AI vendor questionnaires: how to answer questions about your AI (AICM, AI-CAIQ and buyer AI policies)
Short answer: An AI vendor questionnaire is the set of questions a buyer asks about how a supplier's product uses AI: which models and providers you use, whether customer data trains models, where data goes, how outputs are grounded and checked, what humans review, and how you govern AI risk. Buyers use their own AI policy questions, an AI section added to a standard questionnaire (the SIG now has an AI domain), or the Cloud Security Alliance's AI-CAIQ, a questionnaire mapped to CSA's AI Controls Matrix (AICM): 243 control objectives across 18 domains, free to download. Answer with specifics: name model providers, say exactly what is and isn't retained or used for training, and describe the controls customers can switch on.
Who this is for: security, product and sales engineering teams at software vendors whose products use generative AI or machine learning.
On this page
- Where do AI questions come from?
- What do buyers ask about AI? (question themes)
- How do you answer AI questions well?
- A real answer set: how Iris answers about its own AI
- Harder questions (fictional examples)
- Evidence checklist
- How Iris helps you answer AI questionnaires
- FAQ
Where do AI questions come from?
| Source | What it is | Notes |
|---|---|---|
| CSA AI-CAIQ | Questionnaire mapped to the AI Controls Matrix (AICM), for self-assessment or vendor evaluation | AICM released July 2025; v1.1 is current per CSA. A completed AI-CAIQ can be submitted to the STAR Registry (STAR for AI Level 1). |
| SIG AI domain | Artificial Intelligence is one of the SIG's 21 risk domains | See the SIG guide |
| HECVAT 4 | Added AI questions for higher-ed buyers | See the HECVAT guide |
| Buyer AI policy | The buyer's own questions, often from legal and privacy | Most common; least standard |
| Frameworks behind questions | NIST AI RMF, ISO/IEC 42001, EU AI Act | Buyers ask whether you align with or hold these. Only claim what you hold. |
What do buyers ask about AI?
Themes in our own words, not CSA or SIG text:
| Theme | What the reviewer is really checking | Evidence |
|---|---|---|
| Model inventory | Which models and providers, hosted where | AI usage policy; subprocessor list |
| Training and data use | Is customer data used to train or fine-tune any model? | DPA clause; provider terms |
| Data flow and retention | What's sent to the model, what's stored, for how long | Data flow diagram; retention policy |
| Grounding and accuracy | How outputs are kept accurate; what happens when the AI doesn't know | Product docs |
| Human oversight | Who reviews AI output before it's used | Product docs; workflow |
| Access and isolation | Can one customer's data reach another's outputs? | Architecture; isolation controls |
| Security of the AI system | Prompt injection, model abuse, logging | Secure development policy; pen test scope |
| Customer controls | Can customers turn features or data collection off? | Admin settings docs |
| Governance | AI policy, owner, risk assessment, incident handling | AI policy; risk register |
| Regulation and standards | EU AI Act role, ISO/IEC 42001, NIST AI RMF | Statement of alignment (only if true) |
How do you answer AI questions well?
- Name providers. "A leading LLM" reads as evasion. Say which provider and through which service.
- Separate training from processing. "Not used to train models" and "not sent to a third party" are different claims. Most generative AI products do send content to a model provider; say so plainly.
- Don't claim zero retention unless it's true for every system that touches the data, including logs and backups.
- Describe failure behavior. What does the product do when it lacks information? Buyers care more about this than about model names.
- Say what humans do. Is output reviewed before use, and is that enforced or configurable?
- Claim standards only if you hold them. "Aligned with" needs a basis; "certified" needs a certificate.
A real answer set: how Iris answers about its own AI
Iris is an AI vendor too. Here's how we answer the common questions, using the wording our security team approves:
| Question theme | Iris's answer |
|---|---|
| Which models do you use? | Iris's AI runs on Anthropic Claude models through Amazon Bedrock on AWS. Your content isn't sent to OpenAI for drafting. |
| Is customer data used to train models? | No. Your data is never used to train AI models. This is in our DPA. |
| What does the AI draw on? | Questionnaire and RFP answers are drafted only from the customer's own Knowledge Map, approved answers and project context. Public web research is used only in the Proposals feature and onboarding, not in questionnaire answers. |
| What happens when the AI doesn't know? | It answers "There is not enough information to answer the question." Those answers are flagged, can't be approved as-is, and are left blank in exports. |
| How can users check outputs? | Every drafted answer links to its sources and carries a High, Medium, Low or Contradictory confidence label. |
| Human review? | Configurable review and approval before you submit, with assignments, multi-step workflows and an audit trail of who changed and approved what. |
| Tenant isolation? | Customer data is logically isolated per customer. |
| Encryption? | Encrypted in transit and at rest. |
| Customer controls? | Admins can turn on privacy mode ("Disable Data Tracking"), which removes prompts, completions and user identity from AI observability traces and turns off product analytics and session replay for the organization. Tag-based access control limits which knowledge each group can use. |
| Certifications for AI? | We don't claim ISO/IEC 42001 certification. |
Harder questions (fictional examples)
For a fictional vendor, "Fernhill Software":
Do you use AI agents that can take actions in customer systems?
"Fernhill's assistant can propose actions (creating a record, sending a draft), but each one runs only after the user confirms. Agent permissions follow the user's own role. See AI Usage Policy §5."
Have you assessed your obligations under the EU AI Act? (honest partial)
"We have completed an initial assessment and believe our features are not high-risk AI systems under the Act; legal review is ongoing and we will update this answer by [date]."
Evidence checklist
- [ ] AI usage policy (models, approved providers, prohibited uses)
- [ ] Subprocessor list including AI providers
- [ ] DPA clause on training and data use
- [ ] Data flow diagram for AI features
- [ ] Description of grounding, abstention and human review
- [ ] Admin controls documentation (opt-outs, privacy settings)
- [ ] AI risk assessment summary
- [ ] Completed AI-CAIQ, if you maintain one
How Iris helps you answer AI questionnaires
- Upload the questionnaire (Excel, Word, PDF) or answer in a portal with the Chrome extension. Excel files are written back in place.
- Drafts only from your documents (AI policy, DPA, architecture docs, approved answers), each citing its sources with a confidence label.
- Gaps stay gaps: if your AI policy doesn't cover a question, Iris says "There is not enough information to answer the question" rather than inventing a control.
- Keep AI answers current: AI questions change fast. Owners, review cadences and expiry dates on approved answers flag stale ones.
Iris is rated 4.8/5 on G2 from 92 reviews (November 2026). Try Iris free on an RFP you've already answered.
FAQ
What is the AI-CAIQ?
The Cloud Security Alliance's Consensus Assessment Initiative Questionnaire for AI: questions mapped to the AI Controls Matrix, for self-assessment or vendor evaluation.
What is the AICM?
CSA's AI Controls Matrix: 243 control objectives across 18 domains, building on the Cloud Controls Matrix. Free to download.
Do we have to say which LLM we use?
Buyers increasingly require it, and an evasive answer usually triggers follow-ups. Name the provider and the hosting route.
Is "we don't train on your data" enough?
No. Buyers also want to know what's sent to providers, what's retained, and what controls customers have.
How does an AI-CAIQ relate to the CAIQ?
Same publisher and approach, different control set. See the CAIQ guide.
See it live
Respond to RFPs
without the grind.
Book a demo and see your real auto-fill rate on a translation
RFP from your pipeline.
