Someone calls the company and asks to change the bank account on their profile. The voice is calm and polite. It may be the customer. It may be a personal agent running an errand. It may be a cloned voice posing as the customer. According to ElevenLabs, a growing share of the calls reaching businesses and governments are no longer people, and since 9 October the company sells a way to tell which is which.
Fact
The facts: on 9 October 2026 ElevenLabs released synthetic voice detection in its agent platform ElevenAgents. The system analyzes the caller's speech in the first few seconds of a call and returns a determination: human or AI generated. The company's rules decide what happens next. ElevenLabs' examples: humans move up the queue, machines go to an agent with its own instructions and limits, and a synthetic voice requesting a sensitive action such as a bank account change or a password reset is stopped before it gets information. The check runs on the live audio and looks for traces a human listener would not notice. All audio generated with ElevenLabs carries a watermark, but audio from other sources often does not, so detection does not depend on the watermark. For now the feature is available to enterprise customers supported by the Forward Deployed Engineering team; later this month ElevenLabs promises to bring it to a broader group of enterprise customers as an option in ElevenAgents. The same day the company announced it is joining the Personal Agent Protocol working group led by Meta and Sierra as a design partner.
The most honest sentence in the announcement is about the watermark. ElevenLabs marks all the audio it generates, but audio from other sources often is not marked. So the detector does not wait for the mark, it listens to the voice itself. The company that sells synthetic voices now also sells the tool that catches them, and says itself that its watermark is no help with anyone else's audio.
No number. How often the system mistakes a human for a machine and the other way round, the announcement does not say. For a feature that can stop a call, that is the first question you ask, before you switch it on.
The honest agent will introduce itself. The clone will not.
ElevenLabs is taking two roads at once. It joins the Personal Agent Protocol working group, where the rules are being written for how an agent tells a company who it is, whom it represents and what it is allowed to do. And it promises, as such standards take shape, to add support for agents that declare themselves. Voice detection stays for the ones that will not declare.
If you take orders or data changes by phone, the rule can be written today, with or without a detector: an account or password change does not go through on a voice alone, however familiar it sounds.
Official primary sources
ElevenLabs - Defining how AI agents interact with your business, 09.10.2026The visual is generated code-art, no third-party images.