Short answer: yes, you can. GDPR doesn't ban artificial intelligence. It requires three things: that you have a legal basis for processing the data, that you know where it goes, and that a data processing agreement is signed with the vendor. The problem is almost never the AI itself. The problem is data being entered into some random free tool with no agreement at all.
Let me say right away what this text isn't: legal advice. I'm not a lawyer, but someone who connects AI to company systems every week and sees where things actually break. The formal answer for your case should come from your data protection officer.
Where companies really go wrong
When a manager asks “is this legal?”, they're usually thinking of a complex solution with integrations. Meanwhile the real risk in their company is already happening, just in another place.
A sales manager takes a client contract, pastes it into a free chat tool and asks for a summary. The accountant takes a photo of an invoice and uploads it to get it entered faster. Nobody signed any agreement, nobody knows which continent that data ended up on, and nobody recorded it.
That's the case worth worrying about. A deployed business solution with an agreement and a log is, paradoxically, safer than the current situation, because at least you know what's going on.
The question isn't “should we run data through AI” but “do I know where it's already going right now”.
Three questions before connecting data
- 1. Do I have a legal basis? If the AI does the same job an employee has been doing and the purpose doesn't change (answering an order inquiry, entering an invoice), the basis usually stays the same. A new basis is needed when a new purpose appears: for example, you start using client data for profiling in marketing.
- 2. Where does the data physically go? You have to know which country the servers run in and whether the data is taken outside the EU. That's not a technical detail. It's a line in your privacy policy and your list of vendors.
- 3. Is the data used to train the model? The most important and most often skipped question. More on it separately below.
What “the model learns from your data” means
This is where most of the confusion is, so let's keep it simple. There's a difference between two things:
The system uses your data to answer. The agent opens your warehouse system, sees order 4417 and answers the client. The data is used once, for a specific action, and doesn't stay anywhere. That's normal operation.
The model is trained on your data. Your clients' correspondence becomes material from which the vendor improves its model, and in theory that data could affect answers given to others. That's what you don't want, and that's exactly what's switched off in business solutions: by contract and technically.
In free consumer versions this line is often different than in paid business plans. So “we're using the same ChatGPT anyway” and “we have a deployed solution” are two completely different legal cases, even though the technology is similar.
How we do it
So I'm not talking in the abstract, here are our rules. They're the same for every client:
- The data stays yours. We don't let AI tools collect it or train on it. It's used only for that system to work and for nothing else.
- Every action is logged. The action log lives in your system, not ours: you see what the agent did, when and with which data. Our bill is calculated from that same log, by the way, so you can recompute it yourself.
- A person stays in the important decisions. Where a mistake is expensive, the agent prepares and a person approves.
- Minimal access. The agent is given only what it needs for that one job. Not the whole database “just in case”.
The EU AI Act: what actually affects small business
The new EU artificial intelligence regulation sounds threatening, but for most small and medium business automations it touches less than it seems. The strictest obligations apply to high-risk areas, for example when AI decides on people's employment, credit or education.
A typical customer service or invoice entry agent doesn't fall into that category. Of the practical things, two usually remain: transparency (a person has to know they're communicating with a system, not an employee) and human oversight where the decision matters.
The regulation's obligations come into force in stages, and interpretations are still being refined. If your business is in a sensitive area, don't rely on an article on the internet. Check with a lawyer.
When it's really not worth it
It would be dishonest if I spent the whole text only reassuring you. There are cases where I wouldn't run data through AI, and I tell clients so:
- Health, criminal record and similar special data: if you don't have a clear legal basis for it and experience in this area, that's not the place to start automating.
- If you don't know what data you have at all. When nobody in the company can say where the client lists are and who can access them, the first job is order, not AI.
- If the vendor can't say where the data is stored. One question, one clear answer. If there isn't one, you've already been given your answer.
The minimum list worth having
Not perfect compliance, but a realistic minimum you can start with:
- A signed data processing agreement with the vendor that describes where the data is stored, who can access it and how quickly an incident is reported.
- An updated privacy policy and list of vendors.
- An entry in the record of processing activities about the new way of processing.
- An agreement on how long conversations and logs are kept, and when they're deleted.
- Clarity on who in the company is responsible if something happens.
A normal vendor should help you fill in this list, not leave it to you. What else is worth asking before you choose, I've written up in the article on the difference between chatbots and agents.
Questions about AI and GDPR
Can you run client data through AI without breaching GDPR?
Yes. GDPR doesn't ban using artificial intelligence. It requires that the data processing has a legal basis, that you know where the data goes and that a data processing agreement is concluded with the vendor. The problem is almost never the AI itself, but data being entered into a random free tool with no agreement at all.
Does AI learn from my company's data?
It depends on how the system is configured. In free consumer versions, data may be used to improve the model. In business solutions this is switched off by contract and technically. In our case the data stays the client's: we don't let AI tools collect it or train on it, and it's used only for that system to work.
Do I need separate client consent for AI to process the data?
Usually separate consent isn't needed if the AI does the same job an employee has been doing and the purpose of processing doesn't change, for example answering an inquiry about an order. You need to update your privacy policy and list of vendors. But the exact answer for your case should come from your data protection officer or lawyer.
Who is responsible if data leaks through an AI system?
Toward clients and the supervisory authority, the company is responsible as the data controller, and the vendor's liability is set in the data processing agreement. That's why that agreement isn't a formality: it has to describe where the data is stored, who can access it and how quickly the vendor must report an incident.