Cross-Industry & Business Teams

Does Your AI Agency Train AI on Your Data?

We read 90 AI agency and vendor terms. Only 8 say whether client data trains AI. See what they say and what to ask before you share files.

November 22, 20266 min read
Hand-drawn folder of papers with a padlock, illustrating whether an AI agency trains AI on your data

What matters most

  • Only 8 of 90 AI agency and vendor terms say whether client data is used to train AI.
  • Six of those 8 say they do not train on client data, and two software platforms say they may.
  • Only 2 of 41 terms covering client project work address AI training, and both say no.
  • OpenAI and Anthropic say business and developer data is not used for training by default.
  • Ask which AI providers and account types your data passes through, and put the answer in the contract.

You are about to give an AI agency access to your client files, your inbox or your accounting records so it can build an automation. Someone on your team asks the obvious question: will those documents end up training somebody's AI model? You check the agency's terms. They say nothing.

That silence is the norm. We read what AI agencies and vendors publish about using client data to train AI, and almost none of them address it. The scenario above is an illustration, not a real client.

Here is what matters most:

  • We read 90 terms pages from AI agencies and related firms on November 22, 2026. Only 8 say anything about using client data to train AI.
  • Six of the 8 say they do not train on client data. Two, both software platforms, say they may use data to train or improve their own models.
  • Of the 41 pages that cover client project work, only 2 address training at all. Both say no.
  • The large AI providers are clearer than the agencies. OpenAI and Anthropic both say that data sent through their business and developer services is not used for training by default.
  • Because most terms are silent, the answer you need comes from asking, and getting the answer into the contract.

Where these numbers come from

This is part of our study of AI agency terms. We collected the public terms of service, client agreements and privacy pages of 90 firms that AI assistants surface when buyers ask about hiring an AI automation agency. Forty-one cover client project work, 43 cover only a software product or website, and 6 were unclear. We used AI to pull out what each page says about client data and AI training, and a script checked every quote against the page text.

The sample is thin on this question, and we want to be plain about that. Eight pages is not enough to say what is typical. What it does show is that most agencies do not put a training promise in their public terms, so you cannot assume one exists. The same dataset is behind our study of who owns the work in AI agency contracts.

What the few who say anything say

The six that promise not to train are direct about it:

  • "We do not use your data to train AI models." (an agency doing client work)
  • "We use AI coding assistants only under terms that forbid training on what we send them." (an agency doing client work)
  • "We do not train foundation models on client data."
  • "We do not use your Customer Data to create, train, fine-tune, evaluate, or improve generalized or shared artificial-intelligence or machine-learning models." (a software platform)
  • "We do not use your Content, in original, anonymized, or aggregated form, to train AI or machine-learning models." (a software platform)

The two that keep the right to train are software platforms. One lists among its uses of data: "train, fine-tune, and evaluate the internal AI models, classifiers, and guardrails that power the Platform." Another says it "may use interaction data generated by users to train models."

The second agency quote above is the most useful one. It deals with the real risk, which is usually not the agency itself. It is the AI tools the agency sends your data through.

Where your data actually goes

When an agency builds an automation, your documents typically pass through three places: the agency's own systems, the workflow tools it uses, and an AI model provider. Each has its own terms.

The model providers have published clear defaults. OpenAI says data sent through its developer platform and its business plans (ChatGPT Enterprise, Business and Edu) is not used to train its models unless the customer opts in, and that this has applied to developer data since March 1, 2023. It also says it may keep that data for up to 30 days for abuse monitoring. Anthropic's commercial terms, which cover its developer services and Claude for Work, say it may not train models on customer content.

Consumer accounts are different. A free or personal chat account can follow different rules, so the question is which kind of account the agency uses. An agency that pastes your client list into a personal chat account is not covered by the business defaults.

What to ask before you share files

Ask these five questions, and ask for the answers in the contract or a data processing agreement, not only in an email:

  1. Do you use our data to train or improve any AI model, your own or anyone else's? You want a plain no.
  2. Which AI providers will our data pass through, and on which account type? Business or developer accounts, not personal ones.
  3. How long is our data kept, and where? Including logs, copies in workflow tools and test files.
  4. Who at your company can see it? And whether any work is passed to subcontractors.
  5. What happens to our data when the engagement ends? Deleted, returned, or kept, and on what date.

If the agency cannot answer the second question, it does not yet know where your data goes, and that is the answer.

Why this matters more for regulated work

For firms that handle tax records, legal files, health information or financial accounts, a training clause is not a technicality. Your own clients may have agreements with you that limit where their data can go. A vendor that keeps the right to train on what you send could put you in breach of those agreements without either of you meaning to.

Ask for the five answers above before any files move. Our page on document processing automation explains how we handle client documents, and our page on hiring an AI automation agency covers how we work more broadly.

FAQ

Do AI agencies use client data to train AI models?

Most do not say. Of 90 AI agency and vendor terms we read, only 8 address it. Six say they do not train on client data, and two software platforms say they may. Ask the agency directly and get the answer in writing.

Does OpenAI train on data sent through its developer platform?

Not by default. OpenAI says developer platform data and data from its business plans is not used to train its models unless the customer opts in. OpenAI's own data page says it may keep that data for up to 30 days for abuse monitoring.

Does Anthropic train Claude on business customer data?

Anthropic's commercial terms, which cover its developer services and business plans, say it may not train models on customer content. Consumer accounts follow different terms.

What should a data clause in an AI agency contract say?

It should say the agency will not use your data to train or improve any AI model, name the AI providers your data passes through, state how long data is kept, and say what happens to it when the work ends.

Want help checking an agency's data terms?

If you are about to hand an agency access to sensitive files, book a free strategy call. Bring the proposal and we will tell you what to ask for in the contract.

Cross-Industry & Business TeamsAI Pilot or Discovery Sprint: Which Should You Buy First?Cross-Industry & Business TeamsCan You Get a Refund From an AI Agency? What Their Terms SayCross-Industry & Business TeamsHow Much Do AI Integration Services Cost?