AI Due Diligence: Pricing AI Risk and Upside in a Target Before You Close

The market is converging on similar checklists. Bain offers five questions, other advisors propose four axes of AI vulnerability, and the legal field now describes AI diligence as the new IP diligence. The framework below consolidates them into six dimensions that together cover both the liability a buyer might inherit and the asset a buyer might be paying for. Used as a structured question bank rather than a vague unease, it turns a feeling about a target's AI into specific, priceable findings.

Two habits make the framework work. The first is to read the dimensions together rather than in isolation, because they compound: heavy model dependency is more dangerous in a company that also has thin governance, and a brilliant data moat is worth less if the rights to that data cannot be cleanly transferred. The second is to score each dimension against evidence rather than assertion, since the entire point of the exercise is to value what is documented and verifiable instead of what management describes from a slide. The same six dimensions, read from the other side of the table, are exactly what a seller should put in order before a process begins.

Ownership comes first because it decides whether anything else can be valued as an asset at all. The pillar argued that AI earns a premium only when it is owned outright. Diligence is where that ownership is either verified or quietly falls apart. The traditional questions still apply: chain of title, defensible patents, and signed IP assignments from every engineer and contractor. AI adds a layer on top. The license terms for the models in a target's stack do not automatically transfer on a change of control, and many open-source model licenses carry attribution, copyleft, or commercial-use conditions the company has never fully tracked. A rigorous review surfaces every model license in the stack, not just the one on the top-level product, because a single restrictive license buried three layers down can dictate how the acquirer is allowed to deploy the very thing it just bought.

Data is where ownership becomes concrete, and where the sharpest new liability lives. Training data provenance has become material to valuations because of how expensive it is to get wrong. In Bartz v. Anthropic, the use and retention of pirated books in model training led to a settlement of 1.5 billion dollars, on the order of 3,000 dollars per work, and the case crystallized three points that now sit in every careful diligence checklist: sourcing matters, retention matters, and discontinuing a bad practice does not erase the historical exposure. So a buyer inventories the training, fine-tuning, and evaluation datasets, traces each one to its source and license, and confirms there were no shadow libraries or prohibited scraping. The same review weighs the upside, because a proprietary, lawfully sourced dataset that competitors cannot reproduce is precisely the kind of data moat that earns a premium rather than a discount. Privacy sits inside the same review, because models can memorize and later surface sensitive inputs, so a careful buyer maps how personal and customer data flows through both training and inference and asks whether the consent and retention practices behind it would survive a regulator's attention.

Most companies build on someone else's foundation model, and there is nothing wrong with that until it becomes a concentration risk. A target whose product is a thin layer over a single external API inherits that provider's pricing, deprecation schedule, and acceptable-use policy, none of which it controls. Bain's questions about pricing pressure apply directly here: a per-seat or cost-plus model can crack when the capability underneath it is a commodity the customer could call on its own. The diligence question is not whether the company uses third-party models, but whether it has a deliberate strategy around them: portability across providers, ownership of the components that genuinely differentiate the product, and a clear, managed view of how model costs scale with usage. Dependency that has been thought about is manageable. Dependency that has been ignored is a discount waiting to be applied.

AI capability has a habit of concentrating in a small number of people, and that concentration is a genuine risk to the asset. FTI Consulting has identified talent, not capital, as the primary constraint on scaling AI, and in a smaller company the practical form of that constraint is that one or two people understand how the models were built and why they work. If those people leave after close, the institutional knowledge can leave with them, and the acquirer is left holding systems that no one remaining can safely change. Diligence tests whether the knowledge is documented and distributed rather than tribal, and because integration planning belongs in diligence and not after it, retention bonuses, earn-out incentives, and equity for the people who carry the knowledge are standard tools, used precisely because in an AI business the knowledge is much of what is being bought.

Governance is the dimension most likely to be missing entirely in a target, and its absence is both a legal exposure and a cost the acquirer inherits. AI systems make decisions that affect real people, and when a model produces a discriminatory result or a harmful output, someone is accountable. After an acquisition, that someone is frequently the acquirer. The regulatory backdrop is live rather than theoretical. The EU AI Act is already enforceable for prohibited practices and for general-purpose model obligations, with fines reaching 35 million euros or 7% of global turnover. Its heaviest obligations for high-risk systems were deferred under the 2026 Digital Omnibus agreement to late 2027, but the risk-based architecture and the underlying liability are not going away, and parallel rules such as Colorado's AI Act are arriving on their own timelines. Diligence looks for a documented governance program, human oversight in high-stakes applications, evidence of bias audits and red-teaming, and incident logs. A company that has built none of this presents post-close work whose time and cost belong squarely in the deal economics. In regulated sectors the scope only widens, with HIPAA business-associate terms, FDA history for any model treated as a medical device, and a growing set of state health-privacy statutes each adding a diligence thread of its own.

The final dimension is architectural, and it is often the one that separates an asset from a wrapper. Bolted-on AI, added as a feature without rethinking the system underneath, tends to carry technical debt, resist integration, and fail in ways that are expensive to repair. Modular, portable AI woven into core workflows is the opposite, and it is hard for a competitor to dislodge precisely because it is entangled with how the business actually runs. Bain's own diligence work shows the contrast cleanly. On one engagement the team built a prototype of an AI-native healthcare target's technology from the outside, concluded it could be challenged by incumbents and new entrants, and the sponsor walked. On another, a specialty workflow software company made the better case: AI was embedded in customer workflows with high switching costs, the data and workflow moats were defensible, and management treated AI as a board-level priority. Same diligence question, opposite answers, and in both the architecture was where the answer lived.