Every message a customer types into your AI chatbot is personal data - names, order numbers, account details, sometimes payment or health information. When that data flows through a chatbot vendor and its model provider, it's flowing through a chain of processors you're responsible for. In 2026, running an AI support bot in the EU means juggling three obligations at once: a lawful basis under GDPR, a signed data processing agreement with every vendor in the chain, and - from 2 August 2026 - telling customers they're talking to an AI. Here's the checklist.
This is an operational checklist, not legal advice. Where a decision turns on your jurisdiction or industry, talk to counsel.
The three obligations, at a glance
| Obligation | Rule | What it means for your chatbot |
|---|---|---|
| Lawful basis | GDPR Art. 6 | You need a reason to process chat data: contract, legitimate interest, or consent |
| Processing agreement | GDPR Art. 28 | A written DPA with the chatbot vendor (and every sub-processor) before data flows |
| AI disclosure | EU AI Act Art. 50 (from 2 Aug 2026) | Tell users, at the start of the interaction, that they're chatting with AI |
All three apply side by side. Missing the DPA is a breach even if you disclose the AI perfectly - and vice versa.
1. Pick your lawful basis
Which basis fits depends on what the chat does:
- Contract (Art. 6(1)(b)): the chatbot handles orders, bookings, or the run-up to an agreement.
- Legitimate interest (Art. 6(1)(f)): the most common basis for answering general customer enquiries - but you must document the balancing test.
- Consent (Art. 6(1)(a)): required if you use conversation data for anything beyond operations - marketing, profiling, training a model.
Training a model on your customers' conversations is a separate processing purpose from answering their questions. It needs its own basis, and it needs to be disclosed.
2. Get a DPA with every link in the chain
You're the controller; your chatbot vendor is the processor. Under Art. 28, a written data processing agreement must exist before processing begins - and it has to cover the whole chain: the chatbot platform, the hosting provider, and the LLM provider behind it. A chatbot built on RAG touches at least three places where data rests or transits - the vector database, the embedding service, and the LLM - and each one needs to be covered.
Before you buy, ask the vendor for:
- A signable DPA you can review in advance.
- A current sub-processor list, naming every downstream company and its location, with notice before new ones are added.
- A named model provider and confirmation that your data is not used to train its models.
- A lawful transfer mechanism if data leaves the EEA - Standard Contractual Clauses or the EU-US Data Privacy Framework.
- A documented retention and deletion schedule, including backups.
A vendor that can't produce a DPA makes compliance impossible before you've written a single line of policy.
3. Tell users they're talking to an AI
Article 50 of the EU AI Act requires that anyone interacting with an AI system be informed they're doing so, at the latest at the start of the interaction. For a standard support bot - the limited-risk tier - that's essentially the whole obligation. A sentence in the chat greeting ("You're chatting with an automated assistant, not a human agent") satisfies it. A disclosure buried in your terms of service does not.
Two practical notes. First, if there's any realistic chance a user could believe they're talking to a human, the "obvious" exception doesn't apply - disclose. Second, pair the AI disclosure with a clear path to a human: GDPR's transparency principles and basic trust both reward it, and it's what customers expect anyway.
4. Minimize, retain, and honor rights
- Data minimization (Art. 5): collect only what the support interaction needs. Turn off session metadata, device fingerprinting, and browser-data collection that isn't operationally required.
- Retention: 30 days covers most support purposes; 90 days if you need follow-up or pattern analysis. Set a concrete period, document it, and make sure the vendor actually deletes on schedule.
- Data-subject rights: customers can ask to access, correct, export, or delete their chat data. You need a way to find one person's messages and remove them - not just from the live database, but from backups on their own expiry cycle.
- Sensitive data: tell users not to type special-category data (health, etc.) into the chat, and redact payment and account numbers where your platform allows it.
5. Document your classification
A standard support chatbot is limited-risk under the AI Act - no conformity assessment, no technical documentation file, no CE marking. But keep a record of that classification analysis, train staff who monitor the bot (AI literacy obligations applied in February 2025), and review the bot's conversation design against the Act's prohibited practices. If your bot ever triages access to essential services or makes decisions with meaningful consequences, reassess - that's where a support bot can tip into high-risk territory.
What to look for in a platform
- A DPA at account creation, with a published sub-processor list.
- A model provider you can name, with training opt-out confirmed.
- Real deletion - data removal that propagates through logs and indexes, not just a "clear cache" button.
- An AI disclosure you can put in the widget greeting without writing code.
- Retention controls so conversations age out on your schedule.
How Chatonn helps you stay in control
Chatonn's BYOK model is worth a close look for privacy-conscious teams: on the Free plan, you connect your own OpenAI, Gemini, Claude, Groq, or Ollama keys, so the AI layer runs on provider relationships you already hold - you know exactly which model processes your conversations and on what terms. On Pro ($49/mo) and Business ($149/mo), Chatonn manages the AI for you. Either way, flat pricing with no per-seat or per-resolution fees keeps the compliance side simple while you work through the checklist above.
Start on your own keys at chatonn.com and compare the plans on the pricing page.
Put this into practice
Chatonn turns your documentation into an AI support agent that answers instantly and hands off to your team when it should.