AI Chatbot Disclosure Template (EU AI Act Article 50(1))

Document 05 of the RegShelf kit gives you everything between "we should probably label the chatbot" and a documented, defensible implementation: a decision tree to confirm whether the duty applies to each system, ready-to-paste disclosure wording in three tones (neutral, friendly, formal) for four channels (chat first message, persistent UI label, voice opening, AI-sent email and ticket replies), placement rules drawn from the Commission's July 2026 Guidelines, approved short-form snippets in five EU languages, an implementation checklist and an evidence log. All three tones meet the legal standard — the choice is brand voice, not compliance level.

Companies that provide user-facing AI: a support chat widget, a voice bot on the phone line, an AI avatar, or automated email and ticket replies that hold a genuine two-way conversation. The kit's worked examples show where the line runs: a fluent web chatbot needs disclosure (modern bots are easily mistaken for humans); a voice bot needs it in the opening sentence; a feature clearly branded "Ask AI" may be obvious, but a label is still recommended; and AI-drafted replies that a staff member reviews, edits and sends under their own name typically do not trigger the duty at all — the customer is interacting with your colleague, not the AI.

Article 50(1) sits on the provider — it is a design duty. The key parameters, per the Act and the Commission's Article 50 Guidelines of July 2026:

• Scope: AI systems intended to interact directly with natural persons through a genuine bidirectional, conversational exchange. One-way automated messages without a real exchange are not automatically covered.
• Timing: disclosure at the latest at the first interaction — a first bot message that discloses is fine; a line buried in terms of service or an FAQ is not.
• Exceptions: where AI is obvious to a "reasonably well-informed, observant and circumspect" person judged from the circumstances (not from what a tech-savvy founder would notice), and a narrow law-enforcement exemption that almost never applies to an SME product.
• Not required: a persistent label kept visible throughout long conversations. That is recommended practice — the template says so explicitly — beyond the strict legal duty.

Non-compliance falls in the fines framework of up to €15 million or 3% of worldwide annual turnover; for SMEs and small mid-caps the lower of the two applies (Article 99(6)).

Six sections:

• Does the duty apply? — a three-question decision tree plus a table of five worked situations and a border-case rule (if you are debating whether it is obvious, disclose);
• Ready-to-paste copy — twelve disclosure texts covering chat first messages, persistent UI labels, voice openings and email/ticket lines, each in neutral, friendly and formal variants;
• Placement guidance — timing, visibility, voice channels, accessibility (screen-reader equivalents, spoken disclosure), language matching and human-handover rules;
• Multilingual snippets — approved wording in English, German, French, Spanish and Italian for all four placements;
• A ten-step implementation checklist and an evidence log — one row per system and channel, with dated screenshots or recordings, so you can show a market surveillance authority exactly what was live when.

List every channel where people interact with your AI, run each through the decision tree, pick one tone, and paste the wording into the first bot message, voice opening and AI-sent replies. Localise into every interface language, verify accessibility, then screenshot each disclosure and file it in the evidence log with a six-month re-check reminder. If your AI also generates published content, pair this kit with the content marking SOP — disclosure and marking are separate Article 50 duties.