Your AI Employee Now Has to Introduce Itself. Design For That, Not Around It.

As of August 2, if your AI talks to people in the EU, it has to tell them it is an AI. Not in the terms of service. Not on a help page. Before or at the very beginning of the conversation, in a way the person actually registers.

That is Article 50 of the EU AI Act, and it is worth understanding precisely because it does not behave like the rest of the regulation. The high-risk provisions everyone has been panicking about apply to a defined list of use cases: hiring, credit, education, critical infrastructure. Article 50 sits outside that structure entirely. It applies to any system that interacts directly with a person, regardless of how risky that system is. You can have zero high-risk AI in your company and still be fully in scope because you put a chatbot on your pricing page.

Penalties run to 15 million euros or 3% of worldwide turnover. There is a transitional grace period until December 2 for marking and detection obligations on systems already on the market, but the disclosure duty for interactive systems started immediately.

Most vendors are treating this as a compliance chore. Add a banner, log the consent, move on. That reading is a mistake, and companies that make it are going to spend the next two years fighting their own product.

The pretending era is ending anyway

Strip out the legal framing and ask a simpler question. Was the plan ever to have your AI pass as human?

For a lot of the market, quietly, yes. The implicit benchmark for voice agents has been indistinguishability. Vendors demo the pause, the breath, the "um," the barely perceptible typing delay. The unstated pitch is that customers will behave better if they do not know.

That was always a short position on your customers' intelligence, and it is being closed out by events rather than by law. Google's agents now call local businesses on a customer's behalf and hang up if the front desk is too slow. When those agents reach a business running its own voice agent, two machines negotiate and hand humans the outcome. People know. They have been in enough of these conversations to recognise the shape of one, and the tell is never the voice, it is the moment you ask something slightly off-script.

So the choice is not between disclosure and successful deception. It is between disclosure and being caught. And getting caught is expensive in a way a fine is not, because the customer does not file a complaint, they just stop trusting anything else you say.

Disclosure changes the conversation in your favour

Here is what the compliance-chore reading misses. People talk to machines differently than they talk to people, and the machine version is easier to serve.

When someone knows they are talking to an AI, they get more direct. They drop the social preamble. They ask the actual question instead of the polite approximation of it. They are more willing to state a budget, a deadline, or a complaint in plain terms, because there is nobody on the other end to be embarrassed in front of. Anyone who has watched real transcripts knows this: the disclosed conversations are shorter and land on the actual intent faster.

They also calibrate their expectations correctly. A person who thinks they are talking to a human and hits a limitation experiences it as incompetence. A person who knows they are talking to an AI and hits the same limitation experiences it as a boundary, and asks for a human. That escalation is the single most valuable moment in the interaction, and pretending to be human is exactly what suppresses it. You do not want a customer politely working around your agent's blind spot for ten minutes. You want them to say "put me through to someone" the second they need to.

Undisclosed AI optimises for a smooth conversation. Disclosed AI optimises for a finished task. Those are not the same metric, and only one of them shows up in your revenue.

What good disclosure actually looks like

The instinct is to write an apology. Do not. A disclosure that sounds defensive teaches the customer that something is wrong.

Four things, in the first breath:

Name it. "I'm Jessica, the AI assistant for [company]." A name plus the word AI. Not "an automated system," not "a virtual representative," not any of the euphemisms people reach for when they are slightly ashamed of the product.

State the capability, not the limitation. "I can book, reschedule, and answer questions about your account." This sets the frame as competence with edges rather than a lesser substitute for a person.

Offer the exit immediately. "Say 'human' any time and I'll transfer you." Offering it up front means it gets used less, not more, because the customer stops testing whether you will let them out.

Then get to work. No lingering on the AI thing. One sentence, and into the task.

That whole preamble costs about six seconds. In exchange you get a customer who knows what they are dealing with, knows how to escalate, and is not spending cognitive effort trying to work out whether you are real.

Two ways to get this wrong

The first is performative compliance: technically disclosed, practically hidden. A grey line of eight-point text under the chat window. A disclaimer spoken at 1.8x speed. A cookie-style consent nobody reads. This satisfies a checkbox and fails the actual requirement, which is that the person concerned knows. Regulators have been reading badly-implemented cookie banners for six years. They will recognise this genre on sight.

The second is worse: the mid-conversation reveal. The agent presents ambiguously, the customer assumes human, and twelve minutes in something forces the truth out. Every prior statement gets retroactively re-evaluated. That is not a compliance failure, it is a trust event, and it costs far more than the transaction you were trying to close.

The part that is actually strategic

Disclosure is going to become the norm well beyond the EU, because norms follow the strictest large market and because customers will start expecting it. When that happens, the differentiator stops being whether your AI announces itself and starts being what happens in the next sentence.

If your agent introduces itself and then handles the request completely, you have just demonstrated competence under honest conditions, which is the most durable form of trust available. If it introduces itself and then flounders, disclosure did not cause that. It just removed the cover.

That is the real reason to build for this now rather than in November. Disclosure is a stress test on whether your agent is actually good. Vendors resisting it are usually telling you something about their product they would rather you did not notice.

Your AI employee should introduce itself for the same reason a new hire does when they walk into a client meeting. Not because a rule says so, but because anyone worth working with says who they are first.

Want to test the most advanced AI employees? Try it here: https://geta.team