Labelling AI in support: what Article 50 requires
Article 50 of the AI Act requires clear notices whenever artificial intelligence is used, from August 2026. Here is how to implement the labelling duty.
Martin Semmele
Key insights
- From 2 August 2026, the use of artificial intelligence must be recognisable at the first interaction.
- The duty falls on the operator of the site, not on the provider of the tool running behind it.
- A notice in the legal notice page or the privacy policy does not meet the legal requirements.
- Removing vendor branding often deletes the legally required label along with it.
From August 2026: the duty to label
Customer support with artificial intelligence is quickly set up, but from 2 August 2026 it carries a clear statutory duty. Article 50 (1) of Regulation (EU) 2024/1689 (the European AI Act) requires that systems intended to interact directly with natural persons be designed so that the persons concerned are informed that they are interacting with an AI system1. This article sets out what that means for day-to-day support and is not legal advice.
For support leads and anyone running a chat window, the timetable is unambiguous. The duty to label openly in the chat window applies from the cut-off date immediately and with no grace period, for systems already running as well as new ones. Anyone operating an automated dialogue system has to let customers know straight away where they stand.
The cut-off dates at a glance
| Area | Cut-off date | Legal basis | Requirement in support |
|---|---|---|---|
| Transparency notice in the chat window | 2 August 2026 | Article 50 (1) | Visible notice before or at the start of the conversation |
| Machine-readable marking of older systems | 2 December 2026 | Article 50 (2) and Article 111 | Metadata and marking for pre-existing systems |
While technical metadata and machine-readable marking for systems placed on the market before 2 August 2026 get an extension until 2 December 2026, there is no delay for interaction applications under Article 50 (1)2. Whoever fails to display the notice on the cut-off date is in direct breach of the regulation.
This distinction matters, because the debate around the Digital Omnibus package of November 2025 was often described as a postponement of the AI Act. What it concerned were the obligations for high-risk systems. The transparency duty under Article 50 is not among them: it remains anchored to 2 August 2026. So if you have read the news about a postponement, do not apply it to your chat window.
Responsibility: who has to display the notice
Many teams wrongly assume that complying with the transparency duties is solely the job of the vendor whose software they use. The regulation separates the roles precisely, though: providers develop the system or place it on the market under their own name, while deployers use it on their own responsibility in day-to-day business2.
The moment you embed a chat window on your website or in your web application, you are acting as the deployer. Responsibility for correct configuration, for the greeting texts on display and for the visible label rests entirely with your company. The tool vendor merely supplies the technical infrastructure.
- Deployer (your company): responsible for placing the notice, wording it comprehensibly and displaying it correctly in live operation.
- Tool vendor (software manufacturer): responsible for supplying the technical interfaces, configuration options and basic functions.
- Support lead: responsible for checking regularly whether your own text changes have overwritten the legal notice in the widget.
If your team adjusts settings in the admin area, swaps text blocks or changes design specifications, you have to make sure the prescribed information stays visible. A failure here cannot be passed on to the software vendor.
The right place: right at the first interaction
Article 50 (5) of the AI Act provides that the information must be given in a clear and distinguishable manner at the latest at the time of the first interaction1. For the interface, that means: the notice belongs exactly where the conversation begins.
Opening view and role label
An effective notice has to appear in the visible area of the window before the person seeking help sends their message. Ideally the interface combines two elements: a short notice in the header or role label, plus a clear greeting message in the conversation itself.
- Role label in the header: name the function transparently (for instance as virtual assistant or automated help).
- Opening screen in the chat: place an introductory sentence directly above the input field or as the first system message.
- Input area: keep the placeholder text in the input field neutral and factual.
Why the legal notice page and privacy policy are not enough
A reference buried deep in the privacy policy or the legal notice page does not meet the statutory criteria: the notice is meant to appear right at the start of the communication and not only in the terms and conditions or the privacy policy2. Customers cannot be expected to search through legal documents before asking a simple question. The notice has to sit in the immediate field of view of the interaction. The interface must also meet applicable accessibility standards, so that screen readers can pick the label up.
Plain words instead of foreign-language flourishes
Transparency only works with comprehensible language. The regulation requires the information to be provided in a clear and distinguishable manner and to meet the applicable accessibility requirements1. Vague paraphrases or convoluted nested sentences miss that purpose.
Many interfaces carry, by default, English references to the software vendor. Those short lines often read like advertising and rarely explain comprehensibly that the answer is machine-generated. In customer service for a given language, use plain, factual wording in that language throughout.
| Area of the window | Unclear or foreign-language phrase | Clear, plain wording |
|---|---|---|
| Header | Vendor name Support Bot | Automated support help |
| Greeting | I am your personal all-rounder. | Hello. This is an AI-assisted assistant that answers questions from our help documents. |
| Notice before input | English reference to the underlying system | Automated answer. For complex matters our team takes over. |
Avoid exaggerated promises about how human the machine is. Being factual creates more realistic expectations and prevents frustration among customers.
Handover: when a human takes over the conversation
A well-configured support process answers routine questions automatically and passes open cases to the team. The moment a member of staff steps in, the character of the communication changes fundamentally.
The transparency duty requires customers to know at all times who they are talking to. Where a switch takes place, it has to be documented unmistakably in the message history. Getting the handover moment right protects the customer's trust.
- 01Status change in the system: the system signals in the history that automated handling has ended and the matter has been passed on.
- 02Name and role display: in the header, the display switches from the virtual assistant to the real name of the team member.
- 03Personal greeting: the human opens the direct exchange with a short message confirming the transition.
If this transition is obscured or shown unclearly, users do not know whether they are still interacting with a machine or already with a person. Clear status messages provide immediate orientation here.
The trap: your own texts delete the notice
In day-to-day practice the biggest compliance risk rarely arises from bad intent, but from simple mishandling in the admin area of the support software.
Many systems ship pre-configured notice texts that point out the machine assistance by default. When a team adapts the greeting message to its own brand voice or removes visible vendor logos, the standard message is often deleted entirely without anyone noticing. What remains is a neutral text that conceals the machine character of the answer.
- Check the text after every change: after each adjustment in the dashboard, check the live view of your chat window.
- Fixed components in the greeting text: word your own greetings so that the reference to the AI system is a permanent part of the text.
- Four-eyes principle: have configuration changes to widgets and contact forms briefly reviewed before they go live.
Anyone maintaining their own texts has to build the transparency label into those templates actively. Companies should therefore check whether notices actually appear visibly, comprehensibly and in good time in the specific application2. Adapting the visuals to your own corporate design does not release you from the duty under Article 50.
Safe operation with the right platform
A clean implementation calls for software that permits flexible text changes while supporting a clear separation of roles. Good tools keep greeting texts, agent roles and team members in the inbox cleanly apart.
In Comlayer the two kinds of text are separate, and the difference decides the labelling. The role label of the answering side (KI-Assistentin) comes from the widget texts that ship with the product and cannot be overwritten in the dashboard; it does not disappear even when the visible vendor mark is removed on the Pro plan. What is freely editable is the greeting, the headline and the subtitle — and that is exactly where the default wording speaks of the team, not of a machine. So if you rewrite those fields, keep the reference to the automated answer in them deliberately.
As soon as a question gets more complex, the handover to the shared inbox takes effect. The line under the title in the widget switches from the role label of the AI assistant to a note that the team has been informed. The change is therefore visible without asserting a presence that nobody has promised yet.
Frequently asked questions
Is there a transition period for disclosure in the chat window?
No, the duty to be recognisable at the first interaction applies from 2 August 2026 with no transition period. Only the machine-readable marking for existing systems may be retrofitted until 2 December 2026.
Is a general notice in the privacy policy enough?
No. Users have to recognise immediately, at the first interaction, that they are communicating with a machine. A hidden notice in the legal notice page or the privacy terms does not meet this requirement.
Who is responsible for correct implementation?
The duty falls on the operator of the site. If you embed an automated assistant on your web presence, you have to make sure the notices are displayed correctly and clearly visibly.
What happens when the assistant hands over to a human?
As soon as a human member of staff takes over the conversation, that switch has to be immediately apparent to the user. In practice this happens through a clearly changed status or a new name in the history.
Which wordings are suitable for the notice?
Use clear, factual terms in the language of your customer service, such as automated answer or artificial intelligence. Avoid foreign-language marketing flourishes, since they obscure the point unnecessarily for many users.