# Third-party AI due diligence questionnaire

> Send before procurement. The answers become contract schedules and inventory fields.
> Anything marked **⚑** is a common deal-breaker — treat a vague answer as a red flag.

**Vendor:** · **Product:** · **Assessor:** · **Date:**

---

## 1. The system

1. What does the system do, and what decision or output does it produce?
2. What model type and architecture underpins it? Is any component third-party or open-weights?
3. Which of your sub-processors touch our data? List name, location and function.
4. Is the model hosted in a fixed region? Can we pin processing to a region? ⚑

## 2. Data handling

5. What data do you need from us, and what is the minimum viable set?
6. **Do you use our data — including prompts, inputs, outputs or telemetry — to train, fine-tune or evaluate your models?** If yes, under what basis, and can we opt out contractually? ⚑
7. How long is our data retained, including in logs, caches and backups? What is the deletion SLA?
8. Is our data logically or physically segregated from other customers'?
9. Do you process special-category data? Under what safeguards?
10. What happens to our data on termination? Provide the export format and deletion certificate process.

## 3. Legal roles and compliance

11. Do you act as controller or processor for each processing activity? Provide a breakdown.
12. Provide your Art. 28 processing terms and current SCCs / transfer mechanism.
13. Under the EU AI Act, are you a provider? What risk tier does the system fall into, and on what analysis?
14. If high-risk: provide conformity assessment status, technical documentation (Annex IV) summary, and CE marking evidence. ⚑
15. Will you supply the information a deployer needs to meet its own obligations, including for a fundamental rights impact assessment?

## 4. Performance, fairness and limitations

16. What are the documented accuracy/performance figures, and on what benchmark and population?
17. On which populations does performance degrade? Provide subgroup results. ⚑
18. What bias testing has been performed, using what metric, and when was it last run?
19. What are the known failure modes and documented limitations?
20. Provide a model card or equivalent documentation.
21. How is the system evaluated for drift, and how often?

## 5. Transparency and explainability

22. What explanation can be produced for an individual decision, and in what form?
23. Can we surface that explanation to an affected person?
24. Is AI-generated output marked or watermarked?

## 6. Human oversight

25. What controls let a human review, override or reverse an output?
26. What guidance do you supply on avoiding automation bias in reviewers?

## 7. Security

27. Which certifications do you hold (ISO 27001, ISO 42001, SOC 2)? Attach current reports. ⚑
28. How do you defend against prompt injection, data poisoning, model extraction and inversion?
29. Have you been red-teamed? By whom, and when? Can we see the summary?
30. What is your breach notification commitment, in hours? ⚑
31. Do you support SSO, role-based access and audit logging?

## 8. Change management and continuity

32. How are we notified of model updates that materially change behaviour, and how far in advance? ⚑
33. Can we pin a model version, and for how long?
34. What is your deprecation policy and notice period?
35. What are your uptime commitments and remedies?
36. What happens if you cease trading or discontinue the product?

## 9. Incidents and liability

37. Describe your AI incident response process and our escalation path.
38. What indemnities apply to IP infringement arising from model output? ⚑
39. What liability cap applies, and does it carve out data protection breaches?
40. Have you had a regulatory enforcement action or reportable breach in the last 3 years?

---

## Assessment outcome

| Area | Rating (R/A/G) | Notes |
|---|---|---|
| Data handling | | |
| Legal / regulatory | | |
| Fairness and performance | | |
| Security | | |
| Change management | | |
| Commercial / liability | | |

**Decision:** ☐ Approve ☐ Approve with conditions ☐ Reject

**Conditions to land in the contract:**

**Reassessment due:**