AI / LLM Vendor Risk Questionnaire — 30 questions every 2026 procurement team is asking

1. 1.1
   List every foundation model in the data path with provider and version (e.g. Anthropic Claude Sonnet 4.5, OpenAI GPT-5, Llama 4 405B). Update cadence?
   Critical
2. 1.2
   Will customer data be used to train, fine-tune, or improve ANY model — yours or a third party's? Default position and any opt-out.
   Critical
3. 1.3
   If you fine-tune, describe the training data sources, licensing, and consent provenance.
   High
4. 1.4
   Are model weights hosted by you, the foundation provider, or in customer infrastructure? Specify per model.
   High

1. 2.1
   Step-by-step data flow from customer input → your service → foundation model provider → response. Disclose each hop.
   Critical
2. 2.2
   Is zero-retention enabled with the underlying foundation provider (e.g. Anthropic ZDR, OpenAI ZDR)? Provide the contractual evidence.
   Critical
3. 2.3
   Data residency — can processing be pinned to specific regions (EU-only, US-only)? Which models support it?
   High
4. 2.4
   Embedding / vector store retention — how long, where stored, customer-managed deletion?
   High

1. 3.1
   EU AI Act risk classification of the offering (prohibited / high-risk / limited / minimal). Justify the classification.
   Critical
2. 3.2
   If high-risk: confirm conformance with Articles 9 (risk mgmt), 10 (data governance), 13 (transparency), 14 (human oversight), 15 (accuracy & robustness).
   Critical
3. 3.3
   NIST AI RMF profile — have you formally adopted GOVERN/MAP/MEASURE/MANAGE? Provide a summary or ISO/IEC 42001 certification.
   High
4. 3.4
   Model card or system card published per release? Provide the URL.
   High

1. 4.1
   Published hallucination / factual-accuracy metrics for your offering, on what benchmark, last evaluation date.
   High
2. 4.2
   How are outputs grounded — RAG over customer data, web grounding, citations? Describe the architecture.
   High
3. 4.3
   Human-in-the-loop checkpoints for high-impact outputs (legal, medical, financial)? Configurable per customer?
   High

1. 5.1
   Red-teaming process before model upgrades — internal team, external lab, cadence, scope. Provide the most recent summary.
   High
2. 5.2
   Bias evaluations across protected attributes (race, gender, age) — methodology, results, remediation.
   Critical
3. 5.3
   Content safety filters — what categories are blocked (CSAM, weapons, self-harm, PII leakage)? Customer-configurable?
   Critical

1. 6.1
   Prompt-injection defense — describe input sanitisation, instruction hierarchy, and output validation controls.
   Critical
2. 6.2
   Data-exfiltration controls — how do you prevent the model from echoing other customers' data or system prompts?
   Critical
3. 6.3
   Rate limits and abuse detection per API key / customer.
   High

1. 7.1
   Confirm customer owns its inputs and the generated outputs. Carve-outs?
   Critical
2. 7.2
   Output IP indemnification — do you indemnify customers for third-party IP claims arising from model output (similar to Microsoft Copilot Copyright Commitment, AWS Bedrock indemnity)? Cap?
   Critical
3. 7.3
   Open-source licensing — are model weights, fine-tuning data, or generated code subject to copyleft (AGPL, GPL, etc.)?
   High

1. 8.1
   Do you support marking AI-generated content (e.g. C2PA, watermarks) when required by Article 50?
   High
2. 8.2
   If users interact with an AI system (chatbot, agent), is that disclosed in the UX you ship? Configurable copy?
   High

Run a free NIST AI RMF / EU AI Act / ISO 42001 audit. ComplianceIQ uses your audit responses and generated policies to populate an answer pack you can paste into any SIG, CAIQ, or custom enterprise questionnaire — in minutes, not days.