Since 2 August you have to tell customers they're talking to an AI

On 2 August 2026 the transparency rules in the EU AI Act started to apply. Most of the businesses we talk to don't know that, and some of them have had a chat box on their site for years.

This isn't legal advice, and we aren't lawyers. But it's a rule that bears directly on what we build, we've read it closely, and it's simpler than its reputation.

What the rule actually says

Article 50 is about openness. The core of the part that concerns you fits in one sentence: a person interacting with an AI system has to be told.

Two details are worth fixing on.

It has to happen at the start. The information must be given before, or at the very beginning of, the conversation. Burying it in the privacy policy doesn't count, and neither does telling the customer once they finally ask outright.

It applies to you, not only to whoever built the technology. The regulation distinguishes between the provider of an AI system and the party deploying it in their business. Buying the technology from someone else doesn't move the whole obligation across with it. It's your customer talking to your assistant, in your name.

The exception you shouldn't lean on

There is an exception: the information isn't needed where it's already obvious to a reasonably observant person that they're dealing with an AI.

That sounds convenient, and it's a trap. The European Commission reads it narrowly — it's meant for situations where almost no doubt remains about the artificial nature of the interaction.

Think about what that means for a modern voice agent. The entire point of good speech synthesis is that it doesn't sound synthetic. The better your AI sounds, the weaker your exception gets. A chat box with a robot icon in the corner might arguably be obvious. A warm voice answering the phone on the third ring is not.

Our position is simple: the exception isn't worth litigating. Saying it costs half a sentence.

What it means in practice

For chat: it belongs in the first message, not in a tooltip.

For phone: it belongs in the greeting, not after the customer has explained why they called.

And it has to be intelligible. "This service uses machine learning technology" tells nobody anything. "Hi, you're speaking with Kim, the AI assistant at …" does.

The worst case isn't forgetting. The worst case is a guest talking for two minutes, growing suspicious, asking outright, and being told yes. Now you don't just have a compliance problem, you have a customer who feels tricked, and the second one is harder to repair.

How we make it impossible to forget

You write your own greeting with us. The tone is yours, the name is yours, the wording is yours.

What you can't do is save a greeting that doesn't say it's an AI. The field refuses. The check runs as you type, so you see it immediately, and again on the server when it saves — the same check in both places, so they can't drift apart and let something through by a side door. We accept a range of ordinary phrasings, because it should be possible to sound like yourselves rather than like terms and conditions.

There's no setting to turn it off. That's deliberate. A checkbox for not having to mention you're an AI is a checkbox somebody ticks during a rushed onboarding, and then it's us, the people who built the tool, who made it easy to get wrong.

And GDPR, which applied already

The AI Act replaces nothing. A conversation with an AI assistant contains names, phone numbers and personal circumstances — personal data — and the same rules apply as always.

Three things worth checking with whichever vendor you pick.

Where does the data live? With us, in the EU. That's a decision made at the foundation, not something configured away per customer.

What happens to old conversations? They shouldn't sit there forever because nobody decided anything. With us they age out automatically on inactivity, and erasure and export are logged as events in their own right, so you can show afterwards that it actually happened.

Can an individual have their data erased? Yes, and it should be a feature, not a support ticket where somebody goes rummaging in a database.

Here we should be honest about a limit. If you choose to run WhatsApp, Messenger or Instagram as a channel, those are Meta's platforms, and traffic there follows Meta's terms and routing, not ours. That isn't something we can solve for you with a toggle in our portal. It belongs in your own record of processing activities, and it's a deliberate choice worth making with your eyes open. Phone, SMS, email and web chat don't carry that problem.

What to do this week

Four things, and together they take a morning.

Read the first thing your chat or phone agent says, word for word, and decide whether a customer would understand it isn't a person. Check every channel you're running — it's easy for the web chat to disclose it while the phone greeting doesn't.

Find out where your current vendor stores conversations, and for how long. If you don't get a straight answer, that's an answer.

Make sure there's a route to a human, and that the route works. It isn't an Article 50 requirement, but it's the first thing an unhappy customer looks for, and it's cheaper than a complaint.

And talk to your own lawyer before drawing conclusions from a text written by a vendor who wants to sell to you. Including this one.

Why we're writing about it

Honestly: it suits us. We've built for this since before it applied, and a customer who asks our competitors those four questions often comes back.

But there's a better reason. An industry where customers gradually discover they've been tricked into talking to a machine is an industry where nobody gets to have an AI answering the phone in a few years. Saying it costs half a sentence at the start of a call. We've never once seen someone hang up at that moment.

If you'd like to hear what a correct greeting sounds like in practice, book a demo and we'll set Kim up on your own information.