LP AI DILIGENCE QUESTION BANK
Version 1.0 · Informed by AIMM v1.4
AgenticInvestor

Five questions for manager meetings and operational due diligence.
Use these prompts to distinguish repeatable, governed AI capability from polished claims or isolated experimentation.

1. REPEATABILITY
Question: Show me one AI workflow your whole team runs to a single standard.
Evidence to expect: A written workflow, defined human sign-off points and evidence that more than its author can run it.
Red flags: A tool list, a workflow held in one enthusiast's head or reliance on a single advanced user.

2. OWNERSHIP AND BOUNDARIES
Question: Who owns your screening workflow, and what may AI never do without human sign-off?
Evidence to expect: A named owner, written boundaries and consistent answers from senior and junior team members.
Red flags: Ownership belongs to “everyone”, boundaries rely on common sense or different people describe different controls.

3. AUDITABILITY
Question: For a recent deal, show me what AI produced, its sources and who approved it.
Evidence to expect: A live deal record with run history, sources attached to claims and a visible approval trail.
Red flags: A policy document instead of a demonstration, or no answer about which models can see the firm's data.

4. IMPROVEMENT LOOP
Question: What changed last quarter because an AI workflow underperformed—and how did you know?
Evidence to expect: A specific example with a measure, the change made and the result that followed.
Red flags: Improvement described as philosophy, with no numbers, events or examples of anything being rolled back.

5. KNOWLEDGE SOVEREIGNTY
Question: Whose AI accumulates your manager's judgement—and could they take it with them?
Evidence to expect: A contractual answer covering data rights, exportable memory and portable workflows.
Red flags: Unclear ownership of accumulated knowledge, data or workflows—or no considered answer at all.

HOW TO USE THE BANK
- Add the questions to the operational section of the DDQ.
- Ask for demonstrations and records where appropriate.
- Treat confident language as a prompt for evidence, not proof.
- Read trajectory, not just current position.
- Value a candid improvement plan over unsupported perfection.

This is not a pass-or-fail scorecard. The strongest signal is whether a manager can support its claims with evidence, acknowledge material gaps and explain a credible improvement path.

Free to use with attribution to AgenticInvestor.
https://agenticinvestor.org/LPquestions/