Blog
GDPR and AI tools: which business data can you put into them?
· 5 min read
It is Tuesday morning. An employee in your customer service team has received a long, angry email from a customer. She pastes the entire email, including name, address and order history, into a free AI chatbot and asks for a polished reply. The result is fine. It has saved her twenty minutes.
Do you have a problem? That depends on which tool it was, what was in the email and what you agreed beforehand. At most SMEs, nothing has been agreed about this at all, and that is the real risk. Not the AI itself, but the lack of a clear line.
Below is a classification you can use to tell your employees what is and is not allowed. This is practical guidance, not legal advice. If you are unsure about a specific situation, ask your privacy adviser or lawyer.
First: what counts as personal data?
More than most people think. In Article 4, the GDPR defines personal data as "any information relating to an identified or identifiable natural person". A person who can be identified indirectly, for example through an identification number, location data or an online identifier, is covered as well.
That has a practical consequence. Replacing a customer's name with "Customer A" does not automatically make the text anonymous. If it still contains a customer number, an unusual order, an address or a job title at a small company, the person may still be identifiable. Anonymising is harder than removing names.
An email from a colleague, a job application letter or a note about an absence meeting also contains personal data. So it is not just about your customer database.
What happens when you paste something into an AI tool
An AI tool from an external provider is, in most cases, a processor: a party that processes personal data on your behalf. The GDPR defines this in Article 4 as a party that "processes personal data on behalf of the controller". You are then the controller.
What follows from that? Article 28 states that you may only use processors that provide sufficient guarantees, and that the processing must be governed by a contract that sets out, among other things, its purpose, its duration and the type of data. In short: a data processing agreement.
With an employee's free, personal account, that agreement usually does not exist. You then have no agreements about what the provider does with the data, how long it is kept or where it is stored. Whether data ends up outside the EU also plays a role: Article 44 says that a transfer to a country outside the EU is only permitted if the conditions in that chapter are met.
Two other principles from Article 5 help in thinking this through. Data must be "adequate, relevant and limited to what is necessary" (data minimisation), and it may not be used in a way that is incompatible with the purpose for which you collected it. A customer who gave their address for a delivery does not expect it to end up in an external AI tool.
Green, amber, red
For an SME, a simple classification works best. Put it up on the wall, discuss it in a team meeting and adapt it to your situation.
Green: fine in almost any tool.
- Public information, such as your own website, brochures and press releases.
- General questions and brainstorms without data about an identifiable person or customer.
- Text you have written yourself that contains no personal data or confidential agreements.
- Synthetic examples: made-up customers with made-up data to try out a process.
Amber: only in a business environment that you have assessed.
- Customer emails, quotes and contracts. Think of agreements in which you have promised confidentiality, quite apart from the GDPR.
- Names and contact details of customers, suppliers or colleagues.
- Internal documents such as minutes or reports with traceable details.
Amber means: only with a business account covered by a data processing agreement, where you have checked what happens to the data, and where you have decided that it fits the purpose. You choose such a tool as a business, not as an individual employee.
Red: do not do it, unless you have looked into it properly beforehand.
- Health data, such as sick notes stating a reason or medical information about a customer or patient.
- Other special categories from Article 9. These include data on racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership, genetic and biometric data, and data concerning a person's sex life or sexual orientation. Processing this data is prohibited in principle, with exceptions in paragraph 2 of that article.
- Personnel files, performance reviews and anything to do with dismissal or conflicts.
- Passwords, API keys and access credentials. That is not a GDPR question but a security question, and it goes wrong more often than you think.
What you need to arrange beforehand
A ban on its own does not work. Employees use AI because it gets results. If the approved route is cumbersome, they will choose the free route. So give them a business alternative and a short, clear rule.
Also think about what happens when things go wrong. If someone does paste sensitive customer data into the wrong tool, that may be a data breach. Under Article 33, the controller must report a data breach to the supervisory authority "without undue delay and, where feasible, not later than 72 hours" after becoming aware of it, unless it is unlikely to result in a risk to the people concerned. That means employees need to know they can report a slip-up straight away, without being put in the stocks. A culture in which people keep quiet costs you time you do not have.
Are you building an AI application that uses personal data yourself, for example an agent that processes customer emails? Then also look at Article 35: for processing using new technologies that is likely to result in a high risk to people, you must carry out a data protection impact assessment beforehand. Whether that applies in your case depends on the details.
In our own build work, we follow the same line. Before we start, we record what is never shared with production and whether we test with synthetic or real data. That avoids discussions afterwards, and it makes the green category the default path. We wrote earlier about the rules for AI systems themselves in the EU AI Act for SMEs, and about the difference between a chatbot and an agent in chatbot or AI agent.
What does this mean for your business?
- Ask your team which AI tools they already use. Do it without judgement. You need to know what is going on before you can arrange anything.
- Choose one or two business tools and sign a data processing agreement. Also read what the provider says about storing and reusing data, and where the data is kept.
- Make the green, amber and red list and discuss it. One page is enough. Have employees classify examples from their own work, and it will stick.
- Agree what happens after a slip-up. Who reports it to whom, and how quickly? With the GDPR reporting deadline in mind, even a day of silence is too much.
Would you like to know which AI application suits your business and how to set it up responsibly? Get in touch. We are happy to take a look, using real examples from your own processes.