AI is showing up in more deal theses than diligence processes are built to evaluate. A target company’s engineering team can describe a model’s architecture fluently and still not know whether it holds up under production load, adversarial input, or a regulator’s request for an audit trail. The gap is not malicious. Most technical due diligence checklists were written for software, and AI changes enough of the underlying risk that the difference shows up in the first year of ownership, not the data room.
The Lion’s View applies a four-dimension framework in PE and institutional engagements: technical credibility, data and model dependencies, governance and regulatory posture, and commercial execution. Each dimension surfaces a different way an AI or XR investment can look sound on paper and prove fragile once it is inside the portfolio.
The Four Diligence Dimensions
Technical credibility: demonstration versus production
Every AI vendor can show a model performing well in a demonstration; demonstrations are built to succeed. The diligence question is whether the same model performs at the volume, latency, and edge-case exposure the target market will actually generate. The Lion’s View tests this by requesting the model’s performance against data it did not train on, not the benchmark the company chose to present. A target that resists an independent test, or whose only validation comes from its own engineering team, is telling the diligence team something about the boundaries of the technology’s actual state. Architecture dependency matters here too: a product built as a thin layer over a single foundation-model API carries different risk than one with defensible technical differentiation, and the multiple should reflect which one is in the room.
Data and model dependencies
Where the training and inference data came from, and whether the company holds documented, defensible rights to it, is a question that belongs in legal diligence as much as technical diligence. A model that performs well today can lose that performance overnight if a third-party data source changes its terms, or if a foundation-model provider deprecates the version the product depends on. The Lion’s View looks for a documented lineage from raw input to model output. Its absence is common, and it is not disqualifying on its own, but it changes what the buyer is underwriting, from a proven system to an unverified one, and the price should reflect that.
Governance and regulatory posture
A target positioned for today’s regulatory environment is not necessarily positioned for the one it will operate under in twelve to twenty-four months. Risk-tiering obligations under the EU AI Act, sector-specific rules in healthcare and financial services, and the general direction of U.S. federal and state AI policy all bear on whether a company’s current governance posture is an asset or a liability at exit. Board-level oversight of AI risk is the clearest signal: a company where AI risk reporting reaches the board, with a named owner for model risk and drift monitoring, is further along than one where oversight stops at the engineering team, regardless of what the deck says about “responsible AI.”
Commercial execution
The pipeline should be read for what it actually contains: paying customers in production, not pilots that have not converted. Gross margin claims deserve scrutiny once support and compute costs are fully loaded, since AI products carry ongoing inference costs that many software diligence models do not account for. Revenue concentrated in a small number of design partners who received favorable terms to become reference customers is a different risk profile than revenue distributed across a real customer base, and the go-to-market model should be evaluated against the buyer the investment thesis is actually targeting.
Taken together, the four dimensions answer one question a term sheet cannot: whether the company being acquired is the company being described. Technical credibility tests the product, data and model dependencies test the supply chain behind it, governance and regulatory posture test how the company will hold up under future scrutiny, and commercial execution tests whether the revenue is durable. A target can pass three of the four and still carry enough risk in the fourth to change the price.
What the Data Room Doesn’t Show
The materials a target prepares for a raise are built to present the company in its best light, and AI targets have more places to shade the picture than most software companies do. Model performance figures are typically drawn from the company’s own benchmark, run once, under favorable conditions, not from ongoing production monitoring, which would show whether accuracy degrades as real-world data drifts from the training distribution. Compute cost is often presented as a moment-in-time unit economics figure rather than the trajectory that scale will actually produce. Governance maturity is described in policy documents that may not reflect what engineering teams do day to day. None of this is unique to bad-faith actors; it is the natural result of a data room built to close a round, not to survive an operating review. Independent verification is what closes the distance between the two.
Three items rarely appear in a standard data room at all: a record of model performance drift over the trailing twelve months, a documented incident history for any AI system already in production, and a clear answer to what changes in the unit economics if the foundation-model provider raises API pricing. Investors who ask for these three items directly, before the diligence sprint begins, tend to learn more from how the request is received than from the documents that come back.
How Findings Translate to Deal Terms
A diligence finding is only useful if it changes something in the transaction. Technical or data-dependency risk that cannot be resolved before close typically shows up as a price adjustment, an escrow holdback tied to a defined remediation period, or a specific indemnification carve-out rather than a general one. Governance gaps are better addressed through covenants than through price: a ninety-day post-close requirement to stand up model risk ownership and board-level AI reporting costs little and closes the gap the diligence process identified. Where commercial execution is the open question, an earnout tied to production conversion, not booked pipeline, aligns the price with the risk the diligence team actually found.
Red Flags by Deal Stage
Seed and Series A
- The product is a thin wrapper around a single foundation-model API with no defensible technical differentiation.
- Founders cannot answer a direct question about data provenance or rights to the training data.
- Model performance claims rest entirely on a benchmark the company selected and ran itself.
- No one on the founding team owns responsibility for model risk, even informally.
Growth equity
- A low pilot-to-production conversion rate, with revenue projections that assume conversion will improve without a stated reason why.
- Data pipeline sprawl across multiple undocumented sources as the company scaled faster than its infrastructure discipline.
- Governance built for a smaller company that has not been resourced to match current headcount or customer base.
- Customer references concentrated among a handful of design partners who were never charged full price.
Buyout and platform acquisition
- Governance debt accumulated across multiple AI systems built at different times, by different teams, with no shared risk framework.
- Regulatory exposure that was not priced into the multiple, particularly in healthcare, financial services, or public-sector-adjacent markets.
- Integration risk across AI systems inherited through prior acquisitions, each with its own data lineage and vendor dependencies.
- A board that receives no AI risk reporting at all, despite AI being material to the company’s valuation.
When Independent Assessment Pays for Itself
A diligence sprint costs a fraction of a percent of most transaction values. What it protects against is a valuation built on an AI capability that does not hold up, a regulatory exposure priced at zero, or a technical dependency that surfaces eighteen months after close, when the room to renegotiate is gone. The value of independent assessment is structural, not incidental: The Lion’s View holds no platform to sell and no vendor relationship to protect, which means the technology and commercial assessment is not shaped by an incentive to recommend a build, a migration, or a specific vendor relationship.
The clearest signal that assessment is warranted is not deal size. It is how confidently the target can answer a direct technical question without reaching for the deck. A five-minute conversation usually tells an experienced diligence team more than a hundred-page data room, and it tells them what to go verify.
The AI & XR Due Diligence Checklist applies this framework directly to a specific target, question by question. Firms evaluating a diligence engagement can review how The Lion’s View structures this work on the Investor Advisory page.
For a sector-specific application of this framework, see Healthcare AI Diligence Beyond the Demo.