Legal exposure
Where your data goes: what to ask an AI vendor

The short answer
The business versions of the big AI services don't learn from your documents by default. They do keep your questions and answers for a while, and how long, where, and who else handles them differs by vendor.
The bigger risk is often your own people pasting company documents into a free personal account. Get the vendor's terms in writing, and give staff an approved tool so they don't need the free one.
Think of sending your files to an outside print shop
The shop has to read your pages to do the job. That's fine. What you want to know is everything after: do they keep copies, and for how long? Do they use your pages to train their new staff? Where is the shop? Do they send overflow work to another shop you've never heard of? And when you stop using them, do the copies go in the shredder?
Those are exactly the questions to ask an AI vendor.
What the big business services say, in their own terms
All five of these business offerings say they don't use your content to train their AI models by default. Where they differ is how long they keep what people ask and what the AI answers.
Anthropic (Claude), business terms
Deletes questions and answers within 30 days. Content flagged for misuse is kept up to 2 years.
Doesn't learn from itOpenAI, for developers (the API)
Keeps logs for abuse checks for up to 30 days. "Zero data retention" needs OpenAI's approval.
Doesn't learn from itChatGPT Enterprise and Business
Chats are kept until the user deletes them, unless the company sets its own rule (at least 90 days).
Doesn't learn from itMicrosoft 365 Copilot
Kept as activity history under your own retention rules, and searchable in legal discovery.
Doesn't learn from itGoogle Gemini in Workspace
Kept from 90 days to indefinitely, as your administrators decide.
Doesn't learn from it
Personal accounts are different. On free and personal plans, the rules can be the opposite. Anthropic, for example, changed its consumer terms in 2025: people on its free and paid personal plans now choose whether their chats train its models, and if they say yes, chats are kept for five years (Anthropic, August 2025). Business accounts weren't affected.
That's why staff using personal tools is a common way company documents leak. In 2023, Samsung engineers reportedly pasted source code and meeting notes into ChatGPT, and the company then restricted generative AI on work devices (Korea Economic Daily, May 2023).
What people type may become evidence
Questions and answers are records, like email. In May 2025, a US court in The New York Times v. OpenAI ordered OpenAI to keep all ChatGPT output logs that would otherwise be deleted, including ones users had deleted themselves (order of 13 May 2025). The order was narrowed that October, but most of the logs already kept were held.
On the business side, Microsoft treats Copilot questions and answers as records your own legal team can search and hold. Ask your lawyer whether your retention rules and legal holds cover them.
Assume anything typed into an AI tool could be read in a lawsuit one day.
Three ways your data can travel
Straight to an AI company
You, or your builder, connect directly to OpenAI, Anthropic, Google or another model company, under its business terms.
Best for: clear terms, one company to hold to them, and options such as zero data retention.
Gives up: some choice of region. Anthropic's own service, for example, currently processes in the US or globally, not EU-only.
Inside the office software you already use
Microsoft 365 Copilot or Gemini in Google Workspace, under the agreement you already have.
Best for: keeping AI data under the retention, discovery and region rules you already run, such as Microsoft's EU Data Boundary.
Gives up: some visibility. Microsoft now uses Anthropic as a subprocessor for some Copilot features, and those are excluded from its EU Data Boundary (Microsoft).
Through a specialist vendor
A product built for your kind of work, which usually sends your data on to one of the big AI companies behind the scenes.
Best for: features made for your job, and a vendor that will sign your own data terms.
Gives up: simplicity. Your data passes through two companies, so you need both sets of terms to hold.
What goes wrong
- No approved tool, so people use their own. Banning AI rarely stops it. It moves to personal accounts, where the terms are weakest.
- Reading "doesn't train" as "doesn't keep". Not learning from your data and not storing it are separate promises.
- Not asking who is behind the vendor. Most AI products send your data on to a model company. Its terms count too.
- Forgetting the records. Nobody decided how long questions and answers are kept, or whether a legal hold covers them.
- No plan for the end. When the contract ends, who deletes what, by when, and how will you know? Anthropic's data terms, for example, promise deletion within 30 days of termination (Anthropic DPA).
Questions to ask your vendor, and your lawyer
- Do you use our documents, questions or answers to train or improve any AI model, including through feedback buttons? Where does the contract say so?
- How long do you keep our questions and answers, and who at your company can read them?
- Where are they stored and processed, and can we keep that in our region?
- Which other companies handle our data, including the AI company behind you, and how will you tell us when that list changes?
- When the contract ends, what is deleted, by when, and how will we know?
- For your lawyer: do our retention rules and legal holds cover AI questions and answers?
- For your lawyer: which of our documents may we send to an outside service under our existing contracts and privacy commitments?
Other names you may hear
- Training on customer data
- Using what customers send to improve the AI model, so it can resurface in other customers' answers. Business terms usually rule it out.
- Retention
- How long the vendor keeps your questions and answers before deleting them.
- Zero data retention (ZDR)
- An option where the vendor keeps nothing once it has answered. Usually by agreement, not by default.
- Data residency, EU Data Boundary
- A promise about which country or region your data is stored and processed in.
- Subprocessor
- Another company your vendor uses to handle your data, such as the AI company behind it or a cloud host.
- DPA, data processing addendum
- The part of the contract that says what the vendor may and may not do with your data.
- Legal hold, eDiscovery
- Keeping records because of a lawsuit, and searching them for it.