How to conduct due diligence with AI

Define the decision and failure conditions

Write what decision the diligence supports, the scope, review period, materiality rules, owners, and what would stop or change the transaction or initiative. Turn broad topics into answerable questions. “Review the company” is unusable; “Can the stated recurring revenue be reconciled to contracts and collections for the review period?” creates an evidence test.

Build a source map

Inventory documents by source, date, version, authority, question served, and access status. For public-company research, SEC EDGAR provides official filing access, and the SEC’s 10-K guide explains the roles of business, risk, financial, and management-discussion sections. Filings are inputs, not independent confirmation of every management statement.

Separate signed records, system exports, interviews, third-party evidence, and commentary. Note whether each file is complete and current.

Worked synthetic diligence finding

Assume a software supplier claims that 90% of revenue is recurring and that no major customer exceeds 10% of revenue. The data room contains a revenue schedule, twelve contracts, bank summaries, and a management interview.

| Question | Evidence | Exception | Status | | --- | --- | --- | --- | | Is revenue recurring? | Schedule plus contract terms | Two contracts include non-recurring setup fees | Open | | Is concentration below 10%? | Customer schedule | Customer names differ across exports | Open | | Do collections reconcile? | Bank summary | Underlying statements missing | Blocked |

A poor AI summary repeats the 90% claim. A defensible finding says the schedule reports 90%, but classification exceptions and incomplete collection evidence prevent confirmation.

Keep a contradiction log

Record statement A, statement B, sources, dates, materiality, owner, and resolution evidence. Ask AI to compare names, totals, dates, and definitions across files. Require an exact location for every flagged conflict.

Also keep an open-question log with due date and consequence. Missing evidence is not a neutral blank when it prevents a material conclusion.

Search for disconfirming evidence

For each leading conclusion, ask what would make it wrong. Sample items outside management’s examples, examine adverse periods, and compare interview claims with records. Do not let the same person’s statement and spreadsheet count as independent corroboration if both originate from one process.

Draft findings with evidence grades

Each finding should contain the question, current answer, supporting evidence, exceptions, uncertainty, decision implication, and next verification. Use confirmed, partially supported, contradicted, or unresolved according to defined rules.

Verification checklist

Confirm that scope and materiality predate conclusions; files are versioned; citations resolve; samples include adverse cases; interviews are distinguished from records; contradictions remain open until evidence resolves them; calculations reconcile; missing items have consequences; and accountable specialists review legal, financial, tax, security, or technical matters.

Keep the question log open until the recommendation is signed. A late response should be appended with its date, origin, and reviewer, then assessed for whether it changes a finding or only adds context. The final pack should preserve superseded conclusions alongside the evidence that changed them. That history lets decision-makers distinguish new facts from a changed interpretation and prevents a polished summary from erasing uncertainty.

Sources and claim map

The supplier, documents, percentages, and findings are synthetic. This educational workflow is not financial, legal, tax, investment, or transaction advice and does not establish that any diligence scope is complete.

• U.S. Securities and Exchange Commission, EDGAR company filings: supports using the official filing system as a source route for covered public-company records.
• U.S. Securities and Exchange Commission, How to Read a 10-K: supports the limited description of major 10-K information categories.

Sources: [1] [2]