AI and data-use questions in security questionnaires (training, retention, subprocessors)

Security Questionnaires

Author:

Chris Mabry

Short answer: AI data-use questions ask five things: is our data used to train or fine-tune models, which AI providers process it, how long they keep it, where it's processed, and what stops the AI from producing wrong or leaked answers. Answer each one separately, name your model providers as subprocessors, and match your DPA. Avoid absolutes such as "zero retention" or "your data never leaves your environment" unless your architecture and contracts prove them.

In Iris: Iris's own approved wording is "Your data is never used to train AI models," and its answers draft only from your own content.

Who this is for: security and privacy teams at software companies that use AI in their product, and the sales engineers who field these questions.

On this page

  1. How is the question usually phrased?
  2. What is the reviewer really checking?
  3. Which claims should you avoid?
  4. Sample answers (fictional company)
  5. What evidence should you attach?
  6. How Iris answers this about itself
  7. How Iris drafts AI data-use answers
  8. FAQ

How is the question usually phrased?

AI sections are newer and less standardized than encryption or access control, so wording varies widely. These are the recurring intents.

Common phrasing (paraphrased)What it's really asking
"Is our data used to train, fine-tune or improve AI models?"Training on customer content, by you or by your model provider
"Which third-party AI or LLM providers process our data?"Model providers named as subprocessors
"Do AI providers retain prompts or outputs?"Provider-side retention and logging terms
"Where is AI processing performed?"Region and data residency for inference
"Can we opt out of AI features?"An org-level control to disable AI processing or analytics
"How do you prevent one customer's data appearing in another's output?"Tenant isolation in retrieval and prompts
"How do you control inaccurate AI output?"Grounding, citations, human review
"Do humans at your company review our prompts or outputs?"Staff access to AI interaction data, and observability tools
"Do you have an AI governance policy or follow an AI risk framework?"Policy ownership; frameworks such as the NIST AI RMF

What is the reviewer really checking?

  1. Training is a yes/no fact. Buyers want a clean "No" backed by a contract term, and they check whether the promise covers your providers too.
  2. The full processor chain. An AI feature often adds several processors: a model provider, an embedding service, a vector database, a crawler. Reviewers compare your answer to your subprocessor list.
  3. Retention, separately for you and for providers. You store customer data to deliver the service; that's expected. The question is how long, and what providers keep.
  4. Grounding and review. How does the AI decide what to say, and who checks it before it reaches a customer?

Which claims should you avoid?

ClaimWhy reviewers push backSafer pattern
"Zero data retention"You store customer data to provide the service; backups and logs existState the retention period from your DPA
"Your data never leaves your environment"Untrue if a third-party model provider processes itName the provider and the contractual terms
"Our AI never hallucinates"No AI system can promise thisDescribe grounding, citations and abstention
"Trained only on your data"Usually means "retrieves from your data"; the model itself was trained by its provider"Answers are generated from your content using [provider] models"
"Certified for AI" without a named certificateReviewers will ask for itName only certifications you hold, with the certificate

Sample answers (fictional company)

Written for Fernhill Software, a fictional SaaS vendor with an AI assistant. Labelled examples.

Strong answer: training

Answer: No.
Comment: Fernhill does not use customer data to train or fine-tune AI models, and our agreements with our model providers prohibit them from training on it. This commitment is in our DPA §[x].

Strong answer: AI providers

Answer: Fernhill's AI features use large language models from [provider], accessed through [cloud platform], and an embedding service from [provider]. Both are listed on our subprocessor page with their purpose and processing region. Requests are processed in [region].

Weak answer, rewritten: retention

Weak: "We don't retain any of your data."

That contradicts the fact that the product stores customer files.

Rewritten: Fernhill stores customer content to provide the service and deletes it per the retention terms in our DPA §[x] after termination. Our model providers process prompts to generate responses under contractual terms that [describe the provider retention term you can evidence]. See also our data retention answers.

Honest partial: data residency

Answer: No.
Comment: Fernhill processes and stores data in [country] regions only. We do not currently offer EU or other regional data residency.

What evidence should you attach?

QuestionEvidence
TrainingDPA clause; AI or acceptable-use policy
ProvidersSubprocessor list with purpose and region
RetentionData retention policy; DPA retention clause
IsolationArchitecture overview (tenant isolation section)
Accuracy controlsProduct documentation on grounding, citations and review
GovernanceAI governance policy with owner and review date

How Iris answers this about itself

Iris is itself an AI vendor, so here is its own approved wording as a worked example:

  • Training: "Your data is never used to train AI models." This matches Iris's internal policy and DPA, and no training or fine-tuning code exists in the product.
  • Model provider: Iris's AI runs on Anthropic Claude models through Amazon Bedrock.
  • Grounding: RFP and questionnaire answers draw only on the customer's Knowledge Map, approved answers and project context. When those don't cover a question, Iris answers "There is not enough information to answer the question."
  • Isolation: customer data is logically isolated per customer.
  • Privacy control: an org setting, "Disable Data Tracking," removes prompts, completions and user identity from AI observability traces and turns off product analytics and session replay for that org.

Note what's not there: no "zero retention," no "never leaves your environment." Iris uses third-party AI processors and stores customer content to do its job, so it says so.

How Iris drafts AI data-use answers

  • Answers come only from your documents: DPA, subprocessor list, AI policy, architecture notes and past approved answers, each answer cited.
  • Contradiction flags: if your subprocessor page and an old approved answer name different providers, the draft is labelled Contradictory.
  • Abstention: if nothing states your provider's retention terms, Iris says there isn't enough information rather than guessing.
  • Owners and expiry: AI answers change fast. Give each approved answer an owner and an expiration date; expired approved answers move to Needs Review.

Iris is rated 4.8/5 on G2 from 92 reviews (November 2026). Try Iris free on an RFP you've already answered, or a security questionnaire you've already completed.

FAQ

Is using an LLM API the same as training on customer data?
No. Sending content to a model to generate a response is processing; training changes the model. Say which one happens, and what your provider's terms allow.

Do we need to list our model provider as a subprocessor?
If it processes customer personal data on your behalf, generally yes. Confirm with your privacy counsel. See subprocessor questions.

What's a good answer to "how do you prevent hallucinations?"
Describe the controls: retrieval from approved content, citations, confidence or review signals, abstention when sources are missing, and human approval before submission. Don't promise zero errors.

Should we mention the NIST AI Risk Management Framework?
Only if your AI governance policy actually maps to it.

Is there a dedicated AI vendor questionnaire?
Several are emerging. Our framework guides track the main ones.

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.