Blog
Third-Party Risk
Assess AI vendors with 15 questions on data use, providers, permissions, evaluations, and incidents. Know which evidence to request and what to escalate.
A support platform may be approved to store tickets without being approved to send them to a model or let an agent update customer records. When a new AI feature appears, the vendor name may stay the same while the data flow and authority change.
The original assessment might already cover that use. It might not. Start by identifying the feature, account, data, users and permitted actions, then check them against the agreement and configuration your team approved.
AI vendor risk assessment adds that use-specific review to third-party risk management (TPRM). The decision should state what the organization permits and which evidence supports it, rather than assign a single approval to everything a supplier might offer.
The 15 questions below cover data handling, downstream providers, security, performance and governance. Each pairs a question with evidence to request and a reason to escalate. Use them in proportion to the intended use and its consequences.
What does an AI review add to existing vendor assurance?
A SOC 2 report can provide useful assurance about the systems and controls within its scope. Review that scope, the period covered, exceptions and relevant subservice organizations. It does not by itself answer whether a new AI feature performs reliably on your task or enforces your chosen action limits.
The additional questions are practical: which parties receive prompts and files, what is retained, which inputs can influence the system, and what happens when an output is wrong? For an agent, ask what it can change and how the organization can stop it.
Check the contract before assuming a gap. Existing confidentiality, security and service terms may apply, while training rights, feature-specific retention or incident obligations may need clarification. A questionnaire answer is evidence of a claim; it is not automatically a contractual commitment.
NIST’s voluntary Generative AI Profile includes embedded AI in supplier due diligence and addresses third-party contingencies. Its value here is the lifecycle view: approval, changes, incidents and exit all belong in the review.
Which vendor changes should trigger an AI review?
Use the following signals to start triage. First establish whether the change is enabled in your account and whether it changes the approved use:
Release notes that announce AI assistants, copilots, or agents
Updated privacy terms that mention model training
Subprocessor notices that name AI model providers
New AI add-on pricing tiers
A release note need not trigger a full reassessment of the supplier. A new data recipient, broader action permission or materially different use may justify one. Record why the change is covered by the existing approval or which part needs review.
What questions should you ask an AI vendor?
Define the proposed use before sending the questionnaire. A tool limited to public drafting needs different evidence from an agent with customer records and write access. Ask the vendor to identify the version and configuration its answers cover.
Data use and model training
Separate model training from storage and deletion. A no-training commitment does not establish that prompts, files, outputs or operational logs are never retained.
Question | Evidence to request | Escalate when |
|---|---|---|
1. Is our content used to train or improve models? | Terms and settings for this plan, feature and downstream providers. | The permitted use is unclear or conflicts with your requirements. |
2. What is stored, by whom and for how long? | Retention schedule for prompts, files, outputs, logs and feature state. | The vendor cannot identify stored copies or relevant exceptions. |
3. How do export and deletion work? | Administrator controls, deletion timing and termination procedure. | The organization cannot retrieve or remove data as required. |
Model supply chain
Identify which organizations actually receive data. A vendor using a hosted model API has a different dependency from one running downloaded weights in its own environment. Model authorship alone does not establish a downstream processing relationship.
Question | Evidence to request | Escalate when |
|---|---|---|
4. Which models, services and parties receive our data? | A feature-level data-flow map and relevant supplier list. | The vendor cannot explain recipients or their roles. |
5. Where are processing and storage performed? | Locations covering inference, files, logs and support access. | Locations are unknown or conflict with applicable commitments. |
6. Which changes will you notify us about? | Process for model, provider, permission and retention changes. | Material changes can bypass review without notice or control. |
AI-specific security
Security testing should cover the complete deployment. A well-behaved model cannot compensate for a credential that permits the wrong action or exposes another customer’s records.
Question | Evidence to request | Escalate when |
|---|---|---|
7. How do you test and contain prompt injection? | Tests using hostile content and traces of resulting tool calls. | The answer relies only on prompts or a generic firewall claim. |
8. What actions can the AI take, under which identity? | Actual scopes, approval gates, denied-action tests and revocation. | Sensitive actions have no enforceable permission boundary. |
9. How do you test data and tenant separation? | Tests denying access outside the user or organization scope. | Isolation is asserted without deployment-specific evidence. |
Accuracy and human oversight
A useful evaluation describes the task, test material and failure cases. An accuracy percentage without that context is difficult to use in an approval decision.
Question | Evidence to request | Escalate when |
|---|---|---|
10. How is performance evaluated on our intended task? | Representative test cases, scoring method, version and known failures. | A benchmark score substitutes for evidence about the workflow. |
11. What happens when evidence is missing or conflicting? | Examples of uncertainty, escalation and unsupported-answer handling. | The system presents an answer without exposing the evidence gap. |
12. What can people approve, override or reverse? | Demonstration of review gates, corrections and recovery. | High-impact actions cannot be stopped or meaningfully reviewed. |
Governance, incidents and service continuity
Look for operating evidence as well as policies. Framework alignment can organize the review, but it cannot replace a demonstration of the feature’s controls or an agreed incident response.
Question | Evidence to request | Escalate when |
|---|---|---|
13. What assurance covers this AI feature? | Relevant assessment scope, policies, findings and corrective actions. | A certificate or alignment claim has no clear connection to the service. |
14. How are AI incidents reported and contained? | Contacts, notification commitments, preserved logs and access revocation. | No owner or workable process exists for consequential failures. |
15. What happens if a model or provider becomes unavailable? | Fallback, export, termination and recovery plans for this workflow. | The business cannot continue or exit within its required tolerance. |
How do you build AI vendor risk assessment into your TPRM program?
The questionnaire is useful only if its answers change the decision. Carry the evidence through these four steps:
Tier the use by consequence. Consider data sensitivity, decisions affected, action permissions and dependence on the service. Use the relevant questions for a limited public-data tool and deeper testing for sensitive or consequential workflows.
Resolve important claims with evidence. Check the agreement, configuration and test result for the same feature and version. Record an unknown as unknown; do not turn an unanswered question into a pass.
Record a usable decision. Approve the defined use, impose conditions, request remediation or decline. Name the owner, permitted data and actions, exceptions and person accepting residual risk.
Review changes and incidents. Reassess the affected part when a new recipient, permission, model behavior or retention term changes the decision. Keep a review date and a working fallback.
Read the evidence at the level of the feature
For example, OpenAI’s API documentation distinguishes its default no-training policy from abuse-monitoring logs and application state. Retention controls vary by feature and account. This is why a provider-wide statement cannot settle every retention question for a particular integration.
Anthropic’s agent-evaluation guidance makes another useful distinction: the transcript is not the final state of the system. For an agent that updates a record, inspect the resulting record as well as its answer. Ask for failures and denied actions, not just successful demonstrations.
Approve a defined use, with evidence you can revisit
Keep the supplier record linked to the deployment decision. If the feature can take actions, use the AI agent governance guide to define operating limits. If employees are using accounts outside the approved process, review the shadow AI account and data flow before treating the supplier as approved for that use.
Supplier assessment is one part of the program. The enterprise AI governance guide connects purchased AI with internal development, employee use and the decisions needed across their lifecycles.
Certification is one input: ISO/IEC 42001 addresses an AI management system and does not replace applicable law or guarantee a feature’s performance. The practical end point is a decision another reviewer can reconstruct: what was approved, on what evidence, within which limits, and when it must be reconsidered.
For the software supporting these reviews, the AI governance tools buyer’s guide includes a supplier-change scenario and an evaluation sheet for your shortlist.





