- Treat every chat transcript as personal data, and set a retention period that you can state plainly in your privacy notice and configure in the tool.
- Get a data processing agreement from every vendor that stores or processes chat content, and check where the data is hosted and which sub-processors touch it.
- Tell visitors at the start of the chat that they are talking to an AI and what you keep, and collect only the details the conversation actually needs.
Chatbot GDPR compliance is less mysterious than the internet makes it sound, and far more ordinary than most owners expect. Ask five website owners whether their chatbot is GDPR-compliant and you’ll hear five answers: “It’s an AI, so it’s exempt,” “Our vendor handles all that,” “We’re not in Europe,” “I’d have to ask someone,” and “What’s GDPR again?” Only the last is plainly wrong. The others are wrong in ways that can cost real money. The General Data Protection Regulation never mentions chatbots, and it doesn’t need to. A chat transcript is information about identifiable people, and most of the regulation’s duties apply to it just as they apply to a contact form.
This is a practical guide to what those duties mean for a website chat. It isn’t legal advice. Your obligations depend on where you and your visitors are, what you do with the data, and what your vendors actually do. But the core list is short, and most small businesses can work through it in an afternoon with a privacy professional on call for the harder questions.
Why chatbot GDPR starts with the transcript
Start with the basic fact. A conversation with your assistant can contain a name, an email address, a phone number, a booking time, a postcode, an IP address, a device identifier, and whatever the visitor chose to mention. Visitors often share more than they intend. “My son has asthma and needs a sitter on Thursday” is a sentence that contains health information about a child, and a sentence that is likely to end up in a log somewhere.
Even a conversation without a name can identify someone when it’s combined with metadata: the page they were on, the time, the device, the booking reference. The regulation’s test is whether a person can be identified directly or indirectly, and transcripts usually pass that test. Treat them accordingly.
Health information and similar special categories carry extra protection under the regulation. If your assistant operates in a clinic, a counselling practice or anywhere else where visitors discuss health, the level of care you need is higher, and you may need a data protection impact assessment before going live.
Who is responsible for what
The regulation gives names to the roles, and the names matter because they determine who signs what.
- The controller decides why personal data is collected and how it’s used. For a website chat, that’s usually you, the business.
- The processor handles the data on the controller’s behalf. The assistant platform is typically a processor.
- Sub-processors are the further companies the processor relies on, such as hosting providers or the model providers behind the assistant.
The regulation requires a written contract between a controller and each processor. That contract, a data processing agreement, should set out the subject matter and duration of the processing, its nature and purpose, the type of data and the categories of people involved, and the processor’s obligations, including deleting or returning data at the end and helping you answer requests.
| Party | Typical role | What you need from them |
|---|---|---|
| Assistant platform | Processor | A signed data processing agreement, clear retention settings, a list of sub-processors |
| Model or AI provider | Sub-processor, if it receives chat content | Confirmation of where content is processed and whether it’s used for training |
| Calendar or CRM you connect | Processor or separate controller, depending on setup | Your existing contracts, updated to cover the chat’s data flow |
| Your own team | Staff acting for the controller | Access limited to what their role needs |
Lawful basis and the first message
Every use of personal data needs a lawful basis. For most website chats, three bases do the work.
Where the visitor is taking steps towards a contract, such as booking an appointment or requesting a quote, processing needed for that step is generally covered by the contract basis. Answering a visitor’s question about your services can often rest on legitimate interests, provided you’ve weighed the visitor’s reasonable expectations and the impact on them. Consent is the right basis for things the visitor didn’t ask for, such as adding them to a marketing list. Consent has to be specific, freely given and easy to withdraw.
The second obligation is transparency. The regulation requires you to tell people, when you collect their data, who you are, why you’re processing it, on what basis, how long you’ll keep it, and what rights they have. A privacy notice on a separate page helps, but it isn’t enough on its own. Visitors need to see the key points at the moment the chat starts.
A workable first message does the job in a few sentences: “I’m an AI assistant for [business]. Conversations are kept for [period] so we can follow up on your request. Our privacy notice explains your rights [link].” It’s short, true, and visible before the visitor shares anything personal.
Write the retention period in plain words: "chat transcripts are deleted after 90 days" is clear, and then configure the tool to match. A promise you can't configure is worse than no promise at all, because it's a statement you'll have to explain when someone asks.
Ask for less, and keep it for less time
Data minimisation is the most useful principle in the regulation for a chat designer. Collect only what the purpose requires. A booking needs a name, a contact method, the appointment type and a time. It doesn’t need a date of birth, a postcode from three years ago or a full medical history. Each extra field is something you must protect, keep, delete on request and explain in your notice.
This is also where chat is riskier than a form. A form asks for specific fields and nothing else. A conversation invites people to share, and visitors will tell you things you never asked for. Design the assistant so that it steers away from sensitive detail it doesn’t need, and so that a visitor who volunteers a health detail gets a reply that doesn’t repeat it back or store it in a free-text field.
Retention is the other half. The regulation’s storage limitation principle says personal data shouldn’t be kept longer than necessary for its purpose. For transcripts, a short window often makes sense, such as 30 to 90 days, long enough to follow up on an enquiry and short enough not to become a long-term archive of other people’s private conversations. Booking records are different. You may need to keep them for longer for tax or contractual reasons, so the retention rule should distinguish between a transcript and the booking it produced.
Answering visitor requests
Visitors have rights over their data, including the right to see what you hold about them and the right to have it erased in many circumstances. Those rights don’t disappear because the data came through a chat. You need a process for handling them.
A practical process looks like this:
- Verify the person’s identity before releasing anything. A request from someone who is simply guessing a phone number should not produce a transcript.
- Find every record that relates to them: the transcripts, the booking, the lead, and any copies in your inbox or CRM.
- Decide whether an exemption applies, such as a legal obligation to keep a record, and document the reasoning.
- Respond within the regulation’s timescale, which is generally one month, extendable in some circumstances.
- Delete what must be deleted, and confirm that the processor has done the same.
Practise this once before you need it. Many businesses discover, at the moment of a real request, that the transcript lives in a tool they can’t search by email address.
Transfers, hosting and where the data goes
If personal data leaves the European Economic Area, the regulation’s rules on international transfers apply. Some countries have an adequacy decision, which means the regulation considers their protection adequate. Where there isn’t one, transfers generally need appropriate safeguards, such as standard contractual clauses, and a transfer risk assessment may be needed.
For a chat, the practical questions are simple to ask and important to answer. Where is transcript data stored? Where is it processed? Which sub-processors touch it, and where are they located? Is conversation content used to train shared models, and can you opt out in writing? A vendor who answers these clearly is easier to work with than one who responds with a link to a general security page.
Automated decisions and the AI Act
Two further rules are worth understanding, even though most small-business chats won’t trigger them.
The first is the regulation’s rule on automated decision-making. It restricts decisions made solely by automated means that have legal or similarly significant effects on a person. A chat that answers opening hours or books a standard appointment is not in that category. A system that automatically refuses a service, a loan or a job application based on an automated profile of the person may be. If your assistant ever moves from answering questions to making decisions like these, you need human review and a much higher level of scrutiny.
The second is the EU AI Act. Its transparency obligations, set out in Article 50, require people to be informed when they’re interacting with an AI system, unless that’s obvious from the circumstances. The date on which these duties apply has been scheduled for August 2026, but timelines for parts of the AI Act have been moving, so check the current guidance rather than relying on any single date in an article. In practice, telling visitors clearly that they’re talking to an AI is good design and a likely obligation, so you should do it regardless.
A data protection impact assessment is required where processing is likely to result in a high risk to people’s rights. For most small websites, a chat that answers questions and books appointments doesn’t reach that threshold. But the screening question is worth asking, and it becomes more important when the assistant handles health, children’s data, or large volumes of sensitive detail.
Do the screening in writing. Note what data the chat collects, why, who can see it, how long it’s kept and what could go wrong. If the answer is “not high risk,” record that too. A short, dated note is far better evidence than a memory of having thought about it.
AI is not an exemption: using a chatbot never reduces a GDPR duty. It usually adds a processor, a new data flow and a retention question, all of which must be documented.
A short checklist to work through
Here is the list to take to your vendor and your privacy adviser. It’s ordered roughly by urgency.
- Confirm a signed data processing agreement with the assistant platform, and list its sub-processors.
- Check where transcripts are stored and processed, and whether content is used for training.
- Set a transcript retention period in the tool, and match it in your privacy notice.
- Add a clear AI disclosure and a link to the privacy notice at the start of the chat.
- Restrict the data the assistant collects to what each purpose needs, and steer away from sensitive detail.
- Write a short process for access and erasure requests, and test it once.
- Record a dated screening note on whether a deeper impact assessment is needed.
- Review the list when you add a new tool, page or language, or change the retention period.
Keep the compliance proportionate
Chatbot GDPR compliance isn’t a special technical achievement. It’s a set of ordinary habits applied to a new data flow: know who processes the data, tell people what happens to it, keep only what you need, and be able to answer a request. Businesses that already run a tidy contact form process will find most of the work familiar.
Voxy answers from your own content and passes conversations to your team when a person is needed. For the broader reasoning on why grounded, bounded answers are easier to stand behind, read how to keep an AI assistant from making things up. You can see how the product works and check its privacy terms before you go live. Before you connect a chat to live visitor data, make sure your vendor agreements are signed and your privacy notice reflects what the chat actually does. For anything specific to your jurisdiction, take advice from a qualified privacy professional.