Use the questionnaire to separate AI claims from security readiness
A healthcare AI vendor security questionnaire should test more than model quality claims. It should determine whether the vendor can protect ePHI, operate as a HIPAA business associate when applicable, enforce runtime controls over AI agents and tools, preserve source-system access controls, produce usable audit evidence, and support ongoing governance after deployment.
The strongest questionnaires ask for evidence, not assurances. Buyers should request data-flow diagrams, BAA terms, access-control details, logging samples, incident response obligations, change management procedures, secure development evidence, and AI-specific red-team or runtime testing results.
What the questionnaire should validate
| Area | What to validate |
|---|---|
| PHI handling | Where PHI is processed, stored, logged, retained, backed up, deleted, and accessed. |
| Agent permissions | How AI agents are identified, authorized, constrained, approved, monitored, and revoked. |
| Runtime controls | How prompt injection, unsafe tool use, data leakage, and excessive agency are detected or blocked. |
| Audit evidence | Whether security teams can review prompts, retrieved sources, model versions, tool calls, policy decisions, and outputs. |
1. PHI, BAA, retention, and permitted use questions
Healthcare AI due diligence should begin with the vendor’s handling of ePHI and the contractual, operational, and technical evidence that supports those answers. The questionnaire should clarify whether the vendor can operate as a HIPAA business associate when applicable and how it limits permitted use of customer data.
- Ask where PHI is processed, stored, logged, retained, backed up, deleted, and accessed.
- Request BAA terms when the vendor’s role makes them applicable.
- Ask for data-flow diagrams that show how information moves through the AI system, connected tools, and source systems.
- Request retention and deletion details for prompts, retrieved context, model outputs, logs, backups, and support access records.
- Confirm how source-system access controls are preserved when the AI system retrieves or acts on healthcare data.
2. Runtime AI security, agent identity, and tool governance
Healthcare AI security questions should explicitly cover runtime behavior. Static documentation is not enough if the system can retrieve clinical documents, invoke APIs, send messages, write to systems of record, or initiate automated actions. The buyer should understand the boundary between user instruction, agent autonomy, policy enforcement, and human approval.
Agent identity and permission model
Each AI agent should have a clear identity and a defined permission model. Shared service accounts, broad API tokens, and implicit access inherited from an administrator create review and containment problems. The questionnaire should ask how agent identity is represented, how permissions are scoped, how access is reviewed, and how emergency revocation works when a workflow, connector, or agent behavior becomes unsafe.
Runtime review focus
Evaluate whether the vendor can explain how an agent is identified, what it is permitted to do, which tools it can invoke, when approval is required, and how unsafe behavior is contained during live operation.
3. Auditability, monitoring, and evidence your team should require
Auditability is central to healthcare AI review because security teams need to understand what happened, which system or agent acted, what data was involved, and which controls were applied. The questionnaire should ask vendors to provide usable evidence, including logging samples and explanations of how those records support investigations.
- User and agent identity: Logs should show the initiating user, acting agent, delegated identity, session context, and relevant authorization decisions.
- Prompt and output traceability: The vendor should explain what portions of prompts, retrieved context, model responses, and final outputs are retained or redacted.
- Tool-call evidence: Audit records should capture tool name, parameters where appropriate, target system, approval state, execution result, and error conditions.
- Model and policy versioning: Evidence should identify the model or model version, prompt template, policy version, connector version, and material configuration changes.
- Administrative activity: Admin changes to permissions, tools, policies, retention settings, connectors, and support access should be logged and reviewable.
- Export and preservation: The vendor should describe log export options, preservation during investigations, and customer-specific impact analysis support.
4. Governance, secure development, and operational obligations
Healthcare AI vendor review should continue beyond the purchase decision. The questionnaire should test whether the vendor can support ongoing governance after deployment, including incident response obligations, change management procedures, secure development evidence, and AI-specific red-team or runtime testing results.
Operational obligations should be specific enough for security, privacy, clinical, and compliance stakeholders to review. General assurances are less useful than documented processes, sample evidence, and clear responsibility boundaries.
Where Trussed AI fits in the evaluation
Use this questionnaire to evaluate AI vendors at runtime, not only during procurement. The review should focus on PHI exposure, agent permissions, tool governance, policy enforcement, auditability, and operational readiness before approving healthcare AI deployment.