Claude for Due Diligence: Evidence and Open Questions

Use Claude for due diligence by giving it an approved set of sources, a specific checklist, and instructions to separate supported findings from unanswered questions. Have qualified people verify the evidence and approve conclusions before anyone relies on them. The useful output is a reviewable record, not a confident verdict.

Tools like fenbs keep evidence links, follow-up work, and approved decisions outside a private Claude conversation. That helps when a colleague must find the document behind a finding or understand why an issue remains unresolved.

How should you set up the first diligence pass?

1. Define the decision. For example, decide whether a prospective vendor can meet your requirements, rather than asking whether the entire company is trustworthy.
2. Agree on the questions and review boundaries. Name the entities, reporting periods, business activities, and subjects the review covers.
3. Approve the information Claude may receive. Follow your organization's confidentiality requirements and review the account settings and contractual terms before uploading sensitive records.
4. Give Claude a small, labeled source set and request findings with exact document locations. Test its output against those sources before expanding the review.
5. Send unresolved issues to named reviewers. Keep recommendations separate from the person's final decision and any conditions attached to it.

If you are new to the assistant, start with what Claude AI is. The broader AI for due diligence guide covers the wider approach across tools; this workflow concentrates on getting a Claude finding from a document to a defensible review.

What evidence should Claude receive?

Start with a source inventory, not a folder described as everything. Give each document an identifier, title, date, version, location, and access restriction. Mark drafts, unsigned agreements, and management statements so Claude does not silently treat them as final or independently verified.

For a US public company, the SEC's guide to reading a 10-K↗ (opens in a new tab) helps locate business descriptions, risk factors, and financial statements. Match every extract to the correct reporting period. A filing is evidence to examine, not proof that every question about the proposed relationship has been answered.

Anthropic's file upload guide↗ (opens in a new tab) explains supported formats and document-processing limits. Inspect charts, scans, footnotes, and tables yourself when they matter; a successful upload does not establish complete extraction. If you are considering a longer file-based task, our Claude Cowork overview explains that part of the product.

How do you get findings you can verify?

Ask Claude to extract before it evaluates. Every finding should identify the source, exact location, observed fact, proposed interpretation, and remaining uncertainty. Require a page number plus a section heading or table label, since PDF viewer numbering can differ from printed numbering.

Review only the supplied documents against this checklist.
For each question return:
- Finding ID and checklist question
- Source ID, document version, and exact page or section
- Relevant passage or table entry
- What the source directly establishes
- Your interpretation, clearly labeled
- Missing evidence and contradictory information
- Question for the qualified reviewer
Say "not established by the supplied sources" when support is missing.
Do not recommend approval or infer that an absent record does not exist.

Open each cited location for findings that affect the decision. Confirm that the passage exists, supports the wording, and applies to the entity under review. Recalculate important totals from the underlying figures, checking units, currencies, dates, and whether amounts are before or after deductions.

When public research is needed, Anthropic's web search guide↗ (opens in a new tab) explains how Claude supplies source citations. Open the original pages yourself. A citation that discusses a similar company or an old policy does not support the current claim.

Keep verified source locations beside the follow-up work using fenbs Links and listings, so a later reviewer can retrace the finding.

How should you challenge Claude's conclusions?

Run a second pass that tries to disprove each important interpretation. Ask Claude to identify conflicting passages, assumptions that connect the evidence to its conclusion, and documents that could change the answer. Treat that response as another draft requiring verification.

• Ask whether a claim is a contractual promise, a description of past performance, or a statement someone supplied without corroboration.
• Check whether amendments, exclusions, definitions, or exceptions change the apparent meaning of the cited passage.
• Ask what evidence would contradict the finding and whether the review actually included that evidence.

Say a vendor presentation promises continuous support, but the agreement limits service hours. Claude should surface the discrepancy and ask which terms apply. It should not combine the two into a reassuring sentence that hides the unresolved difference.

Which questions must stay open?

A missing record, unreadable attachment, contradictory date, or unanswered management question stays open. Record exactly what was reviewed and what was missing. "No pending dispute found in the supplied documents" is narrower than "there are no pending disputes" and should remain narrower in the final report.

Give every open question an identifier, the finding it affects, the requested evidence, the person following up, and the next action. Distinguish a question answered by evidence from a risk a person knowingly accepts. If you later build a repeatable review process in code, the Claude API guide covers that separate setup.

When a replacement document arrives, record its version and which findings need rechecking. Preserve the earlier conclusion as the answer reached on the earlier evidence, rather than quietly rewriting its basis.

Who reviews and approves the result?

Choose reviewers by the issue: counsel for legal interpretation, an appropriately qualified financial professional for accounting questions, and a security specialist for technical controls. Give each reviewer the original evidence and the open questions, not just Claude's polished summary.

For US lawyers, ABA Formal Opinion 512↗ (opens in a new tab) addresses competence, confidentiality, and independent verification or review of AI output. The necessary review depends on the task and applicable duties. An assistant's confidence does not replace that professional judgment.

Record who approved which conclusion, when, on what evidence, and subject to which conditions. In fenbs Decisions and rules, people decide and sign off; assistants can record what a person decided. Keep a proposed conclusion open until that person acts.

How to do this with fenbs

1. Create a free account, then add a project for the review. Write its scope and the decision it will support.
2. Add approved evidence references in Links and listings. Keep restricted documents in their approved repository and save only information your organization permits in fenbs.
3. Create tasks for unresolved questions, with source references, the requested follow-up, and the person responsible written in the task. Record approved conclusions in Decisions and rules.
4. Connect Claude through Integrations. Ask it to read the project context before working and save progress with checked findings, outstanding evidence, and the next review step before stopping.

fenbs is free for one person, with no credit card. It does not replace a secure deal room, specialist diligence software, or qualified review. If you only need a one-time document summary, a carefully reviewed Claude conversation may be enough.

At the next session, begin with changed documents and unanswered questions. Check whether new evidence invalidates an earlier conclusion before drafting the final report. More practical workflows are in our AI at work guides.