In short
When an AI system works with company data, the data typically passes through two places: the supplier's server and the model provider. AI data protection therefore comes down to four questions: where the system runs and stores data, whether the model provider keeps or deletes the data, who the subcontractors are, and whether there is a data processing agreement. Ask for a written answer to all four.
Key points
- The data passes through both the supplier's system and the model provider: ask about each separately.
- Under their paid business terms, Anthropic, OpenAI and Google do not by default train on the data companies send through the API.
- “Doesn't train on it” is not the same as “doesn't store it”: by default, requests may be kept for a while.
- Under Article 28 of the GDPR, personal data requires a data processing agreement, and subcontractors require the company's written authorisation.
- We run our systems in Hungary, Germany or Finland, and use the models on terms without data retention.
What happens to company data when AI works with it?
The data passes through two places: the supplier's system and the model provider. The supplier's system means the server, the database, the backups and the logs; the supplier chooses where these are. The system sends the model provider the text a task needs and gets the answer back; there, the provider's terms decide.
A model provider is a company that runs the language model, for example Anthropic (Claude), Google (Gemini) or OpenAI (GPT). An API is a connection between two programs; the request and the answer travel through it. In our first conversations this is one of the most common questions, and a fair one: an invoice or a customer e-mail almost always contains personal data.
What is the difference between a free chat and business access?
The difference is whether the provider may train its model on the text you type in. With individual chat plans this typically depends on the user's settings; with business and API access, the big providers do not train on company data by default.
- Anthropic: according to its privacy centre, in the Claude Free, Pro and Max plans it trains on conversations if the user allows it in their settings (with some exceptions, for example safety reviews). For business products, including the API, it does not by default.
- OpenAI: according to its help centre, content in individual services such as ChatGPT may be used for training; this can be switched off in the settings. It does not train on ChatGPT Business, Enterprise or API data by default.
- Google: according to the Gemini API terms, in the paid service Google does not use prompts and responses to improve its products; in the free service it does. In the European Economic Area, Switzerland and the United Kingdom, however, the paid-service data rules apply to the free Gemini API and Google AI Studio too.
Our view on this is simple: company data shouldn't go into employees' own free chat accounts, but into a tool used on business terms.
| Provider, business access | Trains on company data? | Retention by default |
|---|---|---|
| Anthropic API | Not by default; under the commercial terms it may not train on customer content | Deletes requests and responses within 30 days; if a request is flagged for a suspected policy violation, keeps it for up to 2 years |
| OpenAI API | No, unless the company switches it on itself | Abuse monitoring logs for up to 30 days; some features (files, assistants, vector stores) keep data until deleted |
| Google Gemini API, paid | No | Logs prompts and responses for a limited period, solely to detect abuse and for legal obligations |
- Anthropic API
- Trains on company data?Not by default; under the commercial terms it may not train on customer content
- Retention by defaultDeletes requests and responses within 30 days; if a request is flagged for a suspected policy violation, keeps it for up to 2 years
- OpenAI API
- Trains on company data?No, unless the company switches it on itself
- Retention by defaultAbuse monitoring logs for up to 30 days; some features (files, assistants, vector stores) keep data until deleted
- Google Gemini API, paid
- Trains on company data?No
- Retention by defaultLogs prompts and responses for a limited period, solely to detect abuse and for legal obligations
The table is based on the providers' pages as available in October 2026; terms may change.
Does the model provider keep or delete the data?
By default it keeps it for a while. How long depends on the provider, the feature used and the contract. Not training and deleting are two separate questions: a provider can store data without training on it.
Anthropic automatically deletes API requests and responses within 30 days; OpenAI keeps abuse monitoring logs for up to 30 days. Both offer zero data retention, but only after approval, and only for eligible API features. According to Anthropic, the provider then does not store the request and the response, except where the law or the investigation of misuse requires it.
So the question for the supplier is: which model do you use, on what retention terms, and where is that written down?
Where does the system run and store data?
You need to ask about two places: where the supplier's server is, together with the backups and the logs, and where the model processes the request. The first is the supplier's decision; the second depends on the model provider and the settings chosen.
Logs contain data too, so ask what goes into them and how long they are kept. If personal data leaves the European Union, the GDPR's transfer rules apply (Chapter V, from Article 44). There is a choice now: since February 2025, OpenAI has offered European data processing to eligible customers for some API projects. A good answer names a country, not just “in the cloud”.
Who are the subcontractors, and why does it matter?
A subcontractor (in GDPR terms, a sub-processor) is a company that has access to the data on the supplier's behalf: the hosting provider, the model provider, an e-mail or a monitoring service. Under Article 28(2) of the GDPR, the supplier may engage one only with the prior written authorisation of the company as controller. With a general authorisation, it has to give notice of changes in advance so that the company can object.
Ask for a list: name, task, country. Under Article 28(4), the same data protection obligations have to be passed on to the subcontractor by contract, and if it fails to meet them, the supplier remains fully liable.
Do you need a data processing agreement?
Yes, if the supplier processes personal data on the company's behalf. Under Article 28(3) of the GDPR, processing is governed by a contract (or another legal act). A data processing agreement (DPA) sets out what the supplier may and may not do with the data.
Under the regulation, it states, among other things, that the supplier processes the data only on the company's documented instructions, which subcontractors it engages, that it allows checks (audits), and that it deletes or returns the data when the contract ends. Model providers have such agreements too: Anthropic's data processing addendum is part of its commercial terms, and OpenAI signs one on request. These are between the supplier and the provider; the company needs its own agreement with its supplier.
The four questions of AI data protection: what to ask your supplier
- 01
Where does the system run and store data? A country for the server, the backups and the logs, and separately for the model processing.
- 02
Does the model provider keep or delete the data? Which model, on what terms, whether it trains on the data, how long it stores it, and whether there is zero data retention.
- 03
Who are the subcontractors? Name, task, country, and how you will be told when the list changes.
- 04
Is there a data processing agreement? Does it cover deletion when the contract ends, the right to audit and the list of subcontractors?
Ask for the answers in writing. If a supplier can't give you straight answers to these, that is an answer too.
What does this mean for a company of 10 to 250 people?
You don't need a legal department. The four questions and the written answers are enough to base a decision on, and it is worth having a lawyer review the contract.
We would do two things first. Decide which AI tool, and which account, company data may go into; employees' own free chat accounts should not be such a place. And for every automation, check what data passes through the system before it goes live.
This is how we build AI automation: we run the systems in Hungary, Germany or Finland, and use Anthropic, Google and OpenAI models on terms under which the provider does not keep the data. Validation rules run, every step is logged, and where money or a customer is involved, a person approves the action. We describe the details on our data protection page.
This is general information, not legal advice.