data-transactions
AI due diligence is the use of models to scan, extract and flag issues across a deal file — financial anomalies, missing documents, contract clauses — while humans still own scope, materiality and conclusions.
AI due diligence is not a sixth specialist stream beside financial, commercial, legal, tax and operational. It is a set of tools those streams can use: anomaly detection on ledgers, clause extraction, completeness scoring, clustering of Q&A. Invest Europe professional standards and AICPA consulting standards still put judgement — and liability — on the appointed advisor.
The quality ceiling is the data room. Models cannot diligence a file that was never uploaded. They also cannot replace commercial due diligence interviews.
A human-written scope. A structured corpus (ideally a smart data room). Flags with citations back to the source page. A workflow where a financial due diligence manager accepts or rejects each flag. An explicit list of what was not in scope for the model.
The model flags revenue spikes on the last day of each quarter and a cluster of related-party payments. The FDD team already suspected cut-off; the AI shortened the sample. It also flagged a “change of control” in a facilities agreement that counsel had missed in a 90-page scan. It then “concluded” there was no tax risk because the tax folder had two files. The last output is malpractice if taken as a tax due diligence opinion.
Buyers like speed if the citations are real. Sellers should insist that AI-generated questions still go through the Q&A log so disclosure is complete. Nobody should accept a black-box “risk score” as a price chip.
Skipping specialist streams because “we ran AI.” No materiality. Training or prompting on one deal’s confidential data inside another deal’s environment.
Fraud designed to look normal can beat a model the same way it beats a tired associate. Forward-looking commercial claims are not a classification problem.
AI diligence is most honest as a reader of a structured layer already used for valuation and matching. That is the Dazur sequence — structure first, then intelligence — rather than dropping a chatbot on a ZIP of PDFs.