AI in M&A Due Diligence: Faster, Deeper, Riskier?

AI can help search and organise a transaction's documents. Whether it improves due diligence depends on what it misses, how findings are checked and the effort needed to produce a defensible report. Document-processing speed alone does not answer that question.

What the research establishes

ContractNLI, published by Koreeda and Manning in 2021, tested whether statements were supported, contradicted or not addressed in a collection of non-disclosure agreements. It also required identification of supporting passages. The evaluated models struggled, including where negation was expressed through exceptions. This is evidence for testing these failure modes, not a current performance score for commercial M&A platforms.

A benchmark built around one contract type does not establish competence across a multilingual Swiss data room. Ask a supplier to demonstrate the actual tasks and languages you need. Clause extraction, interpretation of a clause and assessment of its transaction consequences are separate tasks.

Define the review before selecting the tool

Specify the transaction structure, materiality criteria and questions the review must answer. These may concern assignment, change of control, termination, intellectual property, employees or data processing. A generated list of clauses is not a complete statement of legal risk.

Identify the permitted documents and users before upload. Check professional secrecy, data protection, contractual restrictions, subcontractors, training reuse and retention. For lawyers subject to the Swiss Lawyers Act (BGFA), professional secrecy under Article 13 BGFA covers everything their clients entrust to them in the course of their professional activities, and they must ensure that their auxiliaries preserve it. The SAV's AI guidance, published in German, French and Italian, calls for examination of input handling and independent critical verification of results. Approval for one matter should not silently authorise access to another.

A five-stage workflow

Intake: preserve amendments and relationships between documents. Record absent or unreadable material. Clean categorisation can make a review easier to audit, but it does not guarantee model accuracy.

Test: agree how to evaluate missed issues as well as false alarms. Include exceptions, cross-references, conflicting versions and clauses whose answer is genuinely absent. Record the model, configuration and document set. A vendor demonstration selected for easy examples is insufficient evidence for your matter.

Review: do not inspect only the documents that AI flags. An omitted issue will not appear in its red-flag report. Base review depth on materiality, observed failure modes and the consequences of a miss. Set the number of contracts reviewed in detail from those factors, not from a fixed rule such as five or ten.

Legal assessment: Swiss corporate, employment, merger-control and form requirements depend on the deal. Assign those questions to suitably qualified counsel and verify the applicable rules. Do not assume every tool has the same training data or that a contract benchmark establishes knowledge of Swiss mandatory law.

Delivery: trace material conclusions to the underlying documents and authorities. Distinguish an extracted passage from the lawyer's interpretation and from an unresolved question. Keep the evidence needed to reproduce significant findings.

Measure the complete job

Compare an approved AI-assisted process with a meaningful baseline on comparable material. Count preparation, checking, correction and documentation as well as extraction time. Record coverage, material omissions and false alarms alongside cost. Savings or additional capacity are conclusions to test, not assumptions to put into a fee proposal.

An illustrative pilot may start with a bounded document set. Its size and acceptance criteria should follow the risks and diversity of the transaction; passing a small sample does not establish completeness across the whole data room.

Agree responsibility and client communication

Review the mandate, client instructions and applicable disclosure requirements. Explain AI's role and the verification process accurately. The CCBE guide on generative AI of 2 October 2025, available on the SAV page, recommends transparency with the client where an informed client could reasonably be expected to object to, set conditions on or have reservations about the use of generative AI for the purpose in question. A general disclosure sentence does not itself permit confidential data to be sent to any provider. Clarify responsibility for the deliverable and any relevant insurance terms with the actual policy and insurer, without assuming all policies exclude or cover AI-assisted work.

Planning AI-assisted transaction work? Get in touch to discuss the review method and evidence your team needs.