AI chatbot disclosure: what your business should check

Your website chat answers questions about services and appointments. Visitors should know from the outset whether a person or AI is replying. The transparency duty under Article 50 of the AI Act has applied since 2 August 2026. A short, visible notice inside the open chat is a sensible first step. You also need to establish whether your business is itself the provider of the system.

We would start with the actual conversation: what does a visitor see before typing a question? What does a caller hear? And is it clear who is replying when a member of staff takes over? Those questions lead to concrete requirements for the tool and its integration.

When does a chatbot need an AI notice?

Article 50(1) addresses providers of AI systems intended to interact directly with people. The European Commission’s FAQ identifies four cumulative criteria: an AI system, a genuine two-way exchange, direct communication by the AI, and natural persons on the receiving end. A conversation about delivery terms can qualify even when visitors are acting professionally.

A fixed email acknowledgement does not become such a conversation simply because it is sent automatically. If AI only prepares a draft which a member of staff checks and sends, it is not communicating directly with the customer either. An email or messaging assistant that independently answers follow-up questions needs a closer look. Assess the exchange itself, rather than the channel’s name.

Read the exception for obviously recognisable AI narrowly. A first name, portrait and personal greeting can suggest a human contact. Our recommendation is to say “AI assistant” explicitly, instead of expecting visitors to work out what is behind the conversation.

Put the notice at the start of the conversation

Paragraph 5 requires clear, distinguishable information no later than the first interaction. For a website chat, we therefore recommend a readable notice inside the open window before visitors send their first question. A reference in the legal notice does little to help at that moment. The law does not prescribe a particular sample sentence.

The last sentence only fits if that contact option exists. If your chat can hand the conversation directly to a member of staff, describe that option instead. Make it visible when a person takes over. These are our recommendations for a clear experience, rather than mandatory wording.

For a telephone assistant, include the information in the spoken greeting before the substantive conversation starts. For example: “Hello, this is Example Company’s AI telephone assistant.” A small visual badge alone may be inadequate for a speaking avatar. Check that the notice actually reaches people in the channel concerned.

Check accessibility in the real widget

Article 50(5) refers to applicable accessibility requirements. In practice, we would present the notice as real text, make it easy to read, and test it with a screen reader and keyboard. An “AI” label embedded in an image, or colour alone, does not communicate the information reliably.

Open the widget on a narrow screen and with enlarged text. Is the notice still visible? Is it read out when the chat opens? Can the input field still be reached? Record the date and widget version you tested. Someone in the business should repeat this short check after a supplier update.

Who is responsible when you use a ready-made widget?

The design duty in paragraph 1 rests with the provider. Under Article 3(3), this can include a business that has a system developed and puts it into service under its own name or trademark. Custom development and white-label arrangements therefore deserve a careful assessment of roles. Using another company’s underlying model does not settle who provides the finished application.

For an unchanged third-party product, ask who provides the interactive system, how the notice is implemented, and which settings can alter it. Then check your public integration. A notice in the product demo does not establish that it still appears on a phone after your customisation.

Our practical recommendation for procurement is to document the roles, wording, placement and acceptance test. The legal provider and deployer roles depend on the system and the actual circumstances. Unclear contractual or white-label arrangements need legal assessment.

What the notice does not resolve

Article 26 concerns deployers of high-risk AI systems. It is not a universal checklist for every service chat. Whether a system falls into that category requires a separate assessment. Article 50(6) preserves other duties; an AI notice does not answer questions about processing purposes, legal grounds or privacy information. The GDPR continues to apply alongside it.

If a system uses emotion recognition or biometric categorisation, paragraph 3 also needs attention. A notice does not make an otherwise prohibited use permissible. Disclosure during a conversation is also distinct from machine-readable marking of generated content. The Commission’s December 2026 grace period concerns only the marking duty under paragraph 2 for systems placed on the market before 2 August. It does not postpone the chatbot notice.

Infringements of Article 50 fall under Article 99(4). Its ceilings are €15 million or 3% of the preceding financial year’s worldwide annual turnover; the higher amount generally applies. For SMEs, paragraph 6 uses the lower of the two ceilings. This is not a fixed fine for a missing sentence: the circumstances determine whether a penalty is imposed and its amount.

How to check the integration before sign-off

1. Walk through the conversation: cover the website, telephone and independently replying messaging channels; distinguish tools that only prepare drafts.
2. Establish the roles: identify the provider of the finished system and those responsible for integration, particularly under your own brand.
3. Test the first contact: check the notice before the substantive conversation, clear wording and a human contact option that actually works.
4. Record the tested state: check mobile display, reading order, keyboard access and handover; name someone responsible for later changes.

Published images, text and deepfakes have their own rules. We cover those in our guide to labelling AI-generated content. Our AI Act overview for smaller businesses explains roles and risk categories.

If you are planning a chatbot for customer communication, design its knowledge sources, response boundaries and handover to your team alongside the notice. As your 360° partner, we connect that planning with technical integration and rollout. This does not replace legal assessment of your specific roles.