Image AI-generated. All text human-written.
Every other obligation in the EU AI Act is invisible from the outside. Yours answers the phone.Consider what it normally takes to catch a company out. Risk classifications sit in internal files. Conformity documentation lives on a shared drive. Governance gaps are visible only to the people who work there. For most of the Act, non-compliance is a private failure. It takes an insider, a leak or an inspection to surface it.
Voice AI breaks that pattern completely.
When an AI voice agent answers a call and never tells the caller it is an AI, the breach is audible. It happens in the open, to strangers, hundreds of times a day. Every one of those callers is holding a recording device.
Since 2 August 2026, Article 50(1) has required providers to design and develop AI systems that interact directly with people so that those people are informed they are dealing with an AI — unless that fact is obvious to a reasonably well-informed, observant and circumspect person. The Commission’s transparency guidelines name AI-enabled voice assistants explicitly. Article 50(5) settles the timing: the disclosure must be clear and distinguishable at the latest at the first interaction.
Not in your terms of service. Not on your website. On the call, at the start. And not as a future obligation — as a live one, enforceable today.
Who actually owns this
The obligation lands on the provider at the design stage. Build the voice agent and the disclosure has to be engineered into the product — a CTO deliverable, not a legal memo. Buy one and point it at your customers, and you carry deployer duties of your own.
The accountable person is the CEO. In the audits ExecLevel has run, the same conversation plays out every time: the product team assumed legal had it, legal assumed the vendor had it, and the vendor’s contract said the customer was responsible. Three parties, one obligation, nobody actually holding it.
The part nobody has told your board
Article 85 gives any natural or legal person who has grounds to consider that there has been an infringement the right to lodge a complaint with the relevant market surveillance authority.
Read that phrase again. Any legal person. A company is a legal person.
Your competitor does not need to have been harmed. They do not need an insider. They do not need sight of your documentation. They need your phone number.
The cost of gathering evidence against you is one phone call. The cost of filing is a web form.
Article 99(4)(g) places Article 50 transparency breaches in the tier of up to €15 million or 3% of total worldwide annual turnover, whichever is higher.
The fine is the slow damage. The fast damage is the investigation. A market surveillance authority can demand documentation, order corrective measures and restrict how you sell. Meanwhile a procurement team at your largest prospect asks one question — is this supplier subject to an open regulatory complaint? — and the deal stalls while you compose an answer.
The deadline that is actually next
There is a follow-on obligation most voice companies have not diarised. Article 50(2) requires providers of systems generating synthetic audio to mark outputs in a machine-readable format, detectable as artificially generated. Cloned and synthesised voices sit squarely inside that.
For generative systems that were already on the market on 2 August 2026, that marking duty runs to 2 December 2026 — which makes it the nearest live deadline on the voice-AI calendar, and a genuine engineering task rather than a script change. Systems placed on the market since 2 August carry the duty already.
The upside nobody is using yet
Here is what makes this market strange. The same asymmetry that exposes you is the cheapest competitive advantage on the table.
Two seconds of audio at the top of every call. “You’re speaking with an AI assistant.” The complaint dies before anyone can file it.
Then say so, in writing, in every tender you enter. Most of your competitors have not thought about Article 50 at all. They read the headlines about high-risk rules being delayed and assumed the whole EU AI Act moved with them. It did not. The high-risk regime moved to December 2027. Transparency did not move at all — and it has been enforceable for a month.
A voice-AI provider that can evidence Article 50 compliance in a procurement pack today is selling against rivals who cannot answer the question.
What to do this month
Call your own product. Record it. Listen to the first ten seconds as a stranger would and ask whether a caller is genuinely told they are speaking to an AI, clearly, at the start.
If the answer is no, or “sort of,” or “it depends which flow they land in,” you have a design task and a fortnight.
That gap — knowing an obligation applies but not what to build, and not wanting a six-figure programme with a large advisory firm to find out — is why ExecLevel AI exists. Classification, evidence and a defensible position, in weeks.
See where your Article 50 position stands at www.execlevelai.eu/article-50
Disclaimer: This article is for general informational purposes only and does not constitute legal or compliance advice. EU AI Act requirements may vary depending on your organisation, AI system, and use case. For advice on EU AI Act compliance, visit www.exclevelai.eu