AI-Assisted Due Diligence: Evidence Checks Before IC
AI-assisted due diligence is safe to bring to an investment committee only when every material claim in the draft has been traced to a specific source and graded on whether that source actually supports it. A citation next to a sentence is not proof. Research on AI systems that cite their sources shows that a large share of cited sentences are not fully supported, and that the wrong answers often look as polished as the right ones.
This post sets out what the evidence says about AI citation errors, a five-grade scale for checking claims, a register template, and a review routine that fits the week before committee.
Why a citation is not the same as support
The main risk is not an invented source. It is a real source cited for something it does not say. That error is harder to catch because the link works and the document is genuine.
Three studies show the pattern:
- Cited answers are often only partly supported. Stanford researchers Nelson Liu, Tianyi Zhang and Percy Liang audited four generative search engines in 2023 (Evaluating Verifiability in Generative Search Engines). On average, only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to. Across the engines they audited, the responses rated most useful tended to have the least accurate citations. These were consumer search tools tested in 2023, so treat the figures as an illustration of the failure mode, not a measure of any current product.
- Retrieval reduces errors but does not remove them. Varun Magesh and colleagues at Stanford tested commercial AI legal research tools that retrieve from curated legal databases before answering (Hallucination-Free?, Journal of Empirical Legal Studies, 2025). Each tool hallucinated more than 17% of the time on their benchmark. Their definition counts two kinds of error: a response that is incorrect, and one that is misgrounded, where the cited source does not support the claim or does not apply. A response can be factually correct and still misgrounded. The authors conclude that users must keep verifying that key propositions are supported by their citations.
- Standards bodies name the same risk. NIST's Generative AI Profile (NIST AI 600-1, July 2024) uses the term confabulation for systems that confidently present erroneous or false content, and notes that outputs may include confabulated logic or citations that purport to justify the answer. It also warns about automation bias, meaning excessive deference to automated systems. Among its suggested actions is to review and verify sources and citations in generative AI outputs.
For a diligence team, the practical reading is simple: checking that links open is not enough. Someone has to read the cited passage and decide whether it says what the memo says.
Which claims need checking
Not every sentence in a memo carries the same weight. Check in order of what would change the committee's decision if it were wrong.
- Numbers. Revenue, growth, retention, burn, runway, customer counts, market size, cap table figures, round terms.
- Statements about third parties. Customer contracts, partnerships, regulatory approvals, litigation, prior investors.
- Statements about the team. Prior roles, exits, patents, degrees, time commitments.
- Comparisons. "Faster than", "first to", "larger than", rankings and market share.
- Inferences. Conclusions the draft reaches by combining facts, such as "the company is likely to need a bridge round."
Tier 1 to 4 claims need a source. Tier 5 claims need their inputs shown, and they should be labeled as analysis, not fact.
The five support grades
Grade each material claim against the exact passage it cites. The grades borrow the distinction between incorrect and misgrounded answers from the legal AI study above, adapted for diligence materials.
| Grade | Meaning | What the reviewer does |
|---|---|---|
| Supported | The cited passage states the claim, with the same figure, period and scope | Keep. Record page or cell reference |
| Partly supported | The source supports part of the claim, or a different period, unit or definition | Rewrite the claim to match the source exactly |
| Misgrounded | The source is real but does not say this, or does not apply (wrong entity, wrong date, superseded version) | Remove the citation. Find a real source or cut the claim |
| Unsupported | No source is given, or the source cannot be opened | Mark as an open question for the company or the deal team |
| Conflicting | Two sources in the data room disagree | Show both. Escalate to the deal lead before committee |
Two rules make the grades useful. First, a claim is only supported if the reviewer has read the passage, not the AI's summary of it. Second, the grade applies to the source the memo cites. If the reviewer finds support elsewhere, the citation is changed and the claim is regraded.
A diligence evidence register you can copy
Keep one row per material claim. This register is the record the committee can audit, and it outlives the memo.
| Field | What to record |
|---|---|
| Claim ID | A short reference used in the memo margin, for example FIN-04 |
| Claim as written | The exact sentence from the draft |
| Tier | Number, third party, team, comparison or inference |
| Source | Document name, version or date, and page, slide or cell |
| Source type | Company-provided, third-party, public record, or interview note |
| Grade | Supported, partly supported, misgrounded, unsupported or conflicting |
| Reviewer and date | Who read the passage, and when |
| Action | Kept, rewritten, cut, or sent to the company as a question |
Record the source type even when a claim is supported. A revenue figure that matches the company's own deck is supported by the deck; it is not independently verified. The committee should be able to see that difference at a glance.
How to run the check in the week before committee
- Freeze the source set. List the data-room documents and versions the draft was built from. Any claim citing something outside that list is regraded.
- Extract claims before reading the narrative. Pull every tier 1 to 4 claim into the register first. Reading the full memo first makes reviewers more likely to accept claims that fit the story.
- Check every number, then sample the rest. Grade all numerical claims. For other tiers, grade all of them on a first deal, then set a sample size your partners agree on, and widen it if the sample turns up misgrounded citations.
- Send open questions early. Unsupported and conflicting claims go to the company or the deal lead the same day, so answers arrive before committee rather than after.
- Mark the memo. In the final draft, flag any claim that is not graded supported. Partners should never have to guess which statements were checked.
- Sign off by name. The deal lead confirms that material claims were reviewed. The investment decision stays with the committee.
Common mistakes
- Trusting the summary of the source. Reading the AI's paraphrase of a document is not checking the document.
- Checking links instead of passages. A working link proves the source exists, not that it supports the claim. This is the misgrounded error the legal study describes.
- Letting fluency stand in for accuracy. Polished prose is not evidence. In the 2023 audit, the responses rated most useful tended to have the least accurate citations.
- Losing the register after the deal. The register shows which document types and claim types failed most often. Use it to tighten the checklist for the next deal, the same way a VC screening rubric improves when its logged reasons are reviewed.
Where software fits
The checks above work with any drafting tool, including none. What software can change is the cost of running them. A system that drafts from approved data-room documents, attaches a page-level reference to each material claim, and lists the claims it could not source makes the register a starting point instead of a blank sheet. Reviewers still read the passages and set the grades.
When a fund wants this built around its own committee template and evidence standards, we design fund-specific diligence and IC preparation systems that produce source-linked summaries, open-question and exception logs, and draft committee materials for review. Material claims are verified by the deal team, and investment approval stays with the fund's authorized decision-makers.