Can you put customer data into ChatGPT? The short answer
Yes, but how you do it depends on the plan you're on. Typing a customer's name or order history into ChatGPT is processing personal data. That's true under UK and EU GDPR alike, whatever you're using it for.
On ChatGPT's personal plans, OpenAI can use that conversation to train its models by default, and there's no contract that makes it your processor. On its business plans, you can sign a data processing agreement. Training is off by default there, and on the higher tiers you can choose to keep the data in Europe.
This guide is for the owner or operations lead of a small UK or EU firm. Your staff already draft and summarise with ChatGPT, and you've started wondering what's going into it. We advise firms on AI, and this question comes up in most of those conversations.
Why this is a data protection question, not a tools question
Processing, under Article 4 of the GDPR, is "any operation ... performed on personal data". That covers typing it into a chat box as much as storing it in a spreadsheet. The European Data Protection Board's ChatGPT taskforce, drawn from regulators across the EU, said as much in its May 2024 report.
It set out four stages where GDPR applies to a large language model: collecting the training data, preparing it, and training the model. The fourth stage is the prompt a user types in, along with what the model gives back. A customer's details typed into a chat box sit in that fourth stage.
The UK's Information Commissioner's Office takes the same view. Its eight questions for organisations using generative AI open with the lawful basis for the processing, not the technology. A December 2024 outcomes report on its year-long consultation kept that position. It added a line worth sitting with: responsibility gets worked out from what actually happens to the data, not assumed from who wrote the contract.
You stay the controller of your customers' data whichever plan you pay for. What changes is what OpenAI becomes alongside you.
Who is the controller, and what OpenAI becomes
On a personal account, whether Free, Plus or Pro, OpenAI keeps a purpose of its own for the conversation. Its "Improve the model for everyone" setting feeds new conversations into training by default. No data processing agreement is on offer for a personal account, at any price.
Paste a customer's details in there, and you've disclosed them to a second organisation making its own decisions about that data. The lawful basis is yours to justify, not OpenAI's.
The position changes across ChatGPT's other plans: Business and Enterprise, plus Edu and the API. None of those trains on your inputs and outputs by default.
You can sign OpenAI's data processing addendum, in effect since 1 January 2026, which sets OpenAI up as a processor acting only on your instructions. EEA and Swiss customers contract with OpenAI Ireland Limited. Other customers contract with OpenAI's US entity.
| Plan | Trains on your data by default | Data processing agreement available | Data residency option |
|---|---|---|---|
| ChatGPT Free, Plus, Pro (personal account) | Yes, unless switched off | No | No |
| ChatGPT Business | No | Yes | No |
| ChatGPT Enterprise, Edu | No | Yes | Yes, including Europe |
| API platform | No, unless you opt in | Yes | Yes, on eligible endpoints |
The renaming is paperwork, not substance. OpenAI renamed ChatGPT Team to ChatGPT Business on 29 August 2025. The features, the admin console and the data terms carried over unchanged, so a policy written under the old name still describes the plan you're on.
A contract alone doesn't settle the question, though. The ICO's guidance on allocating responsibility across the generative AI supply chain says joint controllership can still apply where a business has little real influence over a closed model.
A contract doesn't decide the question on its own. For most small firms using ChatGPT as intended, on an approved plan, for the uses in their own policy, the processor position holds. It's worth rereading if you're doing something unusual with the tool, such as fine-tuning it on your own records.
What a signed plan doesn't cover on its own
A business plan and a training toggle aren't the whole job. Two more pieces complete it.
The first is a lawful basis for the processing itself. Drafting a reply to a customer from their account details rests on the same basis you already use to handle their account. That's usually the contract you have with them, or your legitimate interest in answering them well.
Consent is rarely the right basis here. GDPR consent has to be freely given. It also has to be specific, and revocable. Asking every customer to agree to whichever tool your staff happen to be using that week isn't a workable policy.
The second is a lawful route for the data to reach the US. Most OpenAI processing runs on US infrastructure unless you've set up data residency. OpenAI's addendum relies on the standard contractual clauses for that, with the EU-US Data Privacy Framework's adequacy decision as a second route.
That framework held up against a challenge at the EU's General Court in September 2025. A further appeal is pending at the Court of Justice, and one of the framework's US oversight bodies has lost the members it needs to do its job. Treat the contract's own transfer clauses as the durable protection, not a framework a court could still narrow.
Data residency: what it fixes, and what it doesn't
Data residency decides where the bytes sit. It doesn't decide whether you're allowed to send them there in the first place.
On Enterprise and Edu, plus eligible API endpoints, you can set a new workspace to keep data in Europe rather than the US. That covers conversations, uploaded files and whatever the model generates. ChatGPT Business doesn't offer that choice today, whatever your admin console promises about data controls in general.
Residency helps with a client contract that names where records may be kept, or with your own instinct that customer data belongs closer to home. It doesn't replace the lawful basis and processor contract covered above.
It also doesn't stop OpenAI's own staff or infrastructure providers reaching the data under the terms already agreed; it narrows where the servers physically sit. Ask what a "European data residency" claim locks down before you rely on it for a promise you've made a client.
What the Garante's case against OpenAI shows
Italy's data protection authority, the Garante, opened the highest-profile case against a consumer AI tool in Europe. In a decision dated 2 November 2024, it found several problems at once. OpenAI had processed personal data to train ChatGPT without an adequate legal basis, and it hadn't told users clearly enough what was happening to their data.
It also lacked proper age checks, and hadn't run the public awareness campaign it had been ordered to run after an earlier 2023 data breach. The Garante fined OpenAI a substantial amount and ordered a fresh, six-month information campaign.
On 18 March 2026, the Court of Rome annulled the entire decision. It did so without ruling on any of those findings.
OpenAI had set up its Irish subsidiary, OpenAI Ireland Limited, back in March 2023. Ireland's Data Protection Commission recognised it as OpenAI's lead EU establishment from 15 February 2024. GDPR gives a single lead regulator authority once a company's main European base is settled.
Under that rule, the court found the Garante had already lost the power to decide the case by the time it ruled in November, nine months after jurisdiction had moved. The court had already suspended the fine on the same grounds back in March 2025.
Nothing in that outcome says the original findings were wrong. It says a different regulator, Ireland's DPC, now holds whatever question remains about how OpenAI trains ChatGPT, and that regulator hasn't ruled on it yet. It's also a case about OpenAI as a developer training its models, a different question from whether your firm's use of a business plan is set up properly. Don't read the reversal as an all-clear for either one.
A worked example: sorting what can go where
Say you run a 20-person letting agency in Bristol. A tenant emails to complain about a repair. Your operations lead wants to use ChatGPT to summarise three months of related emails, then draft a reply.
Both jobs need an approved plan with a data processing agreement in place. The tenant's name and address are personal data, and so is their complaint history. There's no basis for running any of it through a tool that might train on it.
Drafting the reply needs the same plan, plus a person reading the draft against the file before it goes. ChatGPT doesn't know your repair obligations or this tenant's history beyond what you paste in.
What shouldn't go in at all: the tenant's bank details from a refund request, or anything from a separate safeguarding note on the file. Those sit outside what the approved use covers, whatever plan you're on. If your team can't tell the difference in the moment, that's a sign the rule needs to be written down. An AI policy earns its place there, and a company second brain can hold the rule where people actually look for it.
When you don't need to change anything
Some firms are already fine. If your staff only put information into ChatGPT that's already public, a published rate card or a job advert already live on your site, none of this applies.
Public information isn't personal data from a customer's file. If your firm holds no personal data that matters, a five-person supplier of parts to other businesses with no consumer customers, say, the processing question mostly falls away too.
And if you've already moved everyone onto an approved business plan with training off and a signed agreement, you've done the substantive work already. The rest of this guide is detail, not a new requirement.
Where it does matter is the everyday middle ground: a firm with real customer records, and staff who reach for whichever AI tool is already open on their phone. That's most firms.
This week's check
Open your ChatGPT account settings. Look at two things: which plan you're actually on, and whether "Improve the model for everyone" is switched on. If it's a personal account and the setting is on, anything typed into it this week could train OpenAI's models, with no contract standing behind that.
Moving the team to an approved plan with a signed data processing agreement is most of the fix. Our guide to the EU AI Act for small businesses covers the separate rules that apply once a tool talks to customers directly, rather than just helping staff draft replies. Our guide to cookie consent covers the same "who else gets the data" question as it applies to your website's own tools.
Questions
If we turn off 'Improve the model for everyone', does that delete what ChatGPT already learned from us?
No. The setting looks forward, not back. Switch it off and your new conversations stop feeding future training. Older conversations already used before you changed it aren't pulled back. Chat history is a separate control too: OpenAI's help pages note it can still show old conversations even with training switched off. Check the setting before anyone pastes something sensitive, not after.
Is this only an OpenAI problem? What about Microsoft Copilot or Google Gemini?
No. Run the same test on any AI tool. Is this a personal account, or an organisation's plan with a signed data processing agreement and training controls you set? Microsoft states that Copilot in Microsoft 365 only draws on content a user is already authorised to see, under the access rules the organisation has already set. The provider's name matters less than which plan it's on.
Do we need a data protection impact assessment before rolling out ChatGPT Business?
Not always. The ICO's screening guidance asks about large-scale processing, systematic monitoring, or new technology likely to affect people's rights. A handful of staff drafting replies from an approved plan falls short of that test. Connecting the tool to your CRM so it reads every customer record often clears it. When in doubt, run the ICO's screening checklist rather than guess.
How we wrote this: drafted with AI, then checked and edited by the people who build our projects. Figures come from our own work or from the linked sources.