How Do I Use ChatGPT or Claude for Due Diligence?

Updated October 2026 · 6 min read · Anweshna

What a general assistant is genuinely good at

A general-purpose model is a fast, fluent reader. On a single document of modest length it does a handful of jobs well:

• Orientation. "Summarise what this agreement does, who the parties are and what each one owes the other." Useful in the first ten minutes with an unfamiliar document.
• Locating. "Where does this contract deal with termination, assignment and change of control? Quote the clauses." Finding the section is the slow part of reading; checking it is fast.
• Explaining. "What does this indemnity basket mean in plain English?" Helpful for a generalist reading outside their specialism.
• Drafting questions. "Draft follow-up questions to the seller about the customer concentration described here." The model writes; your team decides what to send.

All four have something in common: a person checks the output against the document in the same sitting. That is the condition under which a general assistant is safe to lean on.

Prompts that hold up better

Most bad output from a general assistant starts with an open question like "any risks here?" These patterns tend to produce answers you can check:

1. Ask for quotes, not opinions. "List every clause that lets a counterparty terminate or renegotiate on a change of control. For each, quote the sentence verbatim and give the section number." A quote can be checked with Ctrl+F. A characterisation can't.
2. Give it a fixed list. "Check this document against these categories: financial, legal, tax, HR, IP, data privacy. For each, say 'found' with a quote, or 'not addressed'." A fixed list makes two runs comparable and makes silence visible.
3. Allow "not in the document" as an answer. "If the document does not say, answer 'not stated'. Do not infer." Without that permission, a fluent model fills gaps.
4. Separate reading from judging. First ask what the document says, then, in a separate prompt, ask what it might mean for the deal. Mixing the two is where boilerplate turns into alarm.
5. Ask about real events and hypotheticals separately. "Distinguish risks the company says could happen from events it says have happened." Risk-factor sections are written to sound alarming, and most of what they describe has not occurred.

Where it breaks: four failure points

1. Length

Annual reports, credit agreements and disclosure schedules routinely run to hundreds of pages. Even when a document technically fits in the context window, models do not attend evenly across it. Research on long-context behaviour found that performance is often highest when the relevant information sits at the beginning or end of the input, and degrades when it sits in the middle.^[1] A clause on page 140 of 300 is in exactly the position that study describes.

The practical consequence: "I asked about the whole document and it found nothing" is weaker evidence than it feels. Split long documents and ask about each part.

2. Quotes nobody checks

A general assistant can produce a quote that reads plausibly and is not in the document. Purpose-built legal research tools have been measured hallucinating between 17% and 33% of the time on legal research questions, despite vendor claims to the contrary.^[2] That study measured case-law research, not review of a document you supplied, so it doesn't measure your use case directly. It does establish that fluent, confident and wrong is a real failure mode for this category of tool.

3. Repeatability

Run the same open prompt twice and you can get a different list, or a different severity on the same issue. For a one-off question that doesn't matter. For anything that feeds an investment committee, it does, because nobody can reconstruct why the first answer said "material" and the second said "routine". A fixed checklist in the prompt helps. It doesn't make the output auditable.

4. Confidentiality depends on your plan

Deal documents usually arrive under an NDA. Whether pasting them into an assistant is compatible with that NDA depends on the assistant's terms for your specific plan: whether inputs are used for training, how long they are kept, and who at the provider can see them. Those terms differ between consumer and business plans, and they change. Read the current terms for the plan you are actually on before uploading anything a counterparty gave you in confidence, and ask your counsel whether your NDA permits it.

What a screening tool adds, and what it doesn't

A purpose-built screening tool is the same kind of model wrapped in fixed rules. Anweshna, for example, scores each document against a fixed ruleset (15 categories for M&A), refuses to quietly truncate a filing that is too long for one pass (Deep Scan, on Pro, Enterprise and per-deal plans, splits it into chunks and scores every one), and checks every finding against the source text before it reaches a reviewer. It also treats silence in a deal-blocking category as a gap rather than a clean result. We describe those checks in detail in What Makes Anweshna Different from a ChatGPT Wrapper?

What it doesn't add is judgment about your deal. Screening tells you where to look first. It doesn't tell you whether a price adjustment is warranted, and neither does a general assistant.

A workable split

• Use a general assistant for orientation, for locating clauses, for explanations and for drafting questions, always with quotes and always checked.
• Use a fixed, repeatable screen when many documents need the same first pass, when length makes a single prompt unreliable, or when the output will be relied on by people who didn't run it.
• Use people for anything that becomes a conclusion: price, structure, walk-away.

Bottom line

ChatGPT and Claude are excellent reading aids and unreliable sources. Ask them for quotes, give them a fixed list, let them say "not stated", and check every quote before it travels. Once the documents get long, numerous or committee-bound, the work stops being prompting and becomes process. That is the point to use a tool built for repeatable screening.

Compare it against a screening run on the same document →