Most major voice AI platforms can now detect voicemail and either hang up or leave a message: Twilio (Answering Machine Detection, which can wait for the beep), Vapi (a voicemail tool that leaves a configured message or ends the call), Retell (hang up, leave a message or navigate an IVR), and Edesy (hang up, leave a custom message, or schedule a callback). What differs is how each one detects voicemail, how often it gets it wrong, and whether it understands the carrier announcements Indian callers actually hear.
This guide compares the options, shows what voicemail detection is worth on a campaign, and covers when to leave a message and when not to.
How do AI voice platforms detect voicemail?
There are two broad approaches, and many platforms combine them:
| Approach | How it works | Strength | Weakness |
|---|---|---|---|
| Telephony-level answering machine detection (AMD) | Analyses the audio pattern at the start of the call, and can wait for the beep | Decides before the agent speaks | Takes time to be sure; short human greetings can confuse it |
| Conversation-level detection | The AI listens to what is said ("please leave a message after the tone") and decides | Understands words, languages and carrier messages | Needs a second or two of speech to judge |
Both approaches make two kinds of mistake. A false positive hangs up on a real person, which is the expensive error because you lose a live lead. A false negative lets the agent talk to a machine, which wastes minutes. Good systems are tuned to avoid the first even at the cost of a few extra seconds.
Which platforms support voicemail detection and leaving a message?
| Platform | How it detects | What it can do on voicemail |
|---|---|---|
| Twilio | AMD: Enable returns a result as soon as it identifies the called party; DetectMessageEnd waits for the end of the greeting, usually the beep |
Hang up, or play a message after the beep; results include human, machine_start, machine_end_beep and unknown |
| Vapi | Voicemail detection plus a voicemail tool the assistant can call when it hears a greeting | Leave the assistant's voicemail message, a custom script, or end silently |
| Retell AI | Voicemail detection toggle in agent settings | Hang up, leave a message (with dynamic variables), or navigate an IVR menu |
| Edesy | The agent listens for confidence signals and exits only when it sees two or more | Hang up (default), leave a custom message, or schedule a callback |
Sources as of September 2026: Twilio AMD docs,
Vapi voicemail tool,
Retell voicemail handling. Twilio notes a
trade-off worth knowing: the default detection timeout is 30 seconds, and shortening it gives faster
answers but more unknown results.
Try it yourself: Build this agent free on Edesy — self-serve, no demo needed.
How Edesy's voicemail detection works
Edesy treats voicemail as a judgement the agent makes from what it hears, with a deliberate bias against hanging up on real people:
- Two signals before exiting. The agent exits only after it observes two or more signs of a machine, such as a recorded greeting, "please leave a message", a long monologue with no pause, music or hold tones, or a "press 1" menu.
- Human cues keep it talking. A short, questioning "Hello?", or in Hindi "Haan?", "Boliye?" or "Kaun bol raha hai?", counts as a person. If it is unsure in the first few seconds, it repeats the greeting and waits.
- Indian carrier messages. It recognises carrier announcements such as "the number you have called is not reachable or switched off", including Hindi versions like "Aap jis number par call kar rahe hain woh abhi reachable nahi hai".
- Your choice of action, set per agent: hang up (the default), leave a custom message, or schedule a callback so the contact is tried again later.
The voicemail detection use case page summarises the setup, and our engineering deep-dive on voicemail detection approaches goes into the techniques.
Why voicemail detection matters for campaign cost
On an outbound campaign, every second spent talking to a machine is billed. An illustrative example, with assumptions you should replace:
| Assumption | Value |
|---|---|
| Dials in the campaign | 10,000 |
| Reaching a machine or carrier message | 20% (2,000 calls) |
| Time talking to the machine without detection | 30 seconds |
| Time before exiting with detection | 5 seconds |
Without detection: 2,000 x 30 seconds = 1,000 minutes. With detection: 2,000 x 5 seconds = about 167 minutes. That is roughly 830 billed minutes saved, and, just as important, those 2,000 contacts are now correctly marked for a retry at a better time instead of being logged as "completed". Our AI outbound calling guide shows how retries by outcome fit into a campaign.
Voicemail in India is different
Most voicemail detection was designed around US-style greetings. Indian campaigns look different:
- Personal voicemail is less commonly used, so a large share of unanswered calls end in carrier announcements: not reachable, switched off, not in service, busy.
- Those announcements are often in Hindi or a regional language, which a detector tuned on English greetings can miss.
- The right outcome is "unreachable, retry", not "voicemail left" and certainly not a 30-second monologue to a recording.
When you evaluate a platform for Indian calling, test it on real carrier announcements in the languages on your list, not just on a recorded English greeting.
Should the agent leave a voicemail or hang up?
| Call type | Recommendation | Why |
|---|---|---|
| Appointment or payment reminder | Leave a short message | The customer can act on it: date, time, callback number |
| First-touch sales or lead callback | Hang up and retry | A live conversation converts far better than a recording |
| Loan, health or other sensitive calls | Leave only a neutral message | Anyone may hear a voicemail; never include amounts, diagnoses or account details |
| Survey or feedback | Hang up and retry | A voicemail rarely gets a response |
Best practices for voicemail in AI campaigns
- Keep messages under 20 seconds: who is calling, why, and how to call back.
- Leave no sensitive details. "This is ABC Finance, please call us back on this number" is fine; the EMI amount is not.
- Wait for the beep before speaking if your platform leaves messages, or the start gets cut off.
- Retry at a different time of day rather than immediately; people who miss a 10 am call often answer at 6 pm.
- Review recordings of calls marked as voicemail every week to catch false positives, and adjust settings if real people are being dropped.
FAQ
Which voice AI platforms can detect voicemail and leave a message?
Most major platforms can. Twilio's Answering Machine Detection can wait for the beep so you can leave a message. Vapi has a voicemail tool that leaves a configured message or ends the call. Retell can hang up, leave a message or navigate an IVR menu. Edesy lets each agent hang up, leave a custom message, or schedule a callback when it detects voicemail.
How does AI voicemail detection work?
There are two main approaches. Telephony-level answering machine detection analyses the audio pattern at the start of the call and can listen for the beep. Conversation-level detection has the AI listen to what is said, such as "please leave a message after the tone", and decide. Many platforms combine both.
Should an AI agent leave a voicemail or hang up?
Leave a short message when it helps the customer act, such as an appointment reminder with a callback number. Hang up and retry later for first-touch sales calls, where a live conversation converts far better. Never leave sensitive details such as loan amounts or health information in a voicemail.
Is voicemail detection different in India?
Yes. Personal voicemail is less commonly used in India, and many unanswered calls end in carrier announcements such as "the number you are calling is not reachable" or "switched off", often in Hindi or a regional language. Your platform should recognise these and treat them as unreachable attempts to retry, not as conversations.
How much can voicemail detection save on an outbound campaign?
It depends on how many calls reach machines and how quickly the agent exits. As an illustration, if 2,000 of 10,000 dials reach a machine and the agent drops each after 5 seconds instead of talking for 30, you save roughly 830 billed minutes, before counting the retries you can now schedule properly.
Start free
You can run a test campaign with voicemail detection switched on today. Sign up and build your first agent in minutes, no sales call needed. Running large Indian campaigns? Talk to us.