AI due diligence vs technical due diligence: the third workstream in an AI deal

The deal is live, the data room is open and the workstreams are being lettered. Legal diligence has a firm. Technical diligence has a firm. The valuation rests on an AI claim, and the question on the scoping call is whether the tech DD firm already covers it or whether a third line item is needed.

A third line item is needed. Technical due diligence checks security and architecture, legal due diligence checks contracts, and neither is scoped to test whether the model does what the deck says. AI due diligence is a separate workstream with its own scope, its own qualified assessor, its own place in the timetable, its own report and its own set of outcomes. The rest of this piece sets out each of those in turn, for the reader who has to decide whether to commission it.

From the investor's side, diligence exists so that no unexpected risks are hiding away. For a software target, legal and technical diligence are the two workstreams instructed first. Each arrives with a brief refined over many transactions, a known deliverable and a firm that has produced it before.

Tech DD asks about the target's information security posture, how well built the codebase is and how well architected the product is. A published checklist from The Art of CTO scopes the exercise across five areas: codebase health, architecture maturity, security posture, vendor risk and technical debt. Each area has its own evidence requests and its own readout. None of the five carries an item on whether a model performs as claimed. That describes the scope of the exercise, and it is an accurate description of what the buyer is paying for.

Legal DD asks what contracts have been signed and what they commit the company to. Customer agreements, supplier terms, IP assignments, employment contracts, licences, litigation and regulatory correspondence all sit inside the brief. The lawyers read what has been agreed and report what it means for the buyer after completion.

A defined scope is what makes a diligence report usable. The investment committee reads a tech DD report knowing what was examined and to what depth, and reads the legal report the same way. A scope that stops at a stated boundary is a feature of good diligence. The boundary is also where the AI claim sits.

A target can come through tech DD with a clean report: a well-structured codebase, sensible architecture, infrastructure that scales, no critical vulnerabilities. The same target can have an AI claim at the centre of its valuation that nobody has tested. The two findings are compatible, because the first concerns how the product is built and the second concerns whether the intelligence inside it exists in the form described.

Papermark's guide to AI due diligence states the split in two sentences. Technical review asks whether the product can carry the deal thesis, meaning architecture, code quality, security and scalability. AI review asks whether the intelligence inside the product is owned, legal, economic and reproducible. Papermark adds a point about scope rather than skill: a technical team will not read customer contracts for no-training clauses, or reconcile dataset manifests against training-run logs, unless the brief names those tasks.

Owned means the models and the data they were trained on belong to the company rather than to a vendor whose terms could change. Legal means the training data was licensed for that use. Economic means the cost of each inference call is known, including how much of the product's margin routes through a third-party model. Reproducible means the reported performance can be regenerated on data the company has not seen, by someone who does not work for the company. Each is a question about the AI itself, and a code review answers none of them.

The investor's question, as Colin Kelly puts it to clients: "this all sounds good in principle, but how good is this AI? How much of it is made up? How much is a wrapper around Anthropic or ChatGPT, and how much is real?" Those are three separate questions, and each has a method behind it. The tiers and red flags that sort a wrapper from a model are in the thin-wrapper piece.

Foley & Lardner's October 2026 note on AI and autonomy deals lists what buyers now examine: proprietary data as a primary asset, patent characterisation, technical authenticity, regulatory exposure, cybersecurity and training-data rights. On technical authenticity, the note says diligence "must go beyond reviewing marketing materials and pitch decks". It is a thorough list of what to look at. It assigns none of the items to a workstream.

Map the list onto the two briefs already instructed. Cybersecurity sits inside tech DD. Patent characterisation and the regulatory mapping sit largely with the lawyers. The remaining items fall between the two. Whether the data is proprietary is a question about the data and the models trained on it. Whether the training-data rights hold starts in a contract and ends in a training-run log. Technical authenticity is a question about model performance against a held-out set. Each needs an assessor who reads both the contract and the model, and neither existing brief asks for one.

Without a third workstream, those items land in the tech DD report. The firm writing it was briefed on the codebase, and a diligent team will record what it saw of the AI through that lens. What a codebase review surfaces is that model files exist, the inference service runs and a third-party provider's keys are in the configuration. That is the paragraph. It records that AI is present in the product. It cannot record whether the AI performs as claimed, because performance testing was outside the brief and outside the time allotted.

A tech DD firm that stays inside its brief has done the job it was engaged for. The gap exists because the brief was drawn around the codebase and the AI claim sits outside that boundary. Closing it means drawing a second brief, with its own line in the engagement letter and its own assessor, rather than asking the first firm to stretch.

In March 2024 the SEC settled charges against Delphia (USA) Inc. and Global Predictions Inc. The SEC found that Delphia "did not in fact have the AI and machine learning capabilities that it claimed", and that Global Predictions had falsely claimed to be the "first regulated AI financial advisor" offering "expert AI-driven forecasts". The two firms paid $400,000 in penalties between them, settling without admitting or denying the findings, and the SEC Chair's statement on the action was that "such AI washing hurts investors".

The Wall Street Journal reported in August 2019 that human engineers did most of the coding at Builder.ai, against the company's account of an AI-driven development process. The company went on to raise $230m in 2023 from investors including Microsoft and the Qatar Investment Authority, as Bloomberg reported, and entered insolvency in May 2025. The playbook tells that story in full, and the five questions the AI workstream has to answer are in the playbook.

The 2019 report was public four years before the 2023 round and six years before the collapse. The useful question for a reader scoping their own transaction is which workstream, as lettered, would have been scoped to pick it up. A legal review reads contracts. A technical review reads the codebase. Taking a public report that the AI does not do the work, designing a test to check and running it against the live product is an AI review task. In a deal with two workstreams, it belongs to nobody.

Testing a model claim is hands-on work: holding out data the company has not seen, running the model against it, reading training logs, tracing the inference path and pricing it. The assessor who does this has trained and evaluated models as a practitioner, because there is no desk-based substitute for the exercise. That is the qualification test, and it applies whichever firm is being considered. The incentives question, and why an independent assessor is a purchase of scepticism rather than reassurance, is covered in the Big 4 piece.

The investor or acquirer holds the budget. The company under assessment does not pay for the work and does not set its scope. A report commissioned by the party whose claim is being tested carries the wrong signature for an investment committee, however careful the work behind it.

The AI workstream runs before deep financial modelling, because the model inherits the AI claim. Unit economics assume a gross margin, and if a share of inference routes through a hosted foundation model the margin assumption moves. The thin-wrapper piece makes the case that the technical review comes first. In shape, the engagement is time-boxed: a two-week core assessment, sitting inside a two-to-six-week window set by the deal timetable rather than by the assessor.

Both sides of a transaction have a reason to commission the work. A company raising or selling wants external validation on the record before the buyer's questions start. An acquirer wants to know what is real before the price is set. The independence rule above holds in either case: the party paying is the one who needs the answer to stand up to scrutiny.

The output is a report that ranks areas of concern and areas of strength across a few dimensions: what to mitigate and what to worry about. It is written for the investment committee rather than for the engineering team. The engineers will get the detail, but the document that decides the deal has to be readable by the people who decide it, with the dimensions named and each one scored against evidence the assessor produced rather than evidence the company supplied.

The ranking matters because the findings rarely point one way. A target can own its models and have clean training-data rights while its reported accuracy does not hold on a held-out set. Another can route most inference through a vendor while the thing it has built around that vendor is defensible. The report says which findings bear on price, which bear on structure and which are operational matters for after completion. What each finding does to price and structure is in the decision framework.

The workstream ends one of four ways. The deal proceeds as priced. The deal is re-priced to reflect what the AI is worth rather than what it was described as. Conditions are set, such as a retraining programme or a warranty on the training data, with completion tied to them. Or the acquirer walks away. A binary verdict on whether the AI is real would be the wrong deliverable, because most AI claims are partly true and the deal needs to know which part.

The practical step is to add the line before the engagement letters go out. It reads: AI due diligence, independent assessment of the AI claim, two-week core assessment, report to the investment committee. The scoped engagement sets out what those two weeks cover and what the report contains. The first data request to the target can be the checklist, which puts the questions in the target's hands on day one of the timetable rather than week three.

How does AI due diligence differ from classic tech due diligence? Tech DD assesses how the product is built: codebase health, architecture, security posture, vendor risk and technical debt. AI due diligence assesses whether the intelligence inside the product is owned, legal, economic and reproducible, which means testing the model against held-out data, reading the training-data rights and pricing the inference path. The two can both be clean, or one can be clean while the other is not, because they examine different things.

Which consulting firms have genuine expertise in AI due diligence? The test is whether the people doing the work have trained and evaluated models themselves, because checking a model claim means running the model on data the company has not seen and reading its training logs. A firm whose diligence practice grew out of code review or financial modelling can describe the AI, but describing it is what the pitch deck already does. Ask who on the team will run the evaluation and what they have evaluated before.

Who specialises in technology due diligence for AI products? Independent assessors with a practitioner background in model development, engaged directly by the investor or acquirer, run the AI workstream. General tech DD firms cover the codebase and infrastructure of an AI product and should be kept for that. The AI claim itself needs a separate brief and a separate signature on the report. Agathon runs this workstream for investors and acquirers as a two-week core assessment.

How long does AI due diligence take and who pays for it? The core assessment is time-boxed to two weeks, inside a two-to-six-week window set by the deal timetable. The investor or acquirer holds the budget and sets the scope, and the company being assessed does not pay, so that the report stands up to scrutiny at the investment committee.

Does a company raising or selling benefit from commissioning its own AI due diligence? Yes, provided the assessor is independent and the report is on the record before the buyer's questions begin. External validation shortens the buyer's workstream and removes the discount that an untested claim attracts. The findings are the same either way; the difference is who hears them first.

An AI claim that has been tested is worth more to a deal than one that has been assumed, whichever side of the table is paying. Agathon's scoped engagement is a two-week core assessment of the AI claim, run by a practitioner who has trained and evaluated models, with a report written for the investment committee and one of four outcomes at the end. The place to start is the third line in the engagement letter, written before the workstreams are lettered.

• foley.com: what buyers now diligence in ai and autonomy deals lessons from apples ai acquisition
• theartofcto.com: technology due diligence checklist cto guide
• papermark.com: ai due diligence
• 'Still in shock': Builder.ai staff on the collapse into insolvency (Tech.eu)
• sec.gov: press release 2024-36, SEC charges two investment advisers with making false and misleading statements about their use of artificial intelligence (Delphia and Global Predictions, March 2024)