Bjorn, a planner at a care provider, pastes a referral letter into ChatGPT to get a tidy summary. The name, date of birth and diagnosis go in with it. Convenient, and done in two seconds. Under the GDPR, though, that is processing of special-category data, outside the organisation, without anyone having checked it.
The GDPR was written before everyone had an AI assistant in their browser. That does not change the rules: if you put personal data into an AI tool, you are processing it, and every GDPR obligation applies. So the question is not whether the GDPR applies, but which measures you actually have to take.
#Measure 1: lawful basis and purpose limitation
Every processing operation involving personal data needs a lawful basis, such as legitimate interests, a contract or consent. And the data may only be used for the purpose it was collected for. A customer address you hold for invoicing cannot simply go into an AI tool for some other purpose. So start with the question: why is this data in this prompt, and is that allowed?
#Measure 2: data minimisation
Article 5 of the GDPR requires data minimisation: process no more data than the purpose needs. With AI this is the measure with the most impact, because the problem usually lies in the prompt itself. You can rewrite a support reply without a real name, account number or IBAN. You can sharpen an appraisal without the salary or medical background. Keep the task, drop the identifiable data. See also what you can and cannot share with AI.
#Measure 3: clarifying the processor role
Whoever decides why and how data is processed is the controller. An AI vendor that processes personal data on your behalf is a processor, and in that case a data processing agreement (DPA) under Article 28 GDPR is usually required. It sets out what the vendor may and may not do with the data.
Mind the difference between account types. Many free and personal AI accounts run under consumer terms without a DPA and may use your input to improve the model. For business use involving personal data, a managed business environment with a DPA is the starting point.
That distinction only helps if you know which tools people use, including the ones IT never bought a business account for. Most organisations cannot say until they see the full list, signed contract or not, with each tool scored for risk rather than approved or blocked in advance.
A tool without a decision stays in the Inbox until the organisation decides. Once it is allowed, it gets an entry in the AI register, with use cases, owners and documents. That feeds into your record of processing activities but does not replace it.
#Measure 4: international transfers
Many AI providers are based outside the EU, often in the US. Sending personal data outside the EU is only allowed with a valid transfer mechanism: an adequacy decision, standard contractual clauses (SCCs) or another lawful route. So when choosing an AI tool, check where the data is processed. More on this in data sovereignty and AI prompts.
#Measure 5: appropriate security
Article 32 requires appropriate technical and organisational measures. With AI use, that includes approved tools and account tiers, clear rules on what staff may share and a way to stop sensitive data being sent by accident. So security here covers both the vendor and what people do as they work.
#Measure 6: a DPIA for high risk
Processing likely to result in a high risk to individuals requires a data protection impact assessment. Examples include large-scale processing of special-category data and systematic monitoring. Sensitive AI use often calls for a DPIA, and it is the point at which you test whether the processing is proportionate.
#Measure 7: data subject rights
People have the rights of access, rectification and erasure. So you need to know which personal data passes through AI tools and where it ends up. See also how to turn off training on your data and request deletion.
#How to support these measures in practice
Most GDPR measures converge on one moment: the prompt. That is the moment that decides which personal data leaves the organisation. BeeSensible highlights sensitive data as someone types in browser-based AI tools, so it can be masked with a placeholder such as [NAME1], or removed, before the prompt is sent.
An example: an employee pastes in a customer email containing a name, an IBAN and a phone number. The sensitive data is highlighted. The employee masks the name, which becomes [NAME1], and removes the IBAN. The task, drafting a polite reply, stays intact, but the identifiable data does not leave the organisation. That supports data minimisation (measure 2) and security (measure 5), and in practice purpose limitation too.
Here is that same highlight in the text field itself, as a name or IBAN is picked out before the prompt is sent.
Bjorn's referral letter from the start of this piece is the same problem in document form. The name, date of birth and diagnosis are in the file long before anyone opens ChatGPT, so minimisation has to cover the document as well, not just start at the prompt. A tool that opens the file, highlights what is sensitive and lets him export an anonymised copy covers that step, on the same terms as the browser extension: processed in the EU and not kept.
The infrastructure BeeSensible runs on is ISO 27001 certified; BeeSensible itself is not. BeeSensible does not replace your GDPR work, but it puts the most important measure into practice at the moment it counts. So start with the prompt: what is not in it, you never have to account for.
Further reading: the topic page Compliance and AI and our article on the AI Act alongside the GDPR.
Outlook
