← All news and guides
AI governance

Customers Ask If You Train on Their Data: How to Answer and Prove It

How to answer "do you train on our data?" for your own models and your providers, with sample answers, provider terms and the evidence buyers accept.

Mono Team ·

A printed questionnaire on a meeting-room table with a pen resting across it, chairs pushed back and a window behind.

When a customer asks whether you train on their data, answer for each way their data could shape a model, then attach proof. That means your own models, your model providers, and the smaller uses buyers also count, such as embeddings, evaluation sets and human review. A strong answer names the provider, quotes the term that rules out training, shows the account setting, and says what happens if that ever changes. "No, we don't train on customer data" on its own invites a follow-up question.

This guide covers what "training" can mean to a buyer, why the honest answer has several parts, what the main model providers say as of November 2026, weak and strong sample answers, the evidence that proves your claim, and what to do if you do train.

What "training" means to the buyer

Your engineers may hear "training" as pre-training a foundation model. The buyer's security and privacy team may mean something wider: any way their data could end up shaping outputs that other customers see, or being read by someone they didn't approve. Treat each of these as a separate question:

Use of dataWhat it isWhat the buyer wants to know
Pre-trainingWhat it is: Building a foundation model from large datasetsWhat the buyer wants to know: Could our data end up inside a model your provider sells to everyone?
Fine-tuningWhat it is: Further training a model on a smaller datasetWhat the buyer wants to know: Do you fine-tune on our data, and is the result used for anyone else?
Embeddings and RAG indexesWhat it is: Turning documents into vectors so the AI can search themWhat the buyer wants to know: Where are they stored, are they kept per tenant, and are they deleted with the source?
Evaluation setsWhat it is: Saved prompts and outputs used to test model qualityWhat the buyer wants to know: Do our real conversations end up in your test sets?
Human reviewWhat it is: People reading prompts or outputs, for quality or abuse checksWhat the buyer wants to know: Who can read our data, in which company, and when?
Product analyticsWhat it is: Usage data about how the AI feature is usedWhat the buyer wants to know: Is the content of our prompts in your analytics, or only counts and metadata?

Strictly, embeddings and retrieval indexes are not training. The model's weights don't change. Say so, but don't use it to dodge the question. A shared index without tenant filters can still surface one customer's documents in another's answer, which is the outcome the buyer is worried about.

Why the honest answer has several parts

"Do you train on our data?" has at least four honest sub-answers, and they can differ:

  • Your own models. Do you train, fine-tune or evaluate any model you build on customer data?
  • Your model providers. Do the companies whose APIs you call train on what you send them? This depends on their terms and your account settings, not on your intentions.
  • AI features inside your other vendors. Your support desk or CRM may have its own AI features switched on, processing the same customer data under that vendor's terms.
  • Your staff. If engineers or support agents paste customer data into consumer AI tools, the terms of those tools apply, and they can differ from the business terms.

A complete answer covers all four, or says plainly which ones are out of scope and why.

Check your model providers' terms and settings

Printed contract pages fanned out on a wooden desk beside reading glasses and a pen, lit by a window.

Read the current terms for the exact service you use. Consumer apps, free tiers and enterprise APIs from the same company can have different rules. The table summarises what each provider's official pages say as of November 2026. It is a starting point, not a substitute for reading your own contract.

ServiceTraining on your dataRetention to know about
OpenAI APITraining on your data: Data sent to the API is not used to train or improve OpenAI models unless you explicitly opt inRetention to know about: Abuse monitoring logs kept for up to 30 days by default. Some endpoints keep application state. Zero data retention needs OpenAI's prior approval
Anthropic API (commercial terms)Training on your data: Anthropic may not train models on Customer Content from its ServicesRetention to know about: Inputs and outputs deleted within 30 days by default. Longer for flagged usage policy violations and for feedback you submit. Zero data retention is by agreement
Google Vertex AITraining on your data: Google won't use your data to train or fine-tune any AI/ML models without your prior permission or instructionRetention to know about: In-memory caching by default, deleted within 24 hours, and can be turned off. Grounding with Google Search keeps query logs for up to three days. Prompts may be logged for abuse detection
Gemini API (Google AI Studio)Training on your data: Paid services: prompts and responses are not used to improve Google's products. Unpaid services: they are, and human reviewers may read themRetention to know about: Paid services: prompts and responses logged for a limited period to detect policy violations
Amazon BedrockTraining on your data: Inputs and outputs are not used to train base models and are not shared with model providersRetention to know about: Retention is set by a data retention mode per Region. Some models keep traffic for up to 30 days for abuse detection, within AWS
Azure OpenAI (Foundry Models sold by Azure)Training on your data: Prompts, completions, embeddings and training data are not available to OpenAI and not used to train foundation models without your permission or instructionRetention to know about: Flagged prompts and completions may be stored for abuse review. Modified abuse monitoring needs approval, and Microsoft says it will stop being available for models released on or after 1 November 2026, from a date still to be announced

Three points are worth checking:

  • Free tiers. The Gemini API's unpaid tier is the clearest example of a service where your data can be used to improve the provider's products. If anyone on your team prototyped on a free tier with real customer data, say so and say when it stopped.
  • Abuse monitoring is not training, but it is retention. Several of the services above keep some prompts for a period to detect misuse. Buyers will ask how long and who can read them. Quote the term.
  • Opt-ins. Training on API data can be switched on by choice, as OpenAI's "unless you explicitly opt in" shows. Check the organisation settings and confirm that nobody on your team has opted in.

Sample answers, weak and strong

The samples below are generic. Swap in your own facts, and only claim what your systems can show.

1. You call a model provider's API and don't train

  • Weak: "We never use customer data for AI training."
  • Strong: "We do not train, fine-tune or evaluate any model on customer data. Our AI features call [Provider] through its commercial API. [Provider]'s terms state that API data is not used to train its models (clause attached). We have not opted in to any data-sharing programme; a screenshot of the account setting is attached. [Provider] keeps API data for up to 30 days for abuse monitoring. If we ever plan to use customer data for training, it will be opt-in per customer and covered by a contract change."

2. You use retrieval over customer documents

  • Weak: "Your data is only used to answer your questions."
  • Strong: "Our assistant searches your documents using embeddings. This does not change any model's weights. Embeddings are stored per tenant, every search is filtered by tenant ID taken from the signed-in session, and embeddings are deleted within 30 days of you deleting the source document or ending your contract. Automated tests attempt cross-tenant retrieval on every build."

3. You train, with permission

  • Weak: "We may use anonymised data to improve our services."
  • Strong: "We fine-tune our classification model on data from customers who opt in through an admin setting. It is off by default. Before training, we remove names, email addresses, phone numbers and account IDs, and we test that step on every dataset. Customers who opt out are excluded from the next training run. Your current setting is shown in your admin console."

The weak answers aren't lies. They just can't be checked. A reviewer can't approve a claim with no scope, no provider name and nothing attached.

The evidence that proves it

A binder of tabbed documents open on a desk next to a closed laptop and a cup of tea, in soft evening light.

Keep the proof in one place so one pack can answer many questionnaires. Each claim needs a matching piece of evidence:

ClaimEvidenceUsually owned by
Our provider doesn't train on our dataEvidence: The provider's commercial terms or DPA clause, with the date you checkedUsually owned by: Legal and engineering
We haven't opted in to data sharingEvidence: Dated screenshots of the provider account and organisation settingsUsually owned by: Engineering
We don't train on customer dataEvidence: Your DPA or terms clause, and a data flow diagram showing no path from customer data to training jobsUsually owned by: Legal and engineering
Training data is controlledEvidence: Access controls on training buckets, and pipeline code that only reads approved datasetsUsually owned by: Data or ML team
Embeddings stay per tenantEvidence: Retrieval architecture diagram, tenant filter code, and cross-tenant test resultsUsually owned by: Engineering
Our data is deletedEvidence: Retention settings, the deletion job and its logs, and the provider's retention termUsually owned by: Engineering
Other vendors' AI features are controlledEvidence: A list of AI features in your SaaS tools and whether each is onUsually owned by: IT or security
Staff don't use consumer AI tools with customer dataEvidence: Your AI use policy, approved tools list, and any blocking or monitoring in placeUsually owned by: IT or security

Architecture is the strongest evidence, because it shows what can't happen, not just what you promise. If training jobs can only read from a bucket that customer data is never written to, that is easier to defend than a policy. Our guide to answering the AI section of a security questionnaire covers the other question groups that sit next to this one.

If you do train on customer data

Training on customer data isn't forbidden. It has to be disclosed, agreed and controlled. If you do it, or plan to:

  • Put it in the contract. Say what data is used, for which models, and whether the resulting model serves other customers. A change of practice should mean a contract change, not a silent policy update.
  • Prefer opt-in. Opt-in is easier to defend than opt-out, because the customer has made the choice rather than having to find the setting. Whichever you choose, show the setting in the admin console.
  • De-identify, and say how. Remove or replace direct identifiers before training, and test that the step works. Be careful with the word "anonymised": data that can still be linked back to a person may still count as personal data under laws such as GDPR and India's DPDP Act.
  • Use tenant controls. Per-tenant fine-tuned models, or excluding named tenants from shared training, can unblock a cautious buyer.
  • Plan for withdrawal. Removing one customer's data from a model that is already trained is hard. Be honest that withdrawal applies to the next training run, and say how often you retrain.
  • Keep evaluation sets separate. If you save real prompts to test quality, apply the same opt-in, de-identification and retention rules to them.

Some buyers will still say no to any shared training. For them, a model that runs in your own cloud or theirs, with no outside provider at all, removes the question. We cover that setup on own your AI.

How to assess the LLM providers inside your product

The training question is one part of a wider check. Every model provider your product calls processes your customers' data on your behalf, so it belongs in your vendor-risk process like any other sub-processor. Buyers may ask to see that check, not just its result. Run it before you start using a provider, and again whenever its terms change.

AreaWhat to checkEvidence to request or keep
Training and data useWhat to check: Whether business data is used to train or improve models, and whether anyone on your team has opted inEvidence to request or keep: The clause in the commercial terms or DPA, and dated screenshots of your organisation settings
RetentionWhat to check: How long prompts, outputs and abuse-monitoring logs are kept, and whether zero data retention is available to youEvidence to request or keep: The retention term, and written approval if you rely on zero data retention
Where data is processedWhat to check: Which regions store and process data for the exact model and deployment type you useEvidence to request or keep: The provider's data residency page and your account's region setting, checked per model
Security assuranceWhat to check: Independent audits and certifications, and the period they coverEvidence to request or keep: The SOC 2 Type 2 report and ISO/IEC 27001 certificate from the provider's trust portal, and ISO/IEC 42001 where offered
Sub-processorsWhat to check: Who else handles your data, and how you hear about changesEvidence to request or keep: The published sub-processor list, and the DPA clause on change notice and your right to object
Breach noticeWhat to check: How quickly, and through which channel, the provider tells you about an incidentEvidence to request or keep: The DPA clause, checked against the notice you owe your own customers
Deletion at exitWhat to check: What happens to your data when you stop using the serviceEvidence to request or keep: The termination and deletion clause
OwnershipWhat to check: Who owns the inputs you send and the outputs you get backEvidence to request or keep: The commercial terms clause on inputs and outputs
Model changesWhat to check: How you hear about model retirements and changes that alter behaviourEvidence to request or keep: The provider's published notices, and your own test results on each new model version

Public terms already answer much of this. As of 20 November 2026, OpenAI's Services Agreement says it "will not use Customer Content to develop or improve the Services, unless Customer explicitly agrees", and that it deletes Customer Content within 30 days of termination. Its DPA gives you 30 days to object to a new sub-processor, and promises breach notice "without undue delay", with no hour figure, so a tighter limit has to be negotiated. Its trust portal lists SOC 2 Type 2 and ISO/IEC 42001:2023 among other reports, with the SOC 2 report gated behind an access request. Anthropic's commercial terms say it "may not train models on Customer Content from Services".

Residency is where answers most often go wrong, because it varies by model. As of 20 November 2026, OpenAI's API supports storage and processing in India for one model only, and Anthropic's first-party API offers no India option, while Amazon Bedrock routes some Claude models only between its two India regions. Check the exact model and deployment you use before you write "data stays in India".

Keep one record per provider with the date you last checked each row, and recheck it when a provider changes its terms, when you add a model or region, and before a big customer review. For the contract side, see AI clauses in enterprise MSAs. For personal data leaving India, see DPDP cross-border transfers and AI APIs.

Where Mono helps

We help teams answer AI security questionnaires, starting with the training question. We check your provider terms and account settings, trace where customer data goes through your AI features, fix the engineering gaps behind the answer, such as tenant filters, retention and admin controls, and assemble the evidence. If a deal is already waiting, read why enterprise deals stall in AI security review. It all sits inside our wider AI governance work.

We are engineers. This is not legal advice, and your lawyers still own the contract language.

Sources: OpenAI API documentation: Data controls in the OpenAI platform. Anthropic Commercial Terms of Service (effective 17 June 2025) and Anthropic Privacy Center: How long do you store my organization's data? Google Cloud documentation: Vertex AI generative AI data governance. Gemini API Additional Terms of Service. AWS: Amazon Bedrock FAQs, Bedrock data retention and Bedrock abuse detection. Microsoft Learn: Data, privacy and security for Foundry Models sold by Azure, and abuse monitoring. All checked 18 November 2026. OpenAI Services Agreement and Data Processing Addendum, OpenAI Trust Portal, OpenAI API data residency guide, Claude API data residency documentation and AWS: Amazon Bedrock expands Claude model availability to India, checked 20 November 2026.

FAQ

How should a SaaS company answer whether it trains AI on customer data?

Answer for your own models, your model providers, AI features in your other vendors and staff use of AI tools. Name the provider, cite the clause in its commercial terms that rules out training, attach a screenshot of the account setting, state the provider's retention period, and say that any future training would be opt-in and covered by a contract change.

Do OpenAI, Anthropic, Google, AWS and Azure train on API data?

As of November 2026, their official pages say business API data is not used for training by default: OpenAI unless you opt in, Anthropic under its commercial terms, Google Vertex AI without your permission, Amazon Bedrock, and Azure OpenAI without your permission. The Gemini API's unpaid tier is an exception: Google uses that data to improve its products. Check the current terms for the exact service you use.

Are embeddings and RAG the same as training a model?

No. Creating embeddings and retrieving documents doesn't change a model's weights. Buyers still ask about them, because a shared index without tenant filters can surface one customer's documents to another. Store embeddings per tenant, filter every search by the signed-in tenant, and delete embeddings when the source is deleted.

What evidence proves you don't train on customer data?

The model provider's terms or DPA clause, dated screenshots of the provider account settings, your own contract clause, a data flow diagram showing no path from customer data to training jobs, access controls on training data, tenant isolation tests and your deletion job and its logs.

Can a SaaS company train on customer data at all?

Yes, if it is disclosed in the contract and controlled. A defensible setup includes opt-in, tested de-identification, tenant controls, separate handling of evaluation sets, and an honest account of how withdrawal works once a model is trained.

Stuck on the training question in a live questionnaire? Send it to us.