Key takeaways
- Keep human, machine, fax, silence, and unknown as distinct detection results.
- Choose detection sensitivity by the cost of a false human versus a false machine result.
- Make voicemail content, timing, identity, disclosure, opt-out, and retry rules explicit.
- Measure detection errors and buyer response, not messages left alone.
01
The beep is part of the workflow, not an edge case
Outbound voice programs meet personal greetings, carrier announcements, business attendants, long silences, fax tones, and people who answer without speaking. If every answer is treated as a live person, the agent can begin an awkward pitch into a recording. If every pause is treated as voicemail, it can abandon a real caller.
Twilio’s answering-machine detection can return human, machine, fax, and unknown-like outcomes with more detailed reasons depending on configuration. Keep the raw result and timing evidence. Map it to a business action separately so a later tuning change does not erase what the provider actually reported.
Swipe to compare every column
| Detection result | Default response | Important exception |
|---|---|---|
| Human | Begin disclosed live flow | Very short greeting or silence can be misclassified |
| Machine | Leave permitted message or hang up | Do not speak before the greeting ends |
| Fax | Stop voice playback and suppress repeat attempts | Confirm routing policy for shared numbers |
| Unknown | Use a conservative fallback | Do not relabel uncertainty as a human contact |
02
Tune for the mistake that costs more
Detection requires a trade-off between speed and certainty. A longer analysis window may identify a machine more accurately while creating an uncomfortable pause for a person. A faster threshold may improve pickup experience and increase false classifications. Choose by call purpose and consequence, then test on the actual markets and carriers.
Twilio recommends testing parameters such as silence and speech thresholds and notes that asynchronous detection changes when the call begins. Record configuration with every call. A metric that mixes several threshold versions cannot tell the team whether the latest tuning helped.
03
Write a voicemail policy before writing the message
Define which calls may leave a message, who the message may identify, what contact details it can mention, the maximum length, retry limits, local calling hours, suppression, and how a person can stop further contact. Avoid sensitive account or health information when the recording could be heard by someone else.
In the United States, automated and prerecorded calls can fall under the TCPA and Telemarketing Sales Rule depending on their content and context. The FCC’s AI-voice ruling and the FTC’s operational guidance should be reviewed with counsel for the actual program. A technically successful voicemail is not permission to send one.
04
Measure the call after the message ends
Review a labeled sample for false human, false machine, fax, and unknown results. Track time to classification, message start after the beep, truncated messages, hang-ups, repeat attempts, callbacks, opt-outs, complaints, and qualified next steps. Use recordings only under a documented consent, access, and retention policy.
Separate “message left” from “person reached” and from “buyer responded.” If the team rewards voicemail volume, the system will happily leave more messages. The useful question is whether the contact was respectful, compliant for the program, and capable of producing a wanted next step.
Primary sources and further reading
Use the source material to validate details against your own context and current platform configuration.
- Twilio: Answering machine detection
- Twilio: Answering machine detection FAQ and best practices
- FTC: Complying with the Telemarketing Sales Rule
- FCC: AI-generated voices in robocalls
This field note follows the XenGrowth editorial policy: primary sources where available, visible limitations, material review dates, and no invented first-hand experience.
Stay with the problem
