ENGINEERING THE NEXT GENERATION

Logo
Home/Blog/AI Vendor Security Questions: 25 Things to Ask Before You Share Data
AI SecurityVendor SelectionData PrivacyProcurementAI Governance

AI Vendor Security Questions: 25 Things to Ask Before You Share Data

November 18, 2026
Glowing shield with a keyhole in front of an AI core and streams of data
Most AI security risk sits in what the vendor does with your data, not in the model itself.

TL;DR & Quick Summary

When you give an AI vendor your data, most of the risk sits in what they do with it, not in the model. These 25 questions cover seven areas. For each, ask for the answer in writing, and check it against the contract.

  1. Data location and flow (Q1–4)
  2. Training and retention (Q5–8)
  3. Access and people (Q9–12)
  4. What the AI can do (Q13–16)
  5. Logging and monitoring (Q17–19)
  6. Incidents and assurance (Q20–22)
  7. Contract and exit (Q23–25)
  • Key Takeaway: A good answer is specific, names a document or a setting, and matches the contract. A weak answer is a reassurance with no detail.
  • This is not legal advice. Have legal and security review the final terms.
  • Get Started: Want help assessing a vendor? Schedule a strategy call with Cogniq AI or see our AI integration services.

How to Use This List

Don't send it as a form and accept "yes" to everything. Use it in three steps:

  1. Tier the data. Public content needs little scrutiny. Customer records, financial data, health data or trade secrets need all of it.
  2. Ask the questions in writing, and ask for the supporting document each time: a policy, a settings page, a certificate or a contract clause.
  3. Compare the answers with the contract. Marketing pages promise more than contracts commit to.

For a framework to organise this, the NIST AI Risk Management Framework and its generative AI profile are widely used reference points. They're guidance rather than a vendor questionnaire, so treat them as a structure, not a checklist.

1. Data Location and Flow

Q1. Where is our data processed and stored, by country or region? A good answer names regions and shows how they're enforced. Watch for "global infrastructure" with no commitment.

Q2. Which third parties touch our data, including model providers, cloud hosts and transcription or enrichment services? Ask for the full list. For AI products, the chain is often longer than the vendor first mentions.

Q3. How are we told about new sub-processors, and can we object? Under data protection rules such as GDPR Article 28, a processor shouldn't bring in another processor without the controller's authorisation, and should pass the same obligations on to it.

Q4. Is our data encrypted in transit and at rest, and who holds the keys? Ask whether you can bring your own keys for sensitive data.

2. Training and Retention

Q5. Is our data used to train or fine-tune any model, yours or a provider's? Ask about the vendor's own training and each model provider's policy, because they can differ by product tier. Get the answer into the contract.

Q6. Do model providers retain our prompts and outputs, and for how long? Retention periods, abuse-monitoring storage and human review are separate questions. Ask about each.

Q7. How long do you keep our data, and what is deleted when we leave? A good answer gives periods, a deletion process and confirmation of deletion.

Q8. Can we set our own retention and delete specific records on request? This matters for data subject requests under privacy law.

3. Access and People

Q9. Who at the vendor can see our data, and under what conditions? Look for role-based access, approval for exceptional access and logging of every instance.

Q10. Are our users separated from other customers' data, and how is that tested? For multi-tenant products, ask how isolation is enforced and verified.

Q11. How does the product respect our own permissions? If a user can't see a record in your CRM, the AI shouldn't reveal it to them. Ask how permissions are carried through to AI answers.

Q12. What background checks and security training apply to staff with access? Basic, but often skipped.

4. What the AI Can Do

Q13. What actions can the AI take in our systems, and which need a person to approve? List every action. The more it can do, the more controls matter. OWASP's guidance on excessive agency names excessive functionality, excessive permissions and excessive autonomy as the three root causes, which makes a useful lens.

Q14. What permissions does the AI hold, and can they be reduced to read-only? Principle of least privilege: only what it needs. Starting read-only is the cheapest security decision you can make.

Q15. How are instructions hidden in documents or messages handled? This is prompt injection. Ask how untrusted content is separated from instructions, and what stops injected text from triggering actions.

Q16. Can actions be reversed, and is there a limit on how much it can change? Caps on amounts, volumes and frequency limit the damage of any single error.

Our guide to human-in-the-loop approval gates covers how to design the approval steps these questions point to.

5. Logging and Monitoring

Q17. What do you log: inputs, outputs, data used and actions taken? You want to be able to reconstruct what happened and attribute it to a user.

Q18. Can we access the logs, and for how long are they kept? Logs you can't read don't help you in an investigation.

Q19. How do you monitor for quality failures, misuse and unusual activity? Look for alerting and a named person who reviews it, not just a dashboard.

6. Incidents and Assurance

Q20. How quickly will you tell us about an incident, and what will you tell us? Ask for a number of hours or days, and put it in the contract.

Q21. Which certifications or audits cover the product we'd use, and when were they last done? Request the report or certificate. Check the scope covers the actual product and environment, and the date is recent. A certificate for one product doesn't cover another.

Q22. Have you had a security incident involving customer data? What changed? Honest answers are specific. "We have never had any issues" deserves a follow-up.

7. Contract and Exit

Q23. What are our rights to audit, or to receive evidence of your controls? Where a full audit isn't practical, ask what evidence you'll receive and how often.

Q24. What are the liability and indemnity terms for a data breach? Have a lawyer review this. It's often the largest gap between what the vendor says and what it commits to.

Q25. How do we leave, and what do we take with us? Ask for data export in a usable format, code and configuration ownership where relevant, and a transition period. Our guide to writing an AI development statement of work covers ownership and handover terms.

What Good and Weak Answers Sound Like

Topic Good answer Weak answer
Training "No customer data is used for training. See clause 9.2 and our provider terms." "We take privacy very seriously."
Sub-processors A list, a notification period and an objection right "We use industry-leading partners."
Access Role-based, logged, approved for exceptions "Only authorised staff."
Agent actions A list of actions with limits and approval rules "The AI is fully autonomous and safe."
Incidents "Within 72 hours, as in schedule 3" "We'll let you know if needed."
Exit Export format, deletion process, transition period Silence

Matching Effort to Risk

You don't need all 25 for every tool. A rough guide:

Data and capability Effort
Public or low-sensitivity content, no actions Short review: Q1, Q5, Q7, Q25
Internal business data, read-only Sections 1–3, 5 and 7
Customer personal data or regulated data All sections, plus legal review
AI that takes actions in your systems All sections, with extra time on section 4

Common Mistakes

  • Accepting a certificate as the whole answer. Certificates show a security programme exists. They don't answer AI-specific questions such as training, agent permissions and prompt injection.
  • Relying on marketing pages. Only the contract and the signed data processing terms bind the vendor.
  • Forgetting the model provider. Your vendor's promises are only as good as the terms it has with its own providers.
  • Reviewing once. Vendors change sub-processors, models and features. Ask for notification, and re-check yearly.
  • Skipping the exit questions. The time to ask how you leave is before you've entered.

Where This Fits in the Buying Process

Security questions belong early in vendor selection, not after you've chosen. Add the relevant ones to your RFP, and score them. Our RFP checklist for AI development and guide to shortlisting AI agencies show where. For the cost of putting controls in place around an integration, see what AI governance costs with legacy systems.

Conclusion

Assessing an AI vendor comes down to seven areas: where data goes, what's retained and trained on, who can access it, what the AI can do, what's logged, how incidents are handled, and how you leave. Ask in writing, ask for the document behind every answer, and check it against the contract. Specific answers show a vendor has thought about the risk. Reassurances don't.

Schedule a strategy call with Cogniq AI to review a vendor or scope the controls for your project, or explore our AI integration services.

Frequently Asked Questions

Ask where your data is processed and stored, whether it is used to train any model, who at the vendor and its sub-processors can access it, how long it is kept and how it is deleted, what the AI is allowed to do in your systems and who approves sensitive actions, how activity is logged, how incidents are reported, and how you can leave and take your data with you. Get the answers in writing and compare them with the contract.

It depends on the vendor, the product tier and the model provider behind it, so ask directly and get the answer in the contract, not only in marketing material. Ask separately about the vendor's own training, its model providers' retention and training policies, and any human review of conversations.

A sub-processor is another company the vendor uses to process your data, for example a model provider, a cloud host or a transcription service. Under data protection rules such as the GDPR, a processor should not engage another processor without the controller's authorisation, and should flow the same obligations down to it. For AI products the chain is often long, so ask for the full list and how changes are notified.

List every action the agent can take in your systems and what limits apply to each. Ask what it can do without a person approving, how approvals work, how every action is logged and attributed, and how actions can be reversed. OWASP names excessive functionality, permissions and autonomy as the three causes of excessive agency in LLM applications, which gives you a useful checklist.

Certifications are a useful signal that a vendor has a managed security programme, but they do not replace asking AI-specific questions. Ask which certifications cover the actual product and environment you will use, request the report or certificate, and check its scope and date. Then ask the AI-specific questions a general audit may not cover.

Responsibility is shared and depends on the contract and the law. Your organisation usually stays accountable to its customers and regulators for the data it collected, even when a vendor processes it. That is why the contract should cover security standards, incident notification within a defined time, cooperation, liability and the right to audit or receive evidence. Have a lawyer review those terms.