Is our data used to train AI models? How do I answer that on a security questionnaire?

Treat it as several questions, not one. Decide what the row covers (your product, an AI provider you use, your own internal tools), then answer for each from your actual contracts and settings. State the default plainly, say which provider terms or agreement back it, and attach those documents. If you do not know a fact yet, find out before you answer; a guess in a questionnaire can be relied on and written into a contract.

Do you use customer data to train AI models: can I just answer No?

Only if No is true for every scope the row could cover, and you can show it. A short No that is true for your own models but silent about an AI provider you send data to can mislead the reviewer. Name the scope (for example, "We do not train our own models on customer data") and then answer the provider question separately, with the terms that apply to your account.

What do reviewers expect the answer to be?

Two outside sources describe the expectation from the buyer side. Wolfia says the default answer reviewers expect is that customer data is not used to train models, and Delve says vendors should not train on customer data by default. Both are those sources' statements about reviewers, not a rule. What matters is that your answer is true and that you can back it with evidence.

What if we do train or tune on customer data?

Say so, and describe how it works. Wolfia advises naming the opt-in mechanism, how the data is anonymized, and the contract terms that govern the training. Answer only for features and customers where that is true, and make sure the customer's own contract matches what you describe.

We use a third-party AI provider. Is the provider's policy our answer?

Not by itself. The provider's published terms describe the provider's offer, but your answer must reflect your own agreement, plan tier and settings, which can differ from the public page. Read the terms that apply to your account, record where you read them and on what date, and attach them. Your answer is about what happens to the customer's data in your product, so say what you send, to whom, and what the provider may do with it under your terms.

Does this question also cover how our employees use AI tools?

Often it is asked separately, and the two answers must agree. Employee use of AI tools, customer data entered into them, and AI-written code are different rows in the library and in many questionnaires. If your product answer says no customer data reaches an AI provider, your internal-use answer cannot say staff paste customer data into assistants.

How often should these answers be reviewed?

Whenever you change an AI provider, switch a plan or add an AI feature, and on a fixed schedule besides. Name an owner and a last-reviewed date for the set. An answer written before you added a feature or changed a provider is the most common way an AI answer becomes wrong.

Is our data used to train AI models? How to answer on a security questionnaire

Updated 25 October 2026

"Is our data used to train AI models?" looks like a yes or no row, but a one-word answer is rarely accurate. Split the row into what it really asks (your product, your AI providers, your internal tools), answer each part from your own contracts and settings, and attach the documents that back it. This guide gives example wording for three common situations, lists the evidence to attach, and covers the absolute answers that create contract risk. Every fact in the examples is a placeholder for you to replace; nothing here describes what any company or AI provider actually does.

Why this is several questions in one row

Vorp Labs, a vendor of questionnaire software, published a guide on this question on 26 June 2026 (see Vorp Labs on AI training data answers, read 25 October 2026). Its table splits one training-data row into nine sub-questions. They are, in the order the article gives them:

Part of the row	What you need to know to answer
Sending data to a model	Which features send customer data to an AI model at all.
Training your own models	Whether your company trains its models on customer data.
Provider training	Whether the model provider may train on prompts or outputs under your agreement.
Retention	Whether prompts, outputs or logs are kept, where, and for how long.
Retrieval and embeddings	Whether customer content is indexed for retrieval.
Fine-tuning	Whether customer data is used to fine-tune a model.
Evaluation and monitoring	Whether data is used for evaluation or safety monitoring.
Opt-out and settings	What customers can turn off or configure.
Evidence	Which documents support each statement.

Reviewers read a single sentence as covering all of these, so a short answer that is true for one part and silent on another can mislead without meaning to. Nylas lists ten questions a buyer might ask an AI vendor about data use, starting with training and tuning and moving on to sub-processors, data scope, settings and logging (see Nylas on questions for AI vendors, 22 July 2026, read 25 October 2026). Expect follow-ups on any part you leave out.

What reviewers say they expect

Two buyer-side sources describe the baseline, and both are statements about reviewers, not rules. Wolfia says the default answer reviewers expect is that customer data is not used to train models, and that if you do train on it you should describe the opt-in mechanism, how the data is anonymized and the contract terms (see Wolfia on AI training policy, read 25 October 2026). Delve, a compliance vendor, says vendors should not train on customer data by default and suggests asking whether the vendor is contractually prohibited from doing so (see Delve on questions for AI vendors, updated 17 December 2025, read 25 October 2026).

That tells you what the reader is hoping to see, not what you may write. Your answer has to match your contracts and your settings. If the honest answer differs from the default, give it plainly with the scope and the terms behind it; a clear, supported answer reads better than a vague reassurance.

Three situations and an answer shape for each

The wording below is example wording to adapt, not text to paste. Every bracket is a fact only you know. Replace each one, delete any sentence that is not accurate today, and keep the document that supports it. If you cannot fill a bracket, you are not ready to send that answer yet.

1. Your product has no AI features

[If true for every part of the product: [Product] does not include AI or machine-learning features that process customer data, and we do not send customer data to any third-party AI service.] We do not use customer data to train models. [Evidence: our sub-processor list, dated [date], which lists no AI providers.] Our employees' use of AI tools is governed separately by our Acceptable Use Policy [and our AI use policy].

Check the claim against reality before you send it. A support tool, a search feature, an analytics add-on or a code assistant can all involve an AI service, and the row may be read to include them. The last sentence keeps the product answer separate from internal use, which a reviewer may ask about on another row.

2. AI features that use a third-party model provider

[Feature names] send [the data categories sent, for example prompt text and the document the user selects] to [provider name] to produce [what the feature returns]. Under [the agreement or terms that apply to our account, with their date], [provider name] [may / may not] use that data to train its models. [Retention: prompts and outputs are kept by the provider for [period / not kept], and by us for [period / not kept].] [Customers can turn the features off at [account / workspace] level in [where].] We do not use customer data to train our own models [or: We use customer data to tune [model] only for customers who opt in; see [contract section]]. [Provider name] is listed on our sub-processor list and was assessed under our vendor risk process on [date].

This is the situation where the answer is easiest to get wrong. Vorp Labs recommends keeping product behavior, provider behavior, internal use and customer-specific terms as separate statements rather than one blended sentence. Read the provider's terms for your plan, not a public summary, record the date you read them, and attach them. Provider terms differ between consumer, business and API offers, and between plans, so this guide deliberately does not say what any provider does. Your contract and your settings are the source.

3. You train or tune on customer data, with an opt-in

[Customer data of the following kinds: [categories]] is used to [train / fine-tune / evaluate] [model or feature] only for customers that have opted in through [mechanism, for example a contract clause or a settings toggle]. Before use, the data is [anonymized or pseudonymized by [method]]. The terms that govern this use are in [contract or order form section]. A customer can withdraw consent by [method], after which [what happens to the data already used and to future data]. [Customer-specific commitments are reviewed by [role] before we answer.]

This follows the three things Wolfia says to describe if you train on customer data: the opt-in, the anonymization and the contract terms. Vorp Labs adds that customer-specific contractual commitments need privacy or legal review before you state them. Have that review happen before the answer goes out.

What to attach

Vorp Labs lists an evidence packet for these answers, and the Delve and Nylas articles name overlapping documents. Attach the ones that apply to your situation:

Sub-processor list. It names every AI provider that can receive customer data. The trust page generator builds a page that lists the sub-processors you tick.
Data processing agreement. The agreement with each customer, and the one you hold with each provider that handles their data.
Your AI use policy. The free AI acceptable use policy template expands the AI rules in your Acceptable Use Policy.
The provider's data-use terms, the version that applies to your account, with the date you read them.
Customer controls. A screenshot or help-article link showing how a customer turns AI features off or opts in or out, if that option exists.
Vendor assessment. The record that you reviewed the provider under your vendor risk policy before it received data.

Answers that create risk

Customers rely on questionnaire answers and often attach them to the contract. Vorp Labs warns against reusable absolute claims that are not tied to a product, provider, contract or configuration. Its examples include saying the company never uses customer data for AI, that data is never retained, that the provider cannot access customer data, or that all training is disabled. Any of these can be true for one setup and false for another.

Absolutes you cannot support. "Never", "all" and "disabled" need a scope: which feature, which provider, which plan, as of which date.
An old answer that no longer matches. A new provider, a new plan tier or a new feature can make last year's reply wrong. Re-read the answer against the current sub-processor list before reusing it.
Quoting a provider's public page as if it were your contract. Your terms govern, and they may differ from the public page.
A guess. Where you do not know a fact, route the question to whoever can confirm it, and say in the answer what is still being confirmed rather than filling the gap.

Keep AI answers consistent with the rest of the questionnaire

A questionnaire often asks about AI in more than one place, and a careful reviewer compares them. The Policyseed answer library separates the AI topics into rows AI-01 to AI-05:

AI-01: Do your employees use generative AI tools, and how is that controlled?
AI-02: Is customer data entered into AI tools?
AI-03: Is our data used to train AI models?
AI-04: Is code written with AI assistants reviewed before it ships?
AI-05: Does your product use AI or machine-learning features that process customer data?

The row "Is our data used to train AI models?" and the row on product AI features are marked as needing your own facts, because the library cannot know whether you or your providers train on customer data. Before you send, check that these agree: what your product sends to AI providers, what staff may paste into AI tools, whether AI-written code is reviewed, and the sub-processor list. The same approach as in answering a questionnaire without a SOC 2 report applies: state a status, give the evidence, and keep the same scope language across related answers. The SOC 2 page for AI companies covers where these questions come from if you sell an AI product.

Assign an owner and a review date

Put one named person in charge of the AI answers, usually whoever owns vendor review, and give the set a last-reviewed date (Vorp Labs lists a named reviewer and a review date as part of the evidence packet). Review it when a provider, plan or AI feature changes, and on a fixed schedule such as each quarter. Store the answers in one place so sales and engineering copy from the same text.

To start from drafted wording, the free draft answers tool and the library cover the 5 AI rows and mark what only you can fill in. When a customer sends their own spreadsheet, the questionnaire tool drafts answers from your policies, up to 150 questions per questionnaire, and flags what needs your input. It is included with the Audit Kit ($59), or on its own as the Questionnaire Pack ($19). A draft is only a draft: read every answer and replace every placeholder.

Frequently asked questions

Is our data used to train AI models? How do I answer that on a security questionnaire?: Treat it as several questions, not one. Decide what the row covers (your product, an AI provider you use, your own internal tools), then answer for each from your actual contracts and settings. State the default plainly, say which provider terms or agreement back it, and attach those documents. If you do not know a fact yet, find out before you answer; a guess in a questionnaire can be relied on and written into a contract.
Do you use customer data to train AI models: can I just answer No?: Only if No is true for every scope the row could cover, and you can show it. A short No that is true for your own models but silent about an AI provider you send data to can mislead the reviewer. Name the scope (for example, "We do not train our own models on customer data") and then answer the provider question separately, with the terms that apply to your account.
What do reviewers expect the answer to be?: Two outside sources describe the expectation from the buyer side. Wolfia says the default answer reviewers expect is that customer data is not used to train models, and Delve says vendors should not train on customer data by default. Both are those sources' statements about reviewers, not a rule. What matters is that your answer is true and that you can back it with evidence.
What if we do train or tune on customer data?: Say so, and describe how it works. Wolfia advises naming the opt-in mechanism, how the data is anonymized, and the contract terms that govern the training. Answer only for features and customers where that is true, and make sure the customer's own contract matches what you describe.
We use a third-party AI provider. Is the provider's policy our answer?: Not by itself. The provider's published terms describe the provider's offer, but your answer must reflect your own agreement, plan tier and settings, which can differ from the public page. Read the terms that apply to your account, record where you read them and on what date, and attach them. Your answer is about what happens to the customer's data in your product, so say what you send, to whom, and what the provider may do with it under your terms.
Does this question also cover how our employees use AI tools?: Often it is asked separately, and the two answers must agree. Employee use of AI tools, customer data entered into them, and AI-written code are different rows in the library and in many questionnaires. If your product answer says no customer data reaches an AI provider, your internal-use answer cannot say staff paste customer data into assistants.
How often should these answers be reviewed?: Whenever you change an AI provider, switch a plan or add an AI feature, and on a fixed schedule besides. Name an owner and a last-reviewed date for the set. An answer written before you added a feature or changed a provider is the most common way an AI answer becomes wrong.

Policyseed provides governance policy templates and AI tailoring. It is not legal advice and not a compliance guarantee. Management adopts the policies; the CPA firm performs the SOC 2 examination.