Giacomo Mantovani

Article · November 12, 2026 · By Giacomo Mantovani

Using ChatGPT with customer data: what GDPR allows and what it doesn't

Pasting a customer email into ChatGPT can be lawful, depending on the plan. How free accounts, business plans, the API and EU-hosted models differ.

  • GDPR
  • ChatGPT
  • data privacy
  • AI in production
Contract first. Prompt second. Personal account, Business plans, API, Azure, Your own servers.

In most companies someone has already pasted a customer complaint into ChatGPT to get help with the reply, or dropped in an order export to summarise. Usually from a personal account, and usually without telling anyone. The company then has no contract with the provider, no idea where the data went, and nothing to show a customer or a regulator who asks. This guide compares the options available in October 2026 and sets out the rules I apply on projects for banks and healthcare.

Can you put customer data into ChatGPT without breaching GDPR?

Yes, if you pick the right product and configure it. GDPR does not ban external AI services. It requires a lawful basis, a contract with whoever processes data on your behalf (the data processing agreement, or DPA, under Article 28), safeguards for transfers outside the EU, and that you use only the data you need. The UK GDPR keeps the same requirements. Most of the practical risk sits in personal accounts used for work, where there is no contract and conversations can end up in training data.

For individual services, OpenAI says your content may be used to train its models unless you switch off "Improve the model for everyone" in the data controls (OpenAI Help Center, accessed October 2026). Turning it off still leaves the company without a DPA and without any control over employees' accounts.

How do free ChatGPT, business plans and the API differ for customer data?

For customer data the main difference is the contract. OpenAI states that by default it does not use data from ChatGPT Business, Enterprise, Edu or its API platform, inputs or outputs, to train or improve its models (OpenAI, Business data, accessed October 2026). It will sign a DPA for ChatGPT Business, Enterprise and the API, but you have to request it through a form (OpenAI, Enterprise privacy, accessed October 2026).

On location, OpenAI introduced European data residency in February 2025 for the API, ChatGPT Enterprise and Edu. API customers choose Europe as the region when they create a new project; Enterprise and Edu store content at rest in Europe for new workspaces that opt in (OpenAI, February 2025). The other route is a cloud provider. For Azure, Microsoft states that prompts and completions are not available to OpenAI and are not used to train the models, and that a DataZone deployment in an EU member state keeps processing within EU member states (Microsoft Learn, May 2026).

Options compared (checked October 2026)
OptionUsed for training?Where data livesDPAFits
ChatGPT Free or Plus, personal accountYes by default, can be switched offNot chosen by the companyNoPersonal use, never customer data
ChatGPT Business or EnterpriseNo by defaultEnterprise: Europe for new workspaces that opt inYes, on requestDay-to-day staff use under a written policy
OpenAI APINo by defaultEurope for new projects set to the EU regionYes, on requestIntegrations with your CRM, helpdesk or ERP
OpenAI models on AzureNoEU region or DataZone you chooseMicrosoft DPARegulated processes and sensitive data
Open-source models on your own serversNoWherever you run themNo model provider involvedVery strict constraints, if volume covers the cost

Where should data live when you use AI in a regulated industry?

On my projects for banks and financial services the answer was the same every time: models on Azure OpenAI inside the client's EU cloud, no sensitive data in logs, and an audit trail on every production document. The financial statement pipeline, which handles more than 2,500 documents a day, runs with 16 workers on the client's own OpenShift, inside infrastructure the bank's security team already controls (the case study has the details).

Decide provider and region before anyone writes code. Moving a system halfway through from a generic API to an EU cloud means redoing integrations, permissions and tests. A small company drafting customer replies with ChatGPT needs less control than a bank, but should ask the same questions: who the provider is, where the data is processed, and what the contract says.

Which customer data can you strip out before it reaches the model?

More than people expect. To summarise a complaint, the model only needs the text of the problem; the customer's name, account number and date of birth add nothing. Before sending, replace identifiers with tokens such as CUSTOMER_1 or ORDER_2 and put the real values back into the answer afterwards. That is pseudonymisation: under GDPR the data is still personal, because you can link it back, but the provider receives far less.

IGEA, the voice agent that takes bookings for a medical clinic, keeps each clinic's data separate and handles only admin tasks: bookings, changes and cancellations, with no medical triage. Narrowing the job also cut the health information the system has to process (how the voice agent is built).

What should the logs of an AI system that handles personal data contain?

Logs are the part teams forget. Many systems record full prompts and responses for debugging, and the log store becomes a second copy of customer data, often with weaker protection and longer retention than the original. On the banking projects the rule was that no sensitive data went into logs: we recorded the document ID, check results, timings and cost, never the content.

Providers keep data too. OpenAI says API data may be retained for up to 30 days to provide the service and identify abuse, unless you qualify for zero data retention (OpenAI, Enterprise privacy, accessed October 2026). Microsoft describes an abuse-monitoring store kept in the resource's geography and a process to apply to turn it off (Microsoft Learn, May 2026). Read that part of the terms before you sign.

Do you need a DPIA before using ChatGPT at work?

Not always. Article 35 of the GDPR requires an impact assessment when processing is likely to result in a high risk to people, for example large-scale processing of health or other special-category data, or systematic evaluation of individuals. Drafting a reply to a complaint in ChatGPT Business with identifiers removed usually falls outside that. A system that reads medical records or decides who gets credit does not.

If in doubt, run the assessment with your DPO or privacy adviser. It takes a few days, far less than pulling a system out of production.

Have European regulators acted against ChatGPT?

Italy's data protection authority, the Garante, fined OpenAI €15 million in December 2024. Its findings included training ChatGPT on personal data without an adequate legal basis and failing to notify a data breach from March 2023. The Court of Rome upheld OpenAI's appeal in judgment no. 4153/2026, published on 18 March 2026, and the Garante has temporarily removed the decision from its website (Garante, December 2024 release with a March 2026 note).

For a company using ChatGPT, the ruling changes little. That case was about how OpenAI processes data for its own models. When you put your customers' data into an AI tool, the obligations are yours: a contract with the provider, minimal data and an up-to-date privacy notice.

What rules should staff follow when they use ChatGPT?

One page is enough. It lists the approved tools (for example the company workspace only, never a personal account), the data that is never pasted, how to remove names and references before asking for a summary, and who to ask when a case is unclear. A blanket ban does not work: people keep using personal accounts and stop telling you.

A minimal staff policy
FineNever
Ask for a draft reply to a complaint with the customer's name and contact details removedPaste the full email with signature, phone number and contract number
Summarise internal meeting notes in the company workspaceUse a personal account for any work data
Translate a product sheet or an internal procedureUpload ID documents, medical reports or bank statements
Have a pseudonymised text proofreadShare passwords, API keys or credentials

What mistakes do companies make when using AI with customer data?

The most common is letting everyone use their own free account, with no contract and training on by default. Next comes buying a business plan and never requesting the DPA, which does not arrive on its own. Then there are logs holding full prompts and responses for years, and systems that send the whole customer record when one sentence would have done. The most expensive mistake is choosing provider and region after the project has started, when changing them takes weeks.

Next step

Want to use AI on customer data without needless risk?

I help you choose provider, region and architecture, and design logging and checks before any code is written.

Let's talk

If you work in a regulated industry, my page on document AI and RAG for regulated industries explains how I set up pipelines, audit trails and EU cloud deployments.

Frequently asked questions

If I turn off training in my personal account, can I use it for work?

It lowers one risk but leaves the main one: there is no data processing agreement between your company and OpenAI, and the company does not control the account. For customer data, use a business plan or the API with a signed DPA.

Does data sent to the OpenAI API stay in Europe?

It depends on how the project is set up. OpenAI offers European data residency for new API projects created with the EU region; existing projects cannot be switched. On Azure you pick an EU region or DataZone when you deploy.

Does pseudonymisation make data anonymous?

No. If your company can link the tokens back to people, the data is still personal under GDPR. Pseudonymisation reduces what the provider receives and the damage from an incident, but your other obligations remain.

Do I need to update my privacy notice when I use an AI provider?

If the provider processes customer data on your behalf, it belongs among your processors or recipient categories, and any transfer outside the EU or UK has to be described. Settle the exact wording with your DPO or a lawyer.

Author

Giacomo Mantovani

AI Engineer & Consultant for Companies and Agencies

LinkedInGitHub

Next step

Tell me the workflow. We scope it on a free call, then a working version is live within 10 days.

Book a scoping call

Related articles