Definition
Answering machine detection, or AMD, is the step on an outbound call that decides whether a person or a machine, such as voicemail, picked up. The decision comes from the first seconds of audio, so it is a guess.
How answering machine detection works
There is no signal on a phone call that says “this is a machine”. The system infers it from the audio after the call connects. A person usually says a short greeting, such as “Hello?”, and then goes quiet to hear a reply. A voicemail greeting runs longer, has no pause for a reply, and ends with a beep. A detector looks at the length of the first utterance, the silence after it, the pattern of the speech and whether a beep tone arrives. Some carriers offer detection as a service on the call. Other systems run their own classifier on the audio stream.
Two errors are possible, and they cost different things. If a person is labeled a machine, the agent may leave a message to someone who is standing there, or hang up on them. If a machine is labeled a person, the agent starts its opening line over the greeting, so the voicemail records half a sentence and the minutes are wasted. Waiting longer before deciding improves accuracy and lengthens the dead air a real person hears.
Why it matters for voice agents
A human caller can hear a voicemail greeting at once and hang up or leave a message. A voice agent has to be told, either by the telephony layer or by its own listening. Without detection, an agent can talk through a whole greeting, or wait for a reply that never comes, then speak into the beep.
Cartesia Managed Agents have a voicemail detection tool. When enabled, it detects voicemail, speaks a configured message of up to 1,000 characters, and ends the call. Leaving no message is an option.
The message itself needs care. An agent that leaves a recorded message is still placing a call with an AI-generated voice, and consent rules such as the TCPA can apply to it. Ask counsel before you decide to leave one.
Example
An illustrative appointment reminder: the call connects and the agent waits 1.5 seconds. It hears a 9 second greeting followed by a beep. It classifies the call as voicemail, speaks a 15 second message with the clinic’s number, and ends the call. On a person, it would have started its normal opening instead.