AI agent company Sierra said on Oct. 8 it has unveiled a model called Fleming-1 that can tell whether a caller is a person or an AI agent.

Sierra said that as personal AI agents increase, a significant share of calls received by companies could come from AI calling on behalf of consumers.

Sierra on Oct. 7 also announced a "Personal Agent Protocol" in cooperation with other tech companies. If an agent uses the protocol, a company can know the counterparty is an agent and who it represents. The problem is that some agents do not follow the protocol. Fleming-1 was introduced with such cases in mind.

The company said it is becoming increasingly difficult to distinguish human voices from AI voices on phone calls. It said AI voice quality keeps improving and noise or unstable connections can also interfere with judgment. Fleming-1 analyzes the other party's voice in real time during a call and assigns a score on whether the voice was generated by AI. It looks beyond what is audible to the human ear and determines whether it is synthetic before the call ends.

Calls judged to be AI are labeled. The default setting is conservative to prevent cases in which a real person is mistakenly classified as AI, the company said.

Sierra stressed that Fleming-1 does not make decisions and only provides information. It said companies decide how to respond. Fleming-1 can be used in all voice agents built with Sierra.