AI procurement
A practical AI vendor security questionnaire for enterprise procurement: data use, retention, training, SOC 2, ISO, residency, access control, RAG permissions, evals, red teaming, and incident response.
Comparison
| Option | Best for | Strengths | Tradeoffs | Use when |
|---|---|---|---|---|
| Standard SaaS questionnaire | Baseline security, identity, encryption, vulnerability management, and SOC 2 review | Covers familiar enterprise security controls. | Often misses model training, prompt logs, embeddings, tool actions, and AI-specific failures. | The vendor is a normal SaaS tool with limited AI data access. |
| AI-specific questionnaire | AI vendors handling prompts, files, internal knowledge, connectors, agents, or customer-facing output | Targets data use, retention, model behavior, evals, red teaming, and output risk. | Requires cross-functional input from security, privacy, legal, engineering, and business owners. | The vendor processes sensitive prompts, files, or AI-generated decisions. |
| High-risk vendor assessment | Vendors touching regulated data, employment, health, finance, critical workflows, or external side effects | Adds legal review, human oversight, audit evidence, incident playbooks, and deeper testing. | Slower procurement and more evidence required from the vendor. | The AI could affect rights, money, safety, privacy, or regulated decisions. |
The most important vendor questions are about data flow: what goes in, what comes out, what is stored, who can see it, what trains models, and how deletion works.
AI vendors increasingly connect to Slack, Drive, GitHub, CRM, email, databases, and ticketing systems. Permission inheritance and connector auditability can matter more than the model itself.
A mature AI vendor should explain how it evaluates model behavior, red-teams risky workflows, handles safety issues, communicates incidents, and gives customers release-change visibility.
Decision Rules
Use a standard SaaS questionnaire for baseline controls, then add AI-specific questions.
Require deeper review when the vendor connects to internal knowledge or external tools.
Ask for evidence, not only policy statements.
Reject or sandbox vendors that cannot answer training, retention, deletion, and permission questions.
Related Guides
Drill into region, retention, and processing questions.
OpenMap AI-specific controls to SOC 2-style evidence.
OpenReview regional storage, processing, retention, and privacy controls.
OpenMap AI product controls to enterprise assurance evidence.
OpenUse NIST, ISO, and EU risk language in procurement.
OpenChinese Archive
Topic Hubs
Industry Pages
Industry page
Compare AI cybersecurity tools for SOC analysts, SIEM, SOAR, XDR, CNAPP, exposure management, DSPM, DLP, PAM, email security, AI GRC, vendor risk, red teaming, and compliance automation.
OpenIndustry page
Compare AI tools for legal teams, contract review, CLM, eDiscovery, compliance automation, GRC, vendor review, document processing, and high-control review workflows.
OpenFAQ
Include baseline SaaS controls plus AI-specific questions about prompt data, outputs, files, embeddings, training use, retention, residency, connectors, model evals, red teaming, and incident response.
SOC 2 is valuable evidence, but AI procurement also needs model, data, retention, permission, and evaluation questions that may not be obvious in a generic SaaS review.
Security, privacy, legal, engineering, procurement, and the business owner should all participate when the AI touches sensitive data, internal systems, or customer-facing decisions.
Source Links
Reference
Official NIST framework for AI risk management.
OpenReference
OpenAI security, privacy, compliance, and accreditation information.
OpenReference
OpenAI business data privacy, security, and compliance page.
OpenReference
Anthropic compliance artifacts and control information portal.
OpenReference
AICPA overview of System and Organization Controls services.
OpenThe strongest decision is always local to your workflow. Save the vendor links, define a representative task, record the exact prompt or command, and compare the final evidence instead of the marketing claim.
Return to the AI learning map