What's changing
From 2 August 2026, a new part of the European AI Act applies. The rule is simple: if someone is talking to AI, they need to know. And from the very start of the conversation. This covers chat, WhatsApp, email and phone. Including conversations you initiate yourself.
The law doesn't specify what the notification has to sound like. It does specify the outcome: clear, understandable, and audible or visible within the conversation itself. A clause in your privacy policy isn't enough.
We deliberately choose to use the word AI. The law requires that someone knows they're interacting with an AI system — not that the word itself must literally appear. But "digital assistant" or "virtual employee" could just as easily describe a phone menu. That's exactly the confusion the rule is designed to prevent. So: AI.
What's already in the platform
The obligation sits with the provider of the AI system, which is us. We've built the transparency notification into the platform — not as a setting per customer.
In practice:
- The agent states at the opening that it's an AI assistant, in the language of the conversation.
- If someone asks whether they're talking to a person, the agent confirms it's AI. It never claims to be human.
- In chat and WhatsApp, the word AI is also part of the agent's name.
- This applies across all channels, all languages, and outbound conversations too.
- The notification can't be switched off.
The route to a human agent is also built in — but that's our product choice, not a legal requirement. Article 50(1) only requires the opening notification.
Any recording notification you already had is separate and unchanged.
What's coming next: audio watermarking
Article 50(2) also requires watermarking of AI-generated audio. That deadline is 2 December 2026, with a grandfathering provision for systems already on the market before August 2026, giving roughly four additional months. Text-to-speech and voice cloning are explicitly covered. We'll keep you informed when this becomes relevant.
What you'll want to check yourself
Five things that fall outside our platform:
- Other places where you use AI to talk to people, such as a chatbot from another provider on your site.
- Your privacy policy and website. Explain how you use AI and what happens to conversations — not as a replacement for the notification, but in addition to it.
- Whether your staff know what the agent does and doesn't do, and when they need to step in.
- Whether there's a clear route to a human on every channel, including outside office hours.
- Whether your use case involves more than just transparency. More on that below.
About risk categories, honestly
You'll increasingly see claims that AI for customer contact is "limited risk" and that's all you need to worry about. That's too simplistic. The regulation looks at how you're using it, not at the product itself.
It's worth knowing what the law actually means by high risk. Annex III point 5(d) covers systems deployed by emergency services: 112, ambulance, fire brigade, police. Not a garage sorting calls by urgency, or a customer service team triaging enquiries. A standard telephony application doesn't reach that threshold.
And even if a use case were to qualify: the high-risk deadline is 2 December 2027, not August 2026. That pressure doesn't exist yet.
We assess per use case, not in general. If something carries more weight, we'll tell you. For most use cases it comes down to transparency, and that's covered.
What we do with your conversations
Conversations aren't shared with third-party models and are only used for the customer's own agent. Personal data is removed before storage and analysis for improvement. During live handling, names and questions go into the agent in real time — the same as in any conversation with a human.
Not sure what this means for your data processing agreement or contracts? Run it past someone who knows your contracts.
Questions
Want to know what this means for your situation, or want our one-page checklist? Get in touch and we'll send it over.
