An AI chatbot needs a first-message acceptance test

A customer should not have to interrogate a chatbot to discover that nobody is typing on the other side. The notice belongs before trust has formed, not after the bot has collected a name, interpreted a complaint or proposed a next step.

On 20 July 2026, the European Commission published its final guidance on Art. 50 of the AI Act. For systems that interact directly with people, the practical instruction is unusually testable: inform the person from the start of the first interaction, clearly and distinguishably, unless it is obvious that the system is AI. The obligation applies from 2 August 2026.

For an SME, that converts a legal paragraph into a release check. Open every public entry point, start a clean session and record what the user learns before the first substantive exchange.

Four conditions define the interaction

The Commission’s questions and answers set out four cumulative conditions for the chatbot notice. The product must qualify as an AI system. It must support a genuine two-way exchange rather than only collect information or issue fixed automated responses. The AI must communicate directly with the person, not through a human intermediary. And the other party must be a natural person.

That scope excludes some systems commonly called bots. A background classifier that routes support tickets does not communicate directly with the customer. Machine-to-machine traffic is outside this interaction rule. A conventional web form does not become a chatbot because it sends a fixed confirmation.

The opposite mistake is more dangerous: treating a fluent assistant as merely a search box. If a customer can ask follow-up questions and the system generates contextual replies, test it as a direct AI interaction. The product name, channel or supplier category does not decide the point; the exchange does.

Acceptance question	Pass condition	Evidence to retain
Does the system hold a genuine two-way exchange?	The inventory identifies every conversational entry point	URL, channel and product owner
Is the person told from the start?	The notice appears before or with the first AI message	Dated screenshot and transcript
Is the notice clear and distinguishable?	It says plainly that the user is interacting with AI	Approved wording and rendered view
Is the notice accessible?	It remains perceivable and operable in the deployed interface	Keyboard and assistive-technology check
Does a new session behave the same way?	The notice does not depend on old cookies or chat history	Clean-session test result

Source: European Commission final Article 50 guidelines and questions and answers. Last verified 2026-07-22.

The first message is the test boundary

The final guidance removes a tempting ambiguity. A disclosure hidden behind an information icon, placed only in terms, or revealed after several turns is not a reliable way to inform someone from the start of the first interaction.

The smallest durable pattern is a short statement next to the launcher and again in the opening message: this is an AI assistant. Product teams can add useful context — what it can do, what it cannot decide, where a person can take over — but that material should not obscure the basic fact.

Run the test on the deployed journey, not only in a vendor demonstration. A notice configured in the supplier dashboard can disappear when the widget is embedded, when a mobile layout truncates the opening message, or when a messaging channel replaces the welcome screen. A clean browser session catches the first failure. Testing the narrow viewport and the keyboard path catches two more.

Accessibility is part of Art. 50(5), not a decorative improvement. The Commission says the notice must be clear, distinguishable and provided in accordance with applicable accessibility requirements. In practice, do not encode the disclosure only in colour, an unlabeled icon or an image that assistive technology cannot identify. The acceptance evidence should show what a person can perceive, not merely what the configuration field contains.

Do not build around the “obvious” exception

Article 50 does not require a notice when it is obvious to a reasonably well-informed, observant and circumspect person that they are interacting with AI. The Commission says this exception should be interpreted restrictively because it removes transparency from the user.

That makes “the bot sounds robotic” a poor acceptance criterion. Interfaces change, model responses become more natural and customers arrive with different expectations. A clear sentence is cheaper to maintain than an argument about whether a mascot, sparkle icon or product name made the technology self-evident.

There may be narrow cases where the context genuinely makes the interaction unmistakable. Record that reasoning if the business relies on it. For an ordinary support, booking, sales or internal-service assistant, the lower-risk choice is to state the fact.

Provider duty, buyer test

Art. 50(1) places the design and development duty on the provider of the interactive AI system. If an SME buys a standard chatbot, it should not attempt to recreate the provider’s compliance architecture. It should ask the supplier to identify the control and then verify that the control survives deployment.

The distinction becomes less comfortable when the buyer changes the system. A company that places an AI system on the market or puts it into service under its own name or trademark can fall within the Act’s provider definition. Rebranding, substantial modification and bespoke development need a role analysis rather than an assumption that the software vendor owns every duty.

The acceptance test still works in either case. It does not decide legal status; it proves what the person sees. Keep the following in the release record:

• the system and business owner;
• the channels and URLs where interaction starts;
• the exact notice wording;
• one dated screenshot and transcript from a clean session;
• the accessibility check and any defect owner;
• the supplier document or contract term describing the Article 50 control.

This is a compact evidence file, not a new compliance platform. Repeat it when the launcher, welcome flow, channel integration or provider changes.

The August date has no chatbot grace period

The Commission confirms that Article 50 applies from 2 August 2026. Its questions and answers describe a limited transition until 2 December 2026 for the machine-readable marking obligation in Art. 50(2) for generative systems already placed on the market before August. That is a provider-side content-marking issue. It does not postpone the interaction notice in Art. 50(1).

The useful deadline is therefore the next release window. List the customer-facing assistants, test the first clean interaction, fix any missing or inaccessible notice and retain the evidence with the release. If a chatbot cannot state what it is before doing its job, it is not ready to greet a customer.