Buyers ask this as one question, but it's two. "Does AI train on your data?" folds together a question about the model provider your product calls and a question about your own company.

OpenAI's API documentation says that since 1 March 2023, data sent to the OpenAI API is not used to train or improve its models unless the customer explicitly opts in (OpenAI: your data). Anthropic says something similar for its commercial products: by default it won't use inputs or outputs from Claude for Work or the Anthropic API to train its models (Anthropic privacy center). Those two sentences settle the provider half for those vendors, under those terms. The half about your own company is where reps get caught.

The provider question

If your product's AI features run on a third-party model, the buyer's data passes through that provider. The provider's commercial terms decide whether it can learn from that data. Anthropic's commercial terms, for example, state that "Anthropic may not train models on Customer Content from Services" (Anthropic commercial terms).

So the first thing to know is which provider, or providers, your product uses, and which agreement covers it. A consumer chatbot account and a business API contract can have very different rules, and a buyer's security reviewer will know that. Saying "we use OpenAI" without saying it's the API under business terms invites the follow-up you were trying to avoid.

The question about you

The second half is whether your company trains anything on customer data: say, a reply-suggestion model that learns from every customer's tickets. Model providers' policies say nothing about this. Your own terms, DPA or AI addendum do.

Find the sentence in your terms that covers it before the call, and know whether there's an opt-out or an opt-in. If your product improves a shared model using aggregated customer data, that's a legitimate design, but it needs to be stated plainly, because a buyer who later finds it in the contract will treat the earlier conversation as misleading. If the answer isn't written down anywhere, that's a question for your product or legal team to answer in writing.

Retention is a separate question

"Not used for training" doesn't mean "not stored." OpenAI's same documentation explains that, by default, abuse monitoring logs are generated for API usage and retained for up to 30 days, unless longer retention is required by law, and that eligible customers can apply for Zero Data Retention or Modified Abuse Monitoring, subject to OpenAI's approval. Security reviewers who know this will ask whether your account has those controls.

That gives you a cleaner way to answer. Training and retention can both be answered from the provider's documents plus your own account configuration. Access is a different matter, covered by your own security documentation, so keep it out of the AI answer unless the buyer asks.

How the question arrives in writing

Some buyers send AI-specific questions alongside the security questionnaire, and some legal teams attach an AI addendum to the contract. The wording varies, so be ready on the substance: whether customer data is used to train any model, whether prompts and outputs are stored and for how long, which providers are involved, and whether the customer can turn AI features off. Answer them with links to the documents above rather than paraphrases, because a paraphrase that drifts from the provider's wording is exactly what a reviewer will flag.

If your product lets admins disable AI features per workspace, say so early. For some buyers that setting decides whether the features can be piloted at all.

Where it shows up in the paperwork

An AI model provider that processes customer personal data on your behalf belongs on your subprocessor list, alongside your cloud host, because GDPR requires the customer's authorisation before a processor engages another processor. Buyers in the EEA will check that list, and they'll check where the provider processes data. If the provider is outside the EEA, the transfer question applies to it like any other subprocessor, and SCCs and the Data Privacy Framework come up on hosting calls too.

It's worth explaining these features in plain English before the paperwork arrives. VoiceVocab's AI & data course practises explaining what an AI feature actually does, where it can go wrong and why the trade-offs matter to a buyer, out loud, with feedback on whether the answer overpromises. The free lesson shows the format.

The feedback button

One detail catches teams out. Anthropic's policy notes that if a user explicitly reports feedback, for example through a thumbs up or down button, the related conversation may be used for training. Check whether your product's AI features pass that kind of feedback to the provider, and whether your own interface has a rating button that does the same thing internally. A buyer's admin will want to know whether their users can opt their data into training with a single click, and whether they can switch that off for the whole workspace.