∞Direct answer
Direct answer
Use AI to sort work. Do not let it clear a deal. Put each key claim in a queue. Add its source, date, gap, risk, and owner. A reviewer must be able to change the result. NIST says records can help with human checks and accountability. [1] The ICO says a real reviewer needs the power and skill to change a choice. [2]
Set the queue boundary first
Due diligence means checks before you invest. AI can find claims and sort files. It must not turn a weak claim into a checked fact.
Set a rule for the queue. It may sort claims, spot repeats, and flag proof that is not there. A person must say if a claim is checked, open, or wrong.
Add a human gate before a deal memo, a founder ask, or an investment choice. This stops a neat AI note from being the last word.
NIST asks for clear roles in human-AI oversight. It also asks for set risk steps and reviews. [1]
- AI task: sort, pull out, compare, or flag.
- Human task: check proof and pick the next step.
- Gate: no key claim moves on without a named reviewer.
- Scope: say which deal stage and files the rule covers.
Give each item a small proof card
Make one queue item for one claim. Do not put the whole company in one long card. Small items are easy to check and give to one person.
Write the claim in plain words. Add the first source, its date, and the check date. Keep the AI note next to the source. Do not use it in place of the source.
Say what you do not know. A blank may mean the firm is new, private, or not checked. It does not prove a poor deal.
NIST says a clear record can show data, planned use, and who made a choice. It says records can help a person check the work. [1]
- Claim: the exact point that may change the review.
- Proof: link or file, place in it, and source date.
- State: checked, open, in doubt, or wrong.
- AI note: a pull-out, match, or question.
- Owner and due date: one person and the next check.
Send doubt to the right person
Use a few clear routes. Send an item to source check when key proof is not there. Send it to a conflict check when two sources clash. Send it to a deal owner when it may change fit or a deal term.
Do not use one score to pick a route. A high score may hide old proof, a clash, or a lost file. Show why the item went to that route.
Give the reviewer three real choices: confirm, change, or hold. A hold must name the proof that is missing and the next step. A pass must give a short reason.
The ICO says a person must not just rubber-stamp an automated result. A reviewer should weigh the result, look at the data, and have real say in the choice. [2]
- Source check: prove or fix one claim.
- Conflict check: compare sources and save the reason for the choice.
- Deal check: look at a point that may change fit or the deal path.
- Raise it: name who can accept a gap that stays open.
Let the reviewer act
A reviewer needs the source, time to read it, and the right to change the item. A name in a log is not enough.
Keep the review trail. Save who changed the state, when they did it, and why. Keep a link to the proof that led to the change.
If the AI note looks wrong, let the reviewer remove it, fix it, or send it back for more work. Do not make the reviewer defend the tool first.
Article 14 of the EU AI Act applies to high-risk systems in its scope. It calls for human oversight that can help a person step in and, when needed, override or stop a system. [3] This does not say a due diligence tool is high-risk.
- Reviewer: name, role, and right to act.
- Action: confirm, change, hold, reject, or raise.
- Reason: one short note tied to proof.
- Trail: date, old state, new state, and source link.
Test the queue after real work
Each month or quarter, check a small set of items. Compare the AI note with the source and the human choice. Look for items cleared by mistake, held by mistake, or sent to the wrong route.
Count a few things. Count source gaps, clashes, human changes, and late items. These counts show process faults. They do not show future returns or prove a good deal.
Change the rule when the same fault comes back. Save the change and its date. Tell reviewers what changed before they use the new rule.
NIST says use of generative AI may need more human review, tracking, records, and management oversight. [4] Use checks that fit the task and risk.
- Sample: set the date, size, and reviewer.
- Fault: wrong claim, old source, no source, or wrong route.
- Change: new rule, owner, and start date.
- Follow-up: check if the change fixed the repeat fault.
FAQ
Frequently asked questions
What goes in an AI due diligence review queue?
Put one key claim in each item. Add the source, dates, AI note, gap, risk, owner, and next step. Keep the human choice and reason with the item. This lets a person check it later.
Can AI clear a due diligence point?
No. Treat an AI answer as a lead, not clearance. A reviewer should read the key source and pick the state. NIST says records can help human review and accountability. [1]
What makes a human review real?
The reviewer needs the right data, the skill to judge it, and the power to change the result. They must not apply an AI result by habit. The ICO gives this guide for AI decision support. [2]
How often should an investor test the queue?
Pick a set cycle that fits deal flow and risk. Check a small sample after real work. Track repeat faults. Change the rule when you need to. NIST includes checks and set reviews in AI risk work. [1]
REF
Sources
- 1
National Institute of Standards and Technology
Artificial Intelligence Risk Management Framework (AI RMF 1.0) ↗ - 2
Information Commissioner's Office
How do we ensure individual rights in our AI systems? ↗ - 3
- 4
National Institute of Standards and Technology
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile ↗