AI Act: what your e-commerce customer service AI must disclose since August 2026

Since August 2, 2026, an e-commerce customer service AI that interacts directly with your customers must tell them they're talking to an AI, clearly and visibly, at the latest in the first message. This obligation comes from Article 50 of the EU Artificial Intelligence Act (AI Act): it covers the chatbot on your website, emails sent automatically by an AI, and voice agents.
This page explains what applies in practice, which opening messages to use, what the law says (and doesn't say) about handoff to a human, and the upcoming deadlines. It is not legal advice: for a specific case, have your setup reviewed by a lawyer.

The essentials
A chatbot or AI agent that replies to your customers on its own must identify itself as an AI from the very first exchange. A mention in your terms and conditions, a human first name or a plain “Assistant” label is not enough. The postponement voted in 2026 (the “Digital Omnibus”) pushed back the rules on high-risk AI systems, but not this transparency obligation. Fines can reach €15 million or 3% of worldwide turnover, with the lower amount applying to SMEs.
What has the AI Act changed for e-commerce customer service AI since August 2, 2026?
The EU AI Act (Regulation 2024/1689) classifies AI systems by risk level. A customer service chatbot generally falls under “limited risk”: it isn't prohibited and doesn't carry the heavy requirements of high-risk AI systems, but it must be transparent.
In practice, Article 50 requires that people be informed they are interacting with an AI system, unless this is obvious. The information must be provided “in a clear and distinguishable manner”, at the latest at the time of the first interaction. This rule has applied since August 2, 2026.
On July 20, 2026, the European Commission published guidelines on these transparency obligations. They take a narrow reading of the “it's obvious” exception: a helpdesk chatbot that a customer could mistake for a human must disclose that it's an AI.
Which customer service AI systems are covered?
According to the Commission's guidelines, the obligation targets direct, two-way interactions between an AI and a person. For e-commerce customer service, that means:
- Covered: the chatbot on your website or app, messaging channels (WhatsApp, Instagram) handled by an AI, emails written and sent by an AI without review, a voice agent that answers the phone.
- Likely not covered: fixed-rule automatic replies (acknowledgment of receipt, out-of-office message), which aren't AI; internal tools used by your agents.
- The copilot case: when the AI writes a draft that an agent reviews, edits if needed and sends, the exchange is no longer direct between the AI and the customer. That's the reading of several analyses published since the guidelines, with one caveat: a purely formal review, where the agent clicks without reading, isn't enough.
If you're weighing automation against agent assistance, our guide to the customer service copilot covers both modes in detail.

When and how should you tell customers they're talking to an AI?
At the latest in the first message, and for every new customer, not once and for all somewhere on the website. The disclosure must be visible where the customer reads the reply: in the chat's opening bubble, at the top or in the signature of the email, at the start of the call.
What isn't enough, according to the guidelines and the analyses published since:
- a mention in the terms and conditions or the privacy policy;
- a generic name such as “Assistant” or “Advisor”;
- a first name and a human avatar photo;
- a general statement such as “our website uses AI”.
For sensitive topics (complaints, disputes, payment, health), the Commission expects extra care, with reminders throughout the conversation. The same applies if your customers include children or older people.

Which opening messages should e-commerce customer service use?
A good message says three things in one sentence: who is speaking (an AI), on whose behalf (your brand) and how to reach a human. Here are examples to adapt to your tone of voice.
| Channel | Sample message | Avoid |
|---|---|---|
| Website chat | “Hi, I'm the AI assistant for [Brand]. I can track your order or start a return. Type ‘agent’ to talk to someone on our team.” | “Hi, it's Lea!” with a photo |
| Automatically sent email | In the signature: “Reply written by the [Brand] customer service AI assistant. Reply to this email to be contacted by an agent.” | A “Thomas, Customer Service” signature with no other mention |
| Voice agent | “Hello, you're speaking with the [Brand] AI voice assistant. Say ‘agent’ at any time to be transferred.” | A voice that introduces itself with a first name, without saying it's an AI |
| Complaint or dispute | “I'm an AI: your refund request will be approved by a member of our team within 48 hours.” | A refusal sent by the AI with no disclosure and no recourse |
For email, a disclosure in the signature is the easiest to roll out everywhere. Also put it in the first paragraph if the reply deals with a sensitive topic.
Do you need to provide a handoff to a human?
Article 50 of the AI Act does not, on its own, require a handoff to a human for a limited-risk customer service chatbot. It requires disclosure. But there are two reasons to provide a handoff anyway.
First, the GDPR: its Article 22 regulates decisions with significant effects made solely by a machine. A refusal of a refund or goodwill gesture decided by the AI alone falls into that zone. Second, trust: a customer who knows they're talking to an AI wants to know how to get out of it. Say so from the first message, and make sure the agent picks up with the full history, without making the customer repeat themselves.
Our article on support escalation details the rules for handing over.
Who is responsible: the AI vendor or your brand?
Article 50(1) is aimed first at the provider of the AI system, meaning the vendor: it must design its tool so that the disclosure is made. Your brand, which deploys the tool with its customers, chooses in practice the wording of the opening message, the email signature and the displayed name.
In practice, then, check it on both sides: the vendor provides the disclosure and the settings, and your team makes sure they're active on every channel. Ask your vendor, in writing, how its tool discloses the AI on chat, email and voice.
What are the next AI Act deadlines?
The “Digital Omnibus” regulation on AI (Regulation 2026/1744), published in the Official Journal on July 24, 2026, pushed back several dates. It did not change the obligation to tell customers they're talking to an AI.
| Date | What applies | Impact for e-commerce customer service |
|---|---|---|
| February 2, 2025 | Prohibited practices and the AI literacy obligation (Article 4) | Train the teams that use AI |
| August 2, 2026 | Transparency obligations under Article 50 | AI disclosure on chat, automated email and voice |
| December 2, 2026 | Machine-readable marking of generated content, for systems already on the market | Mainly for vendors of generative AI for images, audio or text |
| December 2, 2027 | Rules for high-risk AI systems (Annex III), postponed by the Omnibus | In principle out of scope for standard customer service |
| August 2, 2028 | High-risk AI systems linked to products (Annex I) | Out of scope for standard customer service |
The Omnibus also rewrote Article 4: companies must take measures to build their teams' AI literacy, without an obligation to guarantee a specific level for each person.
What are the penalties for non-compliance?
Breaching the transparency obligations can cost up to €15 million or 3% of total worldwide annual turnover, whichever is higher (Article 99). For SMEs and startups, the lower amount applies. Enforcement is the responsibility of national market surveillance authorities.
How does Klark apply AI transparency?
Our position is simple, and we summed it up this way in a conversation with a brand in September 2026: “It's better to say an AI wrote it when it's an AI that writes and sends the reply to a customer.”
AI Act questions also come in through procurement teams. In 2026, a prospect asked us for a formal risk analysis of our AI, covering model governance and what happens to data sent to third-party models. Prepare these questions for your own vendor: they overlap with GDPR questions, covered in our guide AI in customer service and GDPR and on our security and GDPR page.
The AI Act checklist for your customer service AI
- List the channels where an AI replies on its own: chat, email, messaging, phone.
- Add the AI disclosure to the first message on each channel, with your brand name.
- Remove human first names and photos from the AI's avatars.
- Explain how to reach an agent right from the opening message.
- Keep a human involved in refund refusals, exceptions and disputes.
- Check reviewed drafts: if your agents approve without reading, treat those replies as automated.
- Ask your vendor for its Article 50 documentation and train your teams (Article 4).
To choose a tool, see also our guide AI for customer service: how to choose a solution that works in production.
Frequently asked questions
Does my chatbot have to say it's an AI?
Yes, since August 2, 2026, unless it's obvious to an average customer. For a customer service chatbot, the Commission considers that it generally isn't obvious: display the disclosure from the first message.
Are emails sent by an AI covered?
Yes, when the AI writes and sends them without human review. A signature line such as “reply written by the [Brand] AI assistant” is the simplest way to comply.
Does a reply suggested by AI and sent by an agent need to be disclosed?
In principle no, if the agent actually reviews the reply before sending it: the exchange is no longer direct between the AI and the customer. An automatic or purely formal approval doesn't count.
Did the Digital Omnibus postpone the transparency obligation?
No. Regulation 2026/1744 postponed the rules for high-risk AI systems to December 2, 2027 and August 2, 2028, but the obligation to disclose that people are talking to an AI has applied since August 2, 2026.
Does the AI Act require offering a human agent?
Not directly for a limited-risk chatbot. The GDPR, however, regulates significant decisions made solely by a machine, and your customers expect it: always explain how to reach a human.
What does an online store risk if it doesn't comply?
Up to €15 million or 3% of worldwide turnover, with the lower amount applying to SMEs. The most immediate risk is losing the trust of a customer who discovers they were talking to an AI.
Want to see how Klark discloses its AI replies on your channels? Request a demo.







