How to make AI-assisted investment diligence auditable

The objection to using AI in diligence is rarely about speed. Everyone accepts that a system can read a data room faster than an associate. The objection is about accountability. When a finding appears in an investment paper, someone has to stand behind it, and “the model said so” is not something a partner can say to a committee, an LP or a regulator.

That makes auditability the real design problem. An auditable diligence process is one where a reviewer who was not involved can pick any statement in the output and establish, without asking anyone, where it came from, who checked it and what has changed since. These are the five properties we look for.

1. Every finding points to its source

A finding such as “the ten largest customers account for 46% of revenue” should carry a reference to the specific document, page or spreadsheet tab it was drawn from, not a general pointer to the data room. If a number is calculated rather than quoted, the inputs and the calculation should be visible. The test is simple: can a reviewer get from the sentence to the evidence in one step?

The reverse matters as much. Where the system cannot find support for a claim made in a management presentation, it should say so. “No supporting document found” is a useful finding. A confident sentence with nothing behind it is the failure to design against.

2. Facts, calculations and judgements are kept apart

A diligence report contains three different kinds of statement. An extracted fact: the contract has a twelve-month termination clause. A calculation: gross margin fell by two points. A judgement: customer concentration is a material risk. Each can be wrong in a different way and each needs a different check. Labelling them lets reviewers spend their time where it counts, which is on the judgements.

3. Human review is recorded, not assumed

“Human in the loop” means little unless the loop leaves a trace. For each finding, the record should show who reviewed it, when, and what they changed. Findings nobody has reviewed should be visibly unreviewed, and that status should travel with the finding into whatever document it ends up in. An investment paper that mixes checked and unchecked material without distinguishing them is worse than one written by hand.

4. The record is versioned

Data rooms change. Management uploads revised accounts in week three; a contract schedule is replaced. An auditable process records which version of which document each finding relied on, and flags findings whose sources have since been superseded. Without this, a paper can quietly rest on numbers that no longer exist.

5. Access is controlled and logged

Diligence material is confidential and usually covered by a non-disclosure agreement. Access should be limited to the deal team, and both access and actions should be logged. Where the system runs, where documents are stored and which models process them should be agreed and written down before the first file is uploaded, not reconstructed afterwards.

What this changes

When these properties hold, the question shifts from “do we trust the AI?” to “have we checked the things that matter?” Investment teams already know how to answer the second question. Review effort goes to the judgements and the unsupported claims, and the audit trail becomes a by-product of doing the work rather than a document someone assembles at the end.

These are the principles behind InGen, our product for investment screening and diligence, and they are described in more detail on our governance page.