To answer the AI section of a security questionnaire well, describe what your system does today, in specific terms, and attach proof for each claim. Buyers want to know five things: whether their data trains a model, where it goes and for how long, who else touches it, whether one customer can see another's data through the AI, and who is accountable. A short answer backed by a contract clause, a config or a diagram beats a long answer backed by nothing.
This guide groups the questions you will see, shows what the buyer is really checking, gives a weak and a strong sample answer for each, and lists the evidence to attach. The samples are generic and redacted. Swap in your own facts, and only claim what your systems can show.
Where the AI questions come from
Most enterprise buyers now ask about AI as a separate block, not as a footnote to data security. The questions usually come from one of three places:
- The SIG. Shared Assessments added an Artificial Intelligence risk domain to its Standardized Information Gathering questionnaire in SIG 2024, mapped to the NIST AI Risk Management Framework. If a buyer sends you a SIG, it may include this domain. (Mitratech)
- The CSA AI-CAIQ. The Cloud Security Alliance published its AI Controls Matrix (AICM) in July 2025, with a companion questionnaire, the AI-CAIQ, for self-assessment and vendor review. It maps to ISO/IEC 42001 and NIST AI RMF 1.0. (CSA)
- The buyer's own form. Many security and privacy teams write their own AI questions, often drawing on NIST AI RMF 1.0, its Generative AI Profile (NIST AI 600-1, July 2024), ISO/IEC 42001:2023 and the OWASP Top 10 for LLM Applications.
The wording changes between forms. The underlying questions don't. If you answer the nine groups below once, properly, you can reuse the answers across all of them.
Four rules for every AI answer

Reviewers read hundreds of these. They spot the same weak patterns fast. Four rules avoid most of them:
- Describe the system, not your intentions. "We take privacy seriously" tells the reviewer nothing. "Prompts are stored for 30 days in our EU region, then deleted by a scheduled job" tells them everything.
- Name the scope. Say which features use AI, which model providers they call, and which data reaches those providers. A reviewer can't assess what you haven't named.
- Attach evidence to every yes. A yes without proof becomes a follow-up question. A yes with a clause number or a config screenshot closes the item.
- Never answer yes to a plan. If a control is planned, say so, with a date and an owner. A yes that turns out to be a plan does more damage than an honest no.
The nine question groups, with sample answers
Each group below shows what the buyer is checking, a weak answer, a strong answer and the evidence to attach. Company and vendor details are placeholders.
1. Data use and model training
Typical question: Is customer data used to train or improve AI models? The buyer is checking whether their data could end up inside a model that serves someone else. This is the question that stalls the most deals.
- Weak: "No, we do not use customer data for training."
- Strong: "No. We do not train or fine-tune any model on customer data. Our AI features call [Provider] through its commercial API. [Provider]'s terms state that API inputs and outputs are not used to train its models (DPA section X, attached). We have not opted in to any data-sharing programme. If we ever introduce training on customer data, it will be opt-in per tenant and covered by a contract change."
- Evidence: the model provider's commercial terms or DPA showing no training on API data, your own DPA clause, and a screenshot of the provider account setting showing data sharing is off.
Check the provider's terms before you write this. For example, OpenAI's API documentation says API data is not used to train its models unless you opt in, and Anthropic's commercial terms say it may not train models on customer content from its services. (OpenAI, Anthropic) Consumer plans of the same tools often have different terms, which matters if your team uses them internally.
2. Data residency and retention
Typical question: Where is data processed by AI features stored, and for how long? The buyer is checking regions, copies and deletion. Prompts, outputs, embeddings and logs are all copies of their data.
- Weak: "Data is stored securely in the cloud and deleted when no longer needed."
- Strong: "Prompts and outputs are processed in [region]. We store them in our database for 30 days to support the history feature, then delete them. Embeddings are stored per tenant and deleted within 30 days of a tenant deleting the source document or ending their contract. Our model provider keeps API data for up to 30 days for abuse monitoring under its standard terms. Application logs contain request IDs, not prompt text."
- Evidence: retention configuration or the deletion job's code, a data flow diagram showing every store, and the provider's retention terms.
Provider retention is the detail founders most often miss. OpenAI, for instance, keeps API abuse monitoring logs for up to 30 days by default, with zero data retention available only on approval and not for every endpoint. Quote your provider's actual term, not a guess.
3. Model providers and subprocessors
Typical question: List all third parties involved in AI processing. The buyer is checking whether a new subprocessor has appeared that their privacy team hasn't approved.
- Weak: "We use leading AI providers."
- Strong: "AI features use two subprocessors: [Provider A] for text generation and [Provider B] for embeddings, both listed on our subprocessor page. We do not use any other AI service with customer data. We give 30 days' notice before adding a new AI subprocessor, as set out in our DPA."
- Evidence: your subprocessor list, the notice clause in your DPA, and each provider's security report or certification summary.
4. Tenant isolation
Typical question: How do you prevent one customer's data from appearing in another customer's AI output? The buyer is checking retrieval. If your AI searches a shared vector index, a missing filter can leak another tenant's documents into an answer.
- Weak: "Our architecture ensures data isolation."
- Strong: "Each tenant's embeddings sit in a separate namespace. Every retrieval query is filtered by tenant ID on the server, taken from the authenticated session, never from the request body. Prompts include only data from the requesting tenant. We have automated tests that try cross-tenant retrieval and fail the build if it succeeds."
- Evidence: an architecture diagram of the retrieval path, the test names and a recent passing run, and the relevant section of your latest penetration test.
5. Prompt injection and output safety
Typical question: What controls address prompt injection and unsafe outputs? Prompt injection is first on the OWASP Top 10 for LLM Applications 2025. The buyer is checking whether a crafted document or message could make your AI leak data or take an action it shouldn't.
- Weak: "We use guardrails to block malicious prompts."
- Strong: "We treat all model input from users and documents as untrusted. The model has no direct access to the database or to other tenants' data. Tool calls run with the requesting user's permissions, and actions that change data need user confirmation. Outputs are rendered as text, not executed. We test against the OWASP Top 10 for LLM Applications each release; the latest results are attached."
- Evidence: red-team or test results with dates, a list of the tools the model can call and their permissions, and the relevant pen test findings with their fix status.
Don't claim you have solved prompt injection. Nobody has. Reviewers trust an answer that limits what a successful injection can do far more than one that says it can't happen.
6. Human oversight
Typical question: Are AI outputs reviewed by a human before they affect decisions? The buyer is checking whether your AI makes decisions about people or money without a person in the loop.
- Weak: "Yes, humans are always in control."
- Strong: "Our AI drafts summaries and suggested replies. It does not send, approve or reject anything on its own. Every output is shown to the user as a draft, labelled as AI-generated, and needs a click to use. Admins can turn each AI feature off for their organisation."
- Evidence: screenshots of the draft and label in the product, and of the admin toggle.
7. Logging and monitoring
Typical question: Do you log AI interactions, and can customers access those logs? The buyer is checking two things at once: that you can investigate misuse, and that your logs aren't a second, uncontrolled copy of their data.
- Weak: "All activity is logged and monitored."
- Strong: "We log each AI request with user ID, tenant ID, feature, model, timestamp and token count. Prompt and output text are not written to application logs. Admins can export AI usage per user from the audit log. We alert on unusual volume per tenant. Logs are kept for 12 months."
- Evidence: a sample log entry with data redacted, a screenshot of the admin audit view, and your log retention setting.
8. Incident response
Typical question: Does your incident response plan cover AI-specific incidents? The buyer is checking whether you have thought about a model leaking data, a provider breach, or a feature producing harmful output, and whether you will tell them.
- Weak: "Our incident response plan covers all incidents."
- Strong: "Our incident response plan has an AI section covering data exposure through AI output, model provider incidents and harmful output. Any AI feature can be switched off globally or per tenant within minutes using a feature flag. We notify affected customers of security incidents within the time set in our contract."
- Evidence: the AI section of your incident response plan, the kill-switch mechanism, and the date of your last tabletop exercise.
9. AI governance policy and ownership
Typical question: Do you have an AI governance policy, and who owns it? The buyer is checking whether anyone is accountable. A named owner tells them questions will get answered after the contract is signed.
- Weak: "We follow industry best practices for responsible AI."
- Strong: "Yes. Our AI use policy covers approved models and providers, which data may be sent to them, review before launching a new AI feature, and employee use of AI tools. It is owned by our CTO and reviewed every six months. We use NIST AI RMF as a reference. We are not ISO/IEC 42001 certified."
- Evidence: the policy itself or a summary, the approval record for your current AI features, and your list of approved internal AI tools.
The last sentence of that answer matters. Saying plainly what you don't hold is better than letting a reviewer assume you hold it. Employee use of AI tools comes up here too; if you haven't mapped it, a shadow AI audit is the place to start.
The evidence to keep ready
Most of the evidence is reusable. Build it once and keep it current:
| Question group | Evidence to attach | Who usually owns it |
|---|---|---|
| Data use and training | Evidence to attach: Model provider DPA or terms showing no training on API data; your DPA clause; provider account setting | Who usually owns it: Legal and engineering |
| Residency and retention | Evidence to attach: Data flow diagram; retention config or deletion job; provider retention terms | Who usually owns it: Engineering |
| Subprocessors | Evidence to attach: Subprocessor list; DPA notice clause; provider security reports | Who usually owns it: Legal |
| Tenant isolation | Evidence to attach: Retrieval architecture diagram; cross-tenant test results; pen test section | Who usually owns it: Engineering |
| Prompt injection and output | Evidence to attach: Red-team or OWASP-based test results; tool permission list | Who usually owns it: Engineering and security |
| Human oversight | Evidence to attach: Product screenshots of drafts, labels and admin toggles | Who usually owns it: Product |
| Logging and monitoring | Evidence to attach: Redacted sample log; audit view screenshot; log retention setting | Who usually owns it: Engineering |
| Incident response | Evidence to attach: AI section of the IR plan; kill-switch; tabletop record | Who usually owns it: Security |
| Governance and ownership | Evidence to attach: AI use policy; feature approval records; approved tools list | Who usually owns it: CTO or founder |
If you sell to Indian enterprises too, much of this overlaps with what they ask under the DPDP Act. Our vendor evidence pack covers the privacy side of the same file.
How to answer "no" honestly

Every company has gaps. Reviewers expect some. What they can't accept is finding out later that a yes was really a no. A good no has three parts:
- The fact. "We do not currently offer per-tenant opt-out of AI features."
- The compensating control. "AI features are off by default for new organisations and only enabled when an admin turns them on."
- The plan, if there is one. "Per-feature admin controls are scheduled for next quarter, owned by our Head of Product."
Use "not applicable" only with a reason: "Not applicable. We do not fine-tune or host our own models." A bare N/A reads like you skipped the question.
Some gaps are cheaper to fix than to explain. Missing tenant filters, prompts written to logs, no retention limit and no admin toggle are often small engineering changes. Fixing them before you submit turns a no into a yes you can prove.
Mistakes that create follow-up rounds
- Copying the provider's marketing. The buyer is assessing you, not your model provider. Say what you configured, not what the provider offers.
- Different answers in different forms. Security and privacy teams compare. Keep one answer library and reuse it.
- Forgetting internal AI use. Engineers pasting customer data into a chatbot is in scope for many questionnaires.
- Claiming certifications you don't hold. ISO/IEC 42001 is a certifiable standard. Following it is not the same as being certified, so say which one applies.
If your deal is already waiting on these answers, read why enterprise deals stall in AI security review for what the buyer's side looks like.
Where Mono helps
We help teams answer AI security questionnaires. We review the AI section of a live questionnaire, find the gaps between what you want to claim and what your systems show, and fix the engineering behind those gaps: logging, tenant data isolation, retention, model provider terms and admin controls. Then we assemble the evidence so each answer has proof attached. It sits inside our wider AI governance work for SaaS, and if you also want to check the rest of your stack, a tech audit covers it.
This is engineering and evidence work, not legal advice and not a certification. Your lawyers still own the contract language.
FAQ
What is the AI section of a security questionnaire?
It is a set of questions about how a vendor's AI features handle customer data and risk. It usually covers model training, data retention, model providers, tenant isolation, prompt injection, human oversight, logging, incident response and AI governance. It appears in standard questionnaires such as the SIG and the CSA AI-CAIQ, and in buyers' own forms.
Does the SIG questionnaire include AI questions?
Yes. Shared Assessments added an Artificial Intelligence risk domain in SIG 2024, mapped to the NIST AI Risk Management Framework. If a buyer sends you a SIG that includes this domain, expect questions on AI data use, risk management and governance.
How should I answer whether customer data is used to train AI models?
State plainly whether you train or fine-tune on customer data, name the model provider you use, and cite the clause in its commercial terms or DPA that covers training on API data. Attach that document and a screenshot of the account setting showing data sharing is off. If training is ever planned, say it would be opt-in and covered by a contract change.
Is it acceptable to answer no in an AI security questionnaire?
Yes, and an honest no is far better than a yes you can't prove. Give the fact, any compensating control, and a dated plan with an owner if one exists. Reviewers expect gaps; what damages trust is discovering that a yes was really a plan.
What evidence should I attach to AI questionnaire answers?
The most useful items are the model provider's DPA or terms on training and retention, a data flow diagram, retention settings, tenant isolation tests, prompt injection test results, admin control screenshots, the AI section of your incident response plan and your AI use policy. Build these once and keep them current, because most questionnaires ask for the same proof.
