EU AI Act Article 50: voice agents must disclose they are AI

On 2 August 2026, Article 50 of the EU AI Act became applicable. One of its provisions goes straight to the heart of voice agents: the provider of an AI system that interacts directly with people must design that system so the person is informed they are dealing with artificial intelligence – unless this is obvious. The information has to be given at the start of the first interaction, in a clear and easily perceivable manner.

That is a few lines in a long regulation. But it moves a question many have treated as a matter of taste – should the agent say it is an agent? – into the remit of market surveillance authorities. Non-compliance can trigger fines of up to €15 million or 3% of global annual turnover.

The duty starts at the first second, not at the first complaint

The Commission adopted guidelines on Article 50 on 20 July, and they are clear on two points that are easy to miss.

The first is timing. The disclosure has to come at the start of the first interaction. Not in the terms and conditions, not in a follow-up email, not when the customer asks. On a phone call, that means the opening line.

The second is who the duty falls on. The regulation also applies to providers established outside the EU if the output of their system is used in the EU. It distinguishes between the provider – the party that develops the system and places it on the market – and the deployer, meaning the business that uses it in its own operations. The design obligation sits with the provider. But it is the deployer who faces the customer, and who gets the question in their lap when something goes wrong. That makes your choice of vendor a compliance question, not just a procurement one.

There is one piece of relief, and it does not cover this: the 2 December 2026 deadline is a transitional arrangement solely for the machine-readable marking of AI-generated content in systems already on the market – an obligation that sits with whoever supplies the voice, not with you. The disclosure duty in conversation has no such extension.

Article 50 also has a third paragraph that catches fewer organisations but hits them harder: if you run a system that infers emotion from the voice, a separate disclosure duty applies to it – and that part of the rulebook runs on a very different calendar from the one most people assume.

Norway is not exempt – even though the law is not Norwegian yet

This is where it gets confusing, and the confusion is worth clearing up.

The AI Act has been assessed as EEA-relevant, but it has not yet been incorporated into the EEA Agreement. Because it requires legislative changes and has budgetary and organisational consequences, it carries an Article 103 reservation: the Norwegian parliament has to approve incorporation. A draft Norwegian AI act went out for consultation in 2025, but the timeline has slipped, and the government is now aiming for a bill in spring 2027.

That does not mean Norwegian businesses can wait until 2027.

If your voice agent calls a customer in Denmark, Sweden or Germany, the output of the system is used in the EU – and the regulation applies. If you supply an agent to a customer with operations in the EU, you are a provider into the EU market. And for most Nordic companies the practical line is simpler still: you cannot run one transparency standard for Danish customers and another for Norwegian ones.

"Obvious" is a bad place to hide

The exception for the obvious is the part many hope to lean on. The guidelines shut that door fairly firmly: the assessment is made from the perspective of an average person who is reasonably well-informed, observant and circumspect, and the exception is to be interpreted restrictively – precisely because it deprives people of transparency.

The point is that the exception shrinks as the technology improves. Five years ago a robotic voice was obvious. In 2026, it is exactly because the agent sounds natural that the customer does not realise. The quality of your voice is the argument against you, not for you.

The same quality is available to people who have no intention of disclosing anything at all. The campaign against several of Wall Street's largest hedge funds in August went straight through the help desk, using voices the staff recognised.

How to say it without killing the conversation

This is the part that is actually hard, and it is not legal. It is operational.

The fear is real: that a mandatory disclosure in the opening line makes people hang up. But what makes people hang up is rarely the word "AI". It is long welcome messages, keypad menus, and an agent that cannot help.

Four practical moves:

Say it once, briefly, in the first sentence – then move on. "Hi, you're speaking with an AI assistant here. What can I help you with?" is enough. It is not a disclaimer to be read aloud.

Don't wrap it in legalese. A customer who gets a statutory citation over the phone becomes suspicious, not informed.

Offer a way out immediately. Transparency without an alternative is just a warning. An agent that transfers to a human on request makes the disclosure harmless.

Log it. You should be able to show that the disclosure was actually given – in every call, not in a screenshot from a demo.

The hard part isn't the script

Set against the rest of the AI Act, the disclosure duty is the easiest line to comply with. It is one sentence in a script.

What is hard is the documentation behind it: being able to account for what the agent said, who it spoke to, where the audio stream was processed and who holds it now. That is not decided by the script. It is decided by the architecture – by whether the call runs on infrastructure you know and control, or takes a detour you cannot account for.

Organisations that had already answered that question spent 2 August changing one opening line. The others discovered they have a larger project.

Threll.ai builds autonomous voice agents in Norwegian, Swedish and Danish, on Nordic telecom infrastructure. Transparency is easier to promise when you know where the voice goes.