AI Act chatbot disclosure: how to set up the AI notice

Updated on 8 October 2026: we have added the transition period for Article 50(2).

Why we are looking at the notice in the chat right now

Customers should know whether they are writing to a machine. That is our position, regardless of any deadline.

There is still a reason to look now. According to the European Commission, the transparency rules in Article 50 of the EU AI Act have applied since 2 August 2026. You can read the exact wording in the text of the regulation on EUR-Lex.

Many companies don't have this on their radar. The debate of recent months was almost entirely about the postponement of the high-risk rules by the so-called AI Omnibus, which according to the Commission entered into force on 27 July 2026 (Official Journal of the EU). That has left many people with the impression that the whole AI Act has been postponed. For Article 50, the Commission's questions and answers still give 2 August 2026. It names one exception there: for AI systems placed on the market before 2 August 2026, providers only have to meet the obligation to mark AI-generated content (Article 50(2)) from 2 December 2026.

If someone in your company says the AI Act has been postponed, ask which part they mean.

For the service team, the deadline isn't the interesting part anyway. What matters is whether a customer understands from the first sentence who they are talking to, and how to reach a human. This article describes how a service team sets that up.

What a good AI notice does from the customer's point of view

A notice that nobody reads helps nobody. That is why we test AI notices in customer contact against five questions:

Does it come early? The notice belongs in the first message or the first seconds of a phone call. It does not belong on the legal notice page or behind an info icon.
Is it in the conversation itself? A banner above the chat window is easy to miss. Everyone reads the greeting.
Does it reach everyone? A screen reader has to read it out. If the customer writes in English, the notice should be in English too.
Is it worded honestly? To many customers, a first name with a profile photo comes across as a real person. The word "AI" belongs in the text.
Does it show a way out? The customer should know how and when they can reach a human.

We consider the last question the most important. Frustration in a chat rarely comes from a machine answering. It starts when customers realise after three detours that they are stuck and see no way out.

The AI notice does not replace the data protection notice for the chat. Both should sit side by side.

The legal requirements for content, timing and form are explained in the Commission's guidelines on the transparency obligations of 20 July 2026. For how this applies in your own company, your legal counsel is the right point of contact.

Who is responsible in your company?

The AI Act distinguishes between providers and deployers of AI systems. The Commission describes both roles in its questions and answers on Article 50. Which role your company has for which tool is a legal question. We don't settle it ourselves; we have it reviewed. You can create an overview of your tools with the deployer check for your AI inventory.

In practice, we think another question matters more: who in your company makes sure the notice reaches the customer?

Our recommendation: make sure the notice appears in your customer contact, whoever owes it legally. It costs little. And it catches a typical mistake: the vendor ships a standard notice, your own greeting message overwrites it, and nobody notices.

There are three scenarios you should still pass to your legal counsel separately:

your own solution based on a language model (the technology behind chat assistants), even if a service provider built it,
a purchased tool that the customer only sees under your logo and with your own assistant name,
tools that do more than give information, for example analysing callers' mood or preparing decisions.

The 7-step check for your AI customer service

The checklist is built for one working session with service, IT and purchasing. Each step says what you check, who takes it on and how you can tell it is done.

1. Inventory: where does AI talk to your customers?

What to check: every channel in which an AI system communicates directly with customers. Typical examples are website chat, WhatsApp and other messengers, a voicebot on the hotline and automatically generated email replies. Ask explicitly about pilot projects and tools that departments have introduced themselves. These are the ones most often missing from central overviews.

Who: the head of customer service, with IT. Purchasing supplies the list of software contracts as a cross-check.

Done when: there is a table with channel, tool, vendor, internal contact and start date.

2. Collect borderline cases and send them for review

What to check: in the inventory, mark every system that falls under one of the three scenarios from the previous section: in-house build, third-party tool under your own brand, tool that does more than give information. This list goes to your legal counsel.

Who: management or the head of customer service.

Done when: every borderline case is marked "review commissioned" with a date.

3. Test the notice at first contact

What to check: open each channel from the customer's point of view. Is the notice in the first message, or hidden behind an info icon? Does a screen reader read it out? Is it written in the language the customer writes in?

Who: the head of customer service, with the person responsible for the website.

Done when: for each channel there is a screenshot of the first contact with the notice visible, and the screen reader test is documented.

4. Check voicebots separately

What to check: on the phone, there is no chat window for context, and a good synthetic voice sounds like a human. That is why we recommend an announcement in the first seconds, before the customer describes their request.

Who: the person responsible for the phone system and voicebot, with the head of customer service.

Done when: an internal test call (only staff who are in the loop, no real customer calls) proves the announcement at the start of the call, for example via the voicebot's call log, and the announcement is documented as a fixed part of the call flow.

5. Record AI-generated content

What to check: where do you use AI images, AI audio or AI texts in customer communication, and who approves them? Ask the vendors whether and how their tools label generated content.

Who: marketing and communications, because that is where most published content is produced.

Done when: there is a list of tools for images, audio and text, each with the person responsible for approval and the vendor's answer on labelling.

6. Ask vendors in writing

What to check: ask every tool vendor: how does your product meet Article 50? Is the AI notice active by default, or does it have to be configured? Can it be switched off by accident? How is generated content labelled? Check whether your contract covers these points. "We are AI Act compliant" without details is no answer.

Who: purchasing or the person who manages the contract, with input from service and IT.

Done when: there is a written answer for every tool, and gaps are noted as tasks with a date.

7. Assign responsibility, file the evidence, brief the team

What to check: name one person for AI transparency in customer service. They keep the inventory up to date, check new tools before launch and file the evidence from steps 1 to 6 in one central place. Anyone who configures the chatbot or handles complaints about it should know what the notice is supposed to do and how to tell when it is missing. What belongs in that file is covered in the AI documentation with three documents.

Who: management names the person. That person organises the filing and a short briefing for the teams involved.

Done when: the responsibility is recorded in writing, all evidence is in one place and the next review date is in the calendar. Repeat the check as soon as a new AI tool goes into customer contact.

Three text templates for chat and phone

You decide how the notice sounds. This is where it is decided whether customers skim past it as mandatory text or experience it as good service. A good notice answers two questions: who am I talking to? And how do I reach a human?

Website chat, short and with a way out

"Hello, I'm the digital assistant of [Company], an AI. I can help with questions about orders, delivery and invoices. If you would rather talk to a person, just type 'human'."

The customer knows straight away where they stand, what the assistant can do and how to get past it. A made-up name that pretends to be a human is deliberately left out.

Service with sensitive topics, with a clear limit

"You are writing to an AI assistant. I answer general questions directly. For cancellations, complaints or contract changes, I will put you through to our service team, on weekdays from [time] to [time]."

You decide in advance where the AI stops. That protects customers in delicate moments, and it protects your team from answers nobody has approved.

Phone voicebot

"Hello, you are speaking to the AI voice assistant of [Company]. To be put through to a member of staff, say 'human' or press zero."

The notice comes at the start, spoken, without a detour through a menu. The word AI is what matters. "Automated" on its own sounds like a classic phone menu.

Approval and upkeep: a single standard instead of a patchwork

In many companies, AI in customer contact grew without a plan. A chat widget here, an automatic email reply there, a voicebot from a pilot project. Each system has a different notice, or none at all.

We suggest setting the standard once and then only maintaining it:

One text template per channel. The head of customer service writes it, and the AI transparency owner from step 7 manages the current version. Changes to the greeting go through them.
One fixed route to a human. With clear service hours that are stated in the notice and are accurate.
One approval rule. A new AI channel only goes live once the notice and the handover to a human have been tested from the customer's point of view and documented. That applies to pilot projects too.
One regular review. How often, and at which point, do customers choose the route to a human? That shows you where the AI is out of its depth.

That way the notice stays in place when the next new tool arrives. And you get an instrument for the question that matters beyond any deadline: where does AI help your customers, and where does it cost you trust?

Next step: a first assessment of where you stand

For a first overview you don't need a project. Sit down with the person responsible for customer service and the person who knows the tools. Go through the seven steps and mark each one "done", "open" or "unclear". While you do, open your own chat and call your hotline. Do you learn at first contact that you are talking to an AI? For each open point, note a name and a date. The evidence from the check (screenshots, internal test calls, vendor answers) follows afterwards. How to steer a bot beyond the question of the notice is covered in the containment rate as a governance trap.

The points marked "unclear" are usually the most important. The legal ones go to your legal counsel, the organisational ones to the person from step 7.

If you would like a second opinion: we are happy to go through your results with you in a short conversation and discuss where the notice is missing or unclear, and which questions you should put to your tool vendor. More at /en/ai-diagnose.

Frequently asked questions

How do I test my chat from the customer's point of view? Open each channel in a private browser window or on your own phone, without logging in. Read only the first message. Does it say that an AI is answering, and how you can reach a human? Then switch on a screen reader and repeat the test. For the hotline, an internal test call is enough.

What do I ask my tool vendor? Whether the AI notice is active by default, whether your own greeting texts can overwrite it, whether it can be switched off by accident and how generated content is labelled. Get the answers in writing and check whether your contract covers these points.

Where do I get a binding answer? Not from us. According to the KI-Marktüberwachungs- und Innovationsförderungsgesetz (KI-MIG) (German act on AI market surveillance and innovation promotion, in German), the competent supervisory authority in Germany is the Bundesnetzagentur (Federal Network Agency), unless a sector-specific supervisory authority is responsible. For what applies in your case, your legal counsel is the right point of contact.

This article describes our practical perspective. It does not contain a legal assessment; for that, we refer you to the linked primary sources and your legal counsel.

Does your service chatbot say it is an AI? How to set up the notice in customer service