AI chatbots, GDPR and the EU AI Act: what businesses need to do

On August 2, 2026, one part of the EU AI Act quietly became enforceable for almost every business with a chatbot: you now have to tell people when they are talking to AI. Many teams that switched on an AI agent last year have never looked at the setup since. The greeting may say "Hi, I'm Sam!", the transcripts may be kept forever, and nobody remembers whether the vendor's data processing agreement was signed.

None of this is hard to fix, but it touches two laws that are easy to mix up. The AI Act adds rules about how AI systems behave. GDPR, which has applied since 2018, still governs the personal data in every conversation. This guide explains what each one asks of a business using AI in customer service, and gives you a checklist you can work through in an afternoon.

This article explains the rules in plain terms. It isn't legal advice. If your chatbot makes decisions about credit, insurance, health or eligibility, talk to a lawyer.

Where chatbot setups usually fall short

The gaps we see most often are small and practical:

• The bot hides that it's a bot. A human name, a stock photo and no mention of AI.
• There's no way to reach a person. Customers loop through answers with no exit.
• Transcripts are kept forever, including order details, addresses and the occasional card number someone typed in.
• Nobody checked the vendor. No signed DPA, no idea where data is stored or whether conversations are used to improve the vendor's AI.
• The privacy notice predates the chatbot. It doesn't mention AI, the vendor or how long chats are kept.
• The bot makes decisions it shouldn't, such as refusing refunds or closing accounts with no human review.

Each of these maps to a specific rule, so let's start with what the rules actually say.

What the EU AI Act requires

The AI Act entered into force in 2024 and applies in stages. For customer service, the stage that matters is Article 50, the transparency rules, which apply from August 2, 2026.

Disclosure from the first interaction. AI systems that interact directly with people must be designed so that people are informed they are interacting with an AI system, "from the start of the first interaction in a clear and distinguishable manner", unless this is obvious. The European Commission's guidance says the "obvious" exception should be read narrowly, from the point of view of a reasonably well-informed person (European Commission FAQ).

Who is responsible. Formally, this duty sits with the provider, the company that builds the AI system. In practice your vendor gives you the tools, and you write the greeting and choose the settings. If your greeting hides the AI, the result is non-compliant, whoever is technically liable. Treat it as your job.

What changed this summer. The Digital Omnibus on AI, adopted in June and in force since July 27, 2026, moved the deadlines for high-risk systems to December 2027 and August 2028. It did not delay the Article 50 disclosure rules for chatbots. It gave generative AI systems already on the market a grace period until December 2, 2026 for machine-readable marking of AI-generated content, which matters for image and text generators more than for support bots (Gibson Dunn summary). The Omnibus also softened the AI literacy duty: companies using AI must now take measures to support their staff's understanding of it, rather than guarantee a level of literacy.

Reach and fines. The Act covers AI systems whose output is used in the EU, so non-EU businesses serving EU customers are in scope. Breaches of the transparency rules can be fined up to €15 million or 3% of worldwide annual turnover, whichever is higher.

What it doesn't do. A customer service chatbot answering order and policy questions is not on the Act's high-risk list. That list covers uses such as credit scoring, recruitment and access to essential services.

What GDPR already required

GDPR applies to every conversation that contains personal data, which is nearly all of them. The obligations aren't new, but adding AI makes them easier to miss:

Requirement	What it means for your chatbot	Who does it
Lawful basis	Answering a customer's question usually falls under performance of a contract or legitimate interests	You
Transparency	Your privacy notice names the chatbot, the vendor, the purposes and how long transcripts are kept	You
Data processing agreement	A signed DPA with every vendor that processes chat data, plus a list of their sub-processors	You and the vendor
International transfers	Data leaving the EEA needs a transfer mechanism, such as the EU-US Data Privacy Framework or standard contractual clauses	Vendor, checked by you
Data minimisation and retention	Collect only what you need and delete transcripts after a set period	You, using vendor settings
Data subject rights	You can find, export and delete one customer's conversations on request	You, using vendor tools
Automated decisions	Decisions with legal or similarly significant effects shouldn't be made by AI alone, and customers can ask for a person	You
Security	Encryption, access controls and a record of who can see transcripts	Vendor and you

Step 1: Map where AI talks to your customers

List every place an AI system replies to customers: the website widget, WhatsApp, Instagram and Messenger, email auto-replies, and the help center search. For each one, note the vendor, the language and whether a person can take over. Channels added later, such as WhatsApp, are the ones most often forgotten.

Step 2: Disclose the AI in the first message

Rewrite every AI greeting so the first message says it's AI. Keep it short and friendly:

"Hi, I'm the AI assistant for Northwind Outdoor. I can answer most questions straight away, or connect you with our team."

Do the same on every channel and in every language you serve. If the AI drafts replies that a person checks and sends, you don't need to label each message, but don't let a fully automated reply pretend to be a named employee.

Step 3: Review your vendor before you trust it

Ask every AI vendor the same questions, and keep the answers with your records:

• Where is conversation data stored, and where is it processed, including by the AI model provider?
• Is there a DPA, and is it part of the terms or a separate signature?
• Who are the sub-processors, especially the AI model providers, and where are they?
• Are our conversations used to train or improve models, and can we opt out?
• What retention and deletion controls exist, including for single customers?
• Which certifications or audits does the vendor hold, such as SOC 2 or ISO 27001?

If a vendor can't answer clearly, that's an answer too.

Step 4: Set a retention period and stick to it

Decide how long you need transcripts: long enough to handle disputes, returns and chargebacks, and no longer. Your warranty period and the time customers have to bring a claim in your country are a sensible guide. Set automatic deletion where the tool supports it, and write the period into your privacy notice.

Step 5: Keep sensitive data out of chat

Tell the AI, and your team, never to ask for card numbers, passwords, ID documents or health details in chat. Add a line to the greeting or pre-chat form if customers often volunteer them. If something sensitive arrives anyway, have a routine for redacting or deleting it.

Step 6: Keep a person on decisions that matter

An AI agent can explain your refund policy. It shouldn't be the only thing standing between a customer and a refund outside policy, an account closure or a fraud flag. Route those to a person, and make "talk to a person" available in every conversation. This is good GDPR practice and good service.

Step 7: Update your notice and brief your team

Add the chatbot, the vendor and the retention period to your privacy notice. Then spend twenty minutes with everyone who manages the AI: what it may answer, what it must hand over, and how to handle a request to delete someone's data. That briefing is also the simplest way to show you support staff AI literacy.

Common mistakes

• Assuming the vendor handles compliance. The vendor provides tools; your greeting, settings and retention choices decide the result.
• Disclosing once on a separate page. The disclosure belongs in the conversation, at the start.
• Forgetting WhatsApp and email. The rule applies to every channel where AI replies.
• Keeping everything "just in case". Unlimited retention is a GDPR problem, not a safety net.
• Letting the bot say no to money. Refund and account decisions need a person.

Your one-afternoon compliance check

1. Open your chatbot as a customer on every channel. Does the first message say it's AI? Can you reach a person?
2. Find your DPA for each AI vendor, and the list of sub-processors.
3. Ask your vendor where data is stored and whether conversations train its models.
4. Set a retention period and switch on automatic deletion.
5. Update the privacy notice and brief the team.

If you're still choosing a tool, our guide on how to choose an AI agent for customer service covers what to test before you buy. For the bigger picture of what to automate, see how to automate customer support.