Where Your Data Goes: "We Don't Train" Is a Contract, Not an Architecture
A question we hear in almost every first conversation about AI goes like this. When our people put something into ChatGPT, Gemini or Claude, is it stored? Where? Is it used to train the next model? And then, often, a second sentence: the provider says it doesn't train on business data, but we don't quite believe it.
That second sentence is the interesting one. The instinct behind it is healthy. It is just aimed at the wrong thing.
It depends on the plan, not the provider
"ChatGPT" is not one product with one set of terms. Neither is Gemini, Claude or Mistral. Each is a family of products, and the terms change from one to the next.
On ChatGPT Free, Plus and Pro, training is on by default; the user can switch it off. On ChatGPT Business, Enterprise and the API, it is off by default. On the Gemini app with a personal Google account, conversations train the model when "Keep Activity" is on, which is the default for adults; on Workspace, Google does not train on your data without permission. Mistral Vibe Free and Pro train by default with an opt-out, Vibe Enterprise does not, and Vibe Teams stays on until an admin switches it off.
So the first useful question is not "which provider?" It is "which plan is each of our people actually using?" In most companies, the honest answer includes at least one team on a personal account doing business work. That is the layer with no data processing agreement with your company at all.
Two promises that get blurred
On business plans and APIs, all four major providers make some version of the same promise: we don't train on your content. It is a real, enforceable promise, and it matters.
It is also a promise about what the provider does with your data after it arrives. The data still arrives. It is processed on the provider's infrastructure, it may be logged for abuse monitoring (up to 30 days on OpenAI's API, 55 days on the Gemini API, 30 rolling days at Mistral), and it can be read by staff in defined cases.
"The data doesn't leave" is a different kind of statement. It is not a contract. It is an architecture, and it is only true when the model runs inside your own perimeter, typically an open-weight model on servers you control. With any cloud model, from any provider, in any region, the data passes through the provider.
That is why "we don't train on business data" fails to reassure people. It answers a question about training. They were asking about exposure.
Even the European option needs configuring
A common next step is to pick a European provider and consider the question closed. It is a reasonable instinct, and it still needs work.
Mistral hosts in the EU by default, but its default API endpoint has no committed inference location; EU-only inference requires its EU endpoint, at a 10% premium. Anthropic's own API has no EU option; EU-bound processing goes through AWS or Google, which then become the processor. OpenAI's EU data residency does not cover ChatGPT Business, and even where it applies, OpenAI notes that non-GPU processing may still occur globally.
None of these is a reason to avoid a provider. Each is a reason to decide deliberately, and to write the decision down.
Choose by data class
The question that settles this is not "which provider do we trust?" It is "which class of data may reach which kind of model?"
Public and internal content can go to a business plan or API with training off and a DPA in place. Personal data needs that plus an EU location configured explicitly and a known transfer basis. Special-category data under GDPR Article 9, such as clinical records, belongs by default on a model inside your perimeter, with network egress denied by default, because that is the one arrangement where "it doesn't leave" is true.
Once that policy exists, provider choice becomes something you can revisit. Which brings up the last point: if you want to change models next month, that is not a provider decision either. It is an architecture decision. Agents call one model gateway, the gateway routes each request by data class, and swapping a model means changing a route rather than rebuilding everything.
The full comparison
We put the details in one place: a plan-by-plan comparison of OpenAI, Google, Anthropic and Mistral covering training, retention, human review, EU storage and processing, zero data retention, DPAs and BAAs, with a source for every statement, verified in November 2026. It also covers the legal frame that applies to a European company, the gateway pattern, and the questions to ask any provider before you sign.
