You are running technical diligence on an AI-enabled target. Finance and legal are handled. The question only you can answer is whether the AI is real, and if it is real, whether it is the kind of asset that survives the founders leaving twelve months after close. Those are two different findings, and the second is the one that quietly resets price.
Most diligence on AI washing stops at the first: is the capability as described, or is it marketing over a third-party API. Worth checking, and usually answerable in a day. The more expensive finding is the inverse. The AI is genuinely good, the margin is real, and the value turns out to live in the heads of a few people rather than in anything you are acquiring. That is not a reason to walk. It is a reason to reprice, and the instrument is the key-person discount.
Start with the pause
The fastest read on a target's AI does not come from the architecture diagram. It comes from a management session. Put the domain expert in the room, point at a decision the system makes, and ask why it decides that way. Then watch how the answer arrives.
An instant, crisp answer means it is an empirical fact, codifiable, and a vendor has almost certainly already encoded it. That is a buy for the market, which means it is not a moat for this target. If every "why" is answered without hesitation, the proprietary layer they are charging you for is a rules engine wearing an AI label.
The pause is the opposite signal. The expert reaches, draws on years of doing the thing, and gives you a judgement rather than a rule. That hesitation is the moat. The value is tacit and accumulated, which is exactly what no competitor has and no off-the-shelf model can guess.
“The pause is the moat, but in a deal the same pause is your key-person risk, because the asset just told you it lives in a person's head.”
Here is the turn that diligence usually misses. In a build-or-buy decision the pause is good news. In an acquisition it is two findings at once. The same pause that proves the moat is real also tells you where the moat physically sits. If the answer lives in that expert and has never been encoded anywhere you are buying, the moat walks out with them.
The capability-audit handoff
Targets running an AI narrative tend to hand you the same artefact: a capability deck. A grid of tools and models with ticks, a list of features shipped, seat counts, usage curves. It looks rigorous. It is structurally incapable of answering your question, because it takes the premise as given and only ever returns "yes, the boxes are ticked."
Watch for two specific substitutions. First, the licence count presented as adoption. Broad rollout of an AI seat to the whole company produces a number that goes up while the actual value, which lives in two or three re-engineered workflows, goes untouched. A high seat count with shallow use (summarise this, draft that) is the appearance of doing AI, not evidence of a defensible product. Second, the UI presented as the asset. Executives steer AI builds by the part they can see. So do the decks. A polished interface over a thin value layer is trivial to replicate and defensible against no one, and eighteen months post-close that becomes your problem.
Where the value lives: the system, or three people
Treat "is the AI real" as a question asked layer by layer, not about the whole product. In almost every modern AI system the model itself is bought, an API call, off the shelf. That part is uninteresting and not where any moat lives. The interesting question is which specific layer holds something proprietary, and whether that layer is encoded or embodied.
There is a trap built into how targets present this. The bought model layer gives fast, visible satisfaction, so build energy and demo time flow toward the layer that is already solved, while the unglamorous layer where defensible value actually sits (the data, the retrieval logic, the domain mapping, the way "relevant" gets defined for this specific problem) gets treated as plumbing. A target that has poured its engineering into the model layer and rushed the rest has built the easy thing and starved the hard one.
When you do find a genuinely proprietary layer, the deepest ones resist articulation entirely. The value was never in any single rule. It lived in the interaction of many tacit judgements that do not decompose into instructions, which is why even when the founder tries to write it down it still does not transfer to the model. That is the strongest possible evidence the moat is real. It is also the strongest possible key-person flag, because an asset that cannot be specified cannot be documented, transferred, or retained with a standard earnout.
“An asset that cannot be written down cannot be retained with an earnout, which is precisely why it reprices the deal rather than closing it.”
A spectrum for classifying the claim
To keep the conversation honest, force every "AI" claim onto a spectrum and make the target tell you which rung they are on.
AI-native: the product cannot exist without the model, and a proprietary layer around it encodes something no vendor has. AI-augmented: a real workflow with model-driven features that materially change the output. AI-enabled: a conventional product with model calls bolted on, where removing them would barely change the value. AI-badged: marketing, a rules engine or a thin API wrapper relabelled. Targets price toward native and usually sit a rung or two lower. Your job is to place them by evidence, not by deck.
How to spot the washing in the data room
A few requests separate capability from claim quickly. Ask for the proprietary training or evaluation data and who controls it. Ask how they define "good" output, because a target that cannot articulate what good looks like cannot evaluate its own system, and any self-improvement story sitting on that is noise dressed as signal. Ask to see the layer they call self-learning, and check whether the data volume behind it is anywhere near enough to produce a valid signal rather than months of accumulated low-quality feedback. Probe the gap between what leadership claims the AI does and what the engineers say it does. That perception gap is its own finding: when the people running the company and the people building it disagree about what the valuable thing even is, the moat they are selling you is aimed at a misconception of itself.
The contrarian take: the washing that matters most is internal
The AI washing you should fear in a deal is not the target lying to you. Your diligence is built to catch that. The expensive version is the target lying to itself, and you inheriting the gap.
Internal AI washing is the under-priced risk. Teams overclaim capability upward, to their own board and their own existing investors, because a number that goes up is politically safer than the truth that the deep value sits in a handful of people. By the time you arrive, that overclaim is baked into the equity story and the multiple. The genuinely valuable, genuinely tacit work has been recast as scalable software IP, because nobody inside wanted to say out loud that the company is closer to a consultancy than a platform.
This is where the deal model breaks if you let it. The moat, where it is real, lives in tacit expertise held by people. You cannot extract that value from people who learn at close that the model assumed they were a line item. The ones who hold the irreplaceable judgement do the rational thing: they protect it, coast through the retention period, and leave with the moat in their heads. Evasion does not avoid the cost. It converts a one-time pricing conversation into a permanent erosion of the asset you paid for.
“You are not pricing retention risk. You are pricing the chance that you bought a consultancy and paid a platform multiple.”
From finding to price
The key-person discount is the translation. It is not a generic haircut for "founder risk." It is the measured distance between the multiple on the table and the durability of the thing actually generating the margin. The diligence above tells you which case you are in. If the AI is badged commodity, you reprice for a thinner asset or pass. If the AI is real and encoded in systems and data you are acquiring, the moat travels with the deal and the price holds. If the AI is real but embodied in three people who answered every important "why" with a pause, you are buying judgement, not software, and the structure has to reflect that: discount, retention tied to genuine transfer rather than time served, and an honest internal account of which roles hold the value.
This is worth a second, independent technical read whenever the margin depends on AI that no one on your side of the table can yet separate from the people who built it.
Related insight: The investor’s technical due diligence playbook — where the people-vs-system read fits in a full technical diligence.
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.