-room signal. Run it on three or four people who hold different parts of the system. Uniform instant answers price one way. A consistent reach for the same hard-to-explain judgement prices the other.
What AI washing actually is when you are the buyer
AI washing is the distance between what the target says its AI does and what the system does. The deal-specific version has a particular shape: the AI premium is already in the ask, and the distance is your repricing lever.
It often is not deliberate fraud. There is a spectrum of intent, and where a finding sits on it changes your response. At one end, inadvertent washing, where management genuinely misunderstands what its own engineers built and has been overclaiming up to its own board in good faith. At the other, deliberate misrepresentation engineered for the raise or the sale. Inadvertent washing is a repricing and reps-and-warranties conversation. Deliberate washing is a walk-or-indemnify conversation. You cannot tell which you are looking at from the deck, which is the whole reason an independent read exists.
Watch especially for internal AI washing that has already laundered itself into the documents you are reading. An executive overclaims to the board, the board's belief becomes the CIM, the CIM becomes your assumption. By the time it reaches you, the original exaggeration looks like an established fact with three layers of institutional confidence on top.
Find the layer where the value actually lives
The most common mistake is treating "is the AI real" as a question about the whole system. It isn't. Ask it layer by layer.
In almost every modern AI product the model itself is bought. It is an API call to a frontier lab, and that decision is obvious and uninteresting. The real question is which specific surrounding layer encodes something proprietary. Sometimes the value sits in retrieval, meaning how the system decides what information counts as relevant for this particular domain. Sometimes it sits downstream in the learning layer, the part that is supposed to get better as it sees more outcomes. Sometimes it sits in the data itself. The model was bought in every case. What you are paying a premium for is the one layer that holds something no vendor has.
This is where the demo misleads you. The bought layer, the model, gives the fastest satisfaction and the most impressive show, because frontier labs spent billions making it impressive. Build attention and demo polish flow toward the visible, already-solved layer while the unglamorous layer where defensible value would live stays thin. When you see a target that has poured its engineering into the model and the interface and treats its data and retrieval as plumbing, you are very likely looking at a polished product that is trivial to replicate and defensible against no one.
“The model was bought in every case. What you are paying a premium for is the one layer that holds something no vendor has.”
The strongest moat is also your biggest integration risk
The deepest proprietary value often resists being written down at all, and even forced into words it frequently still doesn't transfer to a model. Not because the model is weak, but 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 an acquirer this cuts both ways. It is the strongest possible signal that the moat is genuine, because it structurally resists the copying that commoditisation depends on. It is also a blunt flag that the asset cannot be documented, packaged, or transferred, because it lives in specific people. That belongs in your structure, not just your price. A real but irreducible moat held by a handful of individuals is a retention and earn-out problem before it is anything else.
The perception gap, and the checklist that hides it
The data room will hand you a capability matrix and invite you to tick boxes. Resist it. A feature audit can only ever confirm gaps in what exists. It can never tell you the AI is solving the wrong problem, or isn't there, because it takes the target's framing as given.
Run the structured challenge instead. Ask why. Ask how they define good output and how they measure it, because a target that can't articulate what good looks like has no way to prove its system is improving and neither will you. Ask to watch the system fail and listen to how they explain the failure.
The sharpest tell is a perception gap. When the target's leadership and the people doing the actual work disagree about what the valuable thing even is, there is no moat to buy. Leadership has been selling a story it does not itself understand, and any value in the building is pointed somewhere other than where the deck says it is.
The read most technical diligence misses
Here is the part standard technology diligence will not give you, because it looks at code and infrastructure and stops there.
The value of any AI capability is gated by adoption, and adoption is gated by trust. If the real moat is tacit expertise held by people, and the target rolled AI across its organisation while dodging the honest conversation about what that meant for jobs, those people did the rational thing. They protected themselves. They used the tools shallowly and they did not surface the workflow knowledge that would make the system genuinely powerful, because that knowledge is their leverage. You cannot extract value from people you are lying to, and you certainly cannot extract it from people heading for the exit the day the deal closes.
“You cannot extract value from people you are lying to, and you certainly cannot extract it from people heading for the exit the day the deal closes.”
So the deepest diligence question is not technical at all. It is whether the people who hold the un-buyable knowledge trust the company enough to stay and keep feeding it. There is a related failure to price too. When a target "did AI" and it quietly became a cost of doing business rather than a transformation, the capability is shelfware, and the organisational conviction needed to make it real has already been spent. You would be buying the deflation, not the upside.
Turning the finding into a price
Four outcomes, and the finding maps cleanly onto each.
Invest at the premium when the pause is real, the proprietary layer holds knowledge no vendor has, and the people who hold it have reasons to stay. The AI line in the model is doing honest work.
Reprice toward a software multiple when the answers come instantly, the model is the product, and the surrounding layers are plumbing. The premium is air. Mark it down to what a competent team could rebuild in a few quarters.
Restructure rather than repricing when the value is genuine but lives in a few irreducible heads. Move that value out of the upfront number and into retention and earn-out, so you pay for the moat only if it walks in through the door each morning.
Walk, or push it into indemnities, when you find deliberate misrepresentation, a hard perception gap, or a chief executive who has already sold the board a vision and is running the process to confirm it rather than test it. The tell for that last one is a leader who treats every necessary complication you raise as an obstacle to a decision already made. When you are being managed toward a conclusion, the diligence is already lost.
When this is worth a second opinion
Bring in an independent technical read whenever the AI premium is large enough to move the deal, the people in the room can answer every "why" without hesitation, or the value depends on a self-improving system you cannot yet see improving.
Related insights: What happens after you discover the AI isn’t real — the reprice-or-walk decision in full; and The investor’s technical due diligence playbook — the assessment that surfaces the finding.
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.