How does an AI voice agent know it has reached voicemail?
Through answering machine detection, usually shortened to AMD. When an outbound call is answered, the system listens to the first seconds and classifies what it hears: a person, a voicemail greeting, a phone menu or an unreachable line. It can do this at the carrier, by timing the greeting, or inside the agent, by understanding the words. Neither is perfect. Twilio's documentation says its detection "will not always return the right answer", and LiveKit publishes no accuracy figure for its own. The practical job is to choose what happens on each result and to make a wrong guess harmless. Botcadence's outbound voice agent has optional voicemail detection on campaigns that hangs up when it hears an answering machine.
Key takeaways
- A voicemail system answers the call just as a person does, so the phone network alone cannot tell them apart.
- Two approaches: carrier detection based on audio timing, and agent detection that classifies the first thing said.
- No vendor we read publishes a dependable accuracy figure. Expect errors both ways.
- Decide five behaviors: person, voicemail, phone menu, unreachable, uncertain.
- Carrier detection has a per-call price: $0.0075 on Twilio and $0.002 to $0.0065 on Telnyx.
- A message left by an AI voice is still an artificial voice call under the TCPA, with the same consent and identification rules.
Why voicemail is hard to detect
LiveKit's outbound calling documentation notes that a voicemail system answers with the same signal as a person picking up. From the network's point of view the call connected. Everything after that is inference from audio:
- A person usually says a short "hello?" and waits.
- A voicemail greeting is longer, continuous, and ends with a beep.
- A phone menu speaks in a different pattern again and waits for a choice.
Short custom greetings ("Hi, it's Sam"), people who answer with a long sentence, and carriers' call-screening features all blur the line.
Two ways to detect it
| Carrier-level detection | Agent-level detection |
|---|
| How it decides | Timing and energy of the first audio: length of speech, pauses, a beep | A model classifies the first utterance by what was said |
| Where it runs | In the carrier's network, before or as the call is bridged | In the voice agent, on the first thing it hears |
| Outcomes | Typically human, machine, fax, unknown | LiveKit's are human, machine-ivr, machine-vm, machine-unavailable and uncertain |
| Cost | Per call: $0.0075 on Twilio; $0.002 standard or $0.0065 premium on Telnyx | Usually part of the agent's normal processing |
| Weakness | Short greetings and long hellos confuse it; it can add a pause before the agent speaks | It needs to hear some speech first, and depends on the model |
LiveKit's documentation says its detection runs once, on the first utterance. That is the approach Botcadence uses on outbound campaigns: the classifier listens to the opening words, with examples of common United States and India voicemail greetings to guide it.
What nobody publishes: accuracy
We looked for accuracy figures on the vendors' own documentation and found none we could rely on. Twilio states only that the result will not always be right. LiveKit lists the outcomes without a rate. Any percentage you see on a sales page should be read with the question "measured on which calls, in which country, on which carrier?"
Two kinds of error matter, and they cost different things:
| Error | What happens | Cost |
|---|
| A person is classified as a machine | The agent hangs up on, or leaves a message for, a live person | A lost conversation and an annoyed contact |
| A machine is classified as a person | The agent talks to a greeting, then to silence after the beep | Wasted minutes and a garbled voicemail |
Tune toward the error you can tolerate. For calls to your own customers, hanging up on a person is worse; lean toward treating uncertain results as human.
What to do on each result
| Result | Sensible behavior |
|---|
| Person | Start the conversation. Identify the business at once |
| Voicemail | Hang up and retry later, or leave one short, approved message |
| Phone menu | Hang up, unless the call is meant to reach an extension and the agent can navigate it |
| Unreachable or disconnected | Mark the contact and stop retrying |
| Uncertain | Treat as a person: greet, and listen for a reply |
On Botcadence, with voicemail detection switched on for a campaign the agent hangs up when it hears voicemail or an unreachable line. With it off, the agent is instructed to leave one short message. Detection is off by default and is set per campaign. Retries follow the campaign's rules: the first call plus up to three retries by default, one a day per contact, at least four hours apart.
Should the agent leave a message?
Arguments for hanging up: no risk of a clumsy half-message, no minutes spent on a greeting, and the contact sees a missed call from a number they can call back.
Arguments for a message: the contact learns who called and why, which raises callbacks for expected calls such as appointment confirmations.
If you leave one, keep it under twenty seconds: who is calling, why, and the number to call back. And remember the law. The FCC treats AI-generated voices as artificial voices under the TCPA, and its rules require an artificial voice message to state the business's identity at the beginning and give a phone number. Consent requirements apply to the call whether or not anyone answers. See are AI phone calls legal?.
What it costs
| Item | List price, read November 12, 2026 |
|---|
| Twilio answering machine detection | $0.0075 a call |
| Telnyx answering machine detection | $0.002 a call standard; $0.0065 premium |
| Retell AI batch calling | $0.005 a dial |
| Time on voicemail | Billed as call time on most platforms; on Botcadence at least 15 seconds for an answered call |
A campaign of 2,000 dials with Twilio's detection adds $15. The larger cost is minutes spent listening to greetings when detection is off or wrong.
How to test it
- Call ten phones you control: some answered, some sent to voicemail, one with a short custom greeting, one with a carrier's default greeting, one with call screening on.
- Note the result for each and how long the agent took to speak.
- Answer with a long sentence ("Hello, this is Maria speaking, who's calling please?") and see whether you are treated as a machine.
- Check what was left on each voicemail.
- Repeat on each carrier you call most.
Where Botcadence fits
Botcadence's outbound campaigns include optional voicemail detection based on LiveKit's answering machine detection, with hang-up on voicemail, plus calling hours by time zone, limited retries and automatic opt-out when a contact asks not to be called. It does not publish an accuracy figure, and you should test it on your own contact list as described above. For keeping outbound numbers in good standing see how to reduce Spam Likely labels on AI calls, and for what the calls can do see the outbound sales voice agent page.