Voicemail detection for AI calls: why AMD isn't enough

On a typical outbound campaign, many of the calls that connect don't reach a person. They reach a voicemail greeting, a full mailbox, or a phone's screening assistant. What your system does in those first two seconds affects cost, answer rates, your number's reputation and, for some industries, compliance.

Most guides to answering machine detection were written for diallers that pass calls to human agents. AI calling changes the problem, because the AI is the agent and can listen too.

How answering machine detection works

Classic AMD listens to the first seconds of audio after the call is answered and looks at the pattern of speech and silence. A person tends to say something short, "Hello?", then wait. A greeting tends to talk on: "Hi, you've reached Sam, I can't come to the phone right now, please leave a message after the tone."

Twilio's AMD documentation shows the knobs. By default it treats more than 2,400 ms of continuous speech as a machine, waits for 1,200 ms of silence before deciding that speech has ended, and gives up with "unknown" after 30 seconds. Each setting is a trade. Lowering the speech threshold catches short voicemails but labels more people as machines; raising it does the reverse. Twilio says plainly that a very short, two-second greeting may be taken for a human.

Carriers sell premium versions that add speech recognition. Telnyx's premium AMD returns results such as human_residence, human_business, machine, silence, fax_detected and not_sure, and recommends treating not_sure as a human. It can also report when a voicemail beep has sounded and when an iPhone call-screening prompt ends.

The two mistakes, and which one hurts more

Every detector makes two kinds of error, and they're not equally bad.

For human-agent diallers, regulators focus on the second one. In the US, the FTC's Telemarketing Sales Rule treats a call as abandoned if the person answers and isn't connected to a sales representative within two seconds of their completed greeting, and its safe harbour allows no more than 3% of answered calls to be abandoned per 30-day period. In the UK, Ofcom's persistent misuse policy has long treated AMD false positives as a cause of repeat silent calls: when it tightened the rules in 2010, it required that a number flagged as a machine be called back only when an agent was available, as Pinsent Masons reported.

Those rules were written for diallers that hold a call while waiting for a free human. An AI agent that speaks immediately isn't the same thing, and how the rules apply to it hasn't really been tested. But the principle carries over: don't leave people in silence, and don't hang up on them.

The problem with waiting for a verdict

Here's the trap. The more accurate classic AMD is, the longer it waits before deciding. While it waits, the person who answered hears nothing. They say "hello?" twice and hang up. Your call log records a two-second call, which also looks like spam to the analytics engines that label numbers (see STIR/SHAKEN and Spam Likely).

Newer approaches cut the wait. LiveKit's May 2026 write-up describes a detector that combines rules with a language model reading the transcript, with a median of 840 ms to a decision, and that "errs on the side of treating the call as human" when the opening is short or unclear. It reports a 97.3% F1 score for voicemail and 95.7% for humans on its own test set. Like every vendor figure, including the 97% to 99.7% claims you'll see in press releases, that's measured on the vendor's own data and isn't directly comparable to anyone else's.

What we learned in a real pilot

In October 2026 we ran a two-round outbound pilot for an online casino operator: 61 players, 121 dial attempts, $11.53 all in. Most attempts were unanswered or reached voicemail, which is why the average came to about 9.5¢ per attempt.

The lesson that changed the product: carrier voicemail detection mislabelled some voicemail greetings as humans. Short personal greetings, the kind that sound like someone picking up and saying their name, came back as "human". The AI then started its opening line into a mailbox.

Two other lessons came from the same calls. Waiting for the detector to finish left several seconds of silence, so real players heard nothing and hung up. And because identity had to be confirmed before the brand could be named, an opening line spoken to a voicemail had to be neutral enough that nothing sensitive was left behind.

How Dial Milo handles it now

Dial Milo uses two listeners rather than one gate.

1. Milo starts within a second. It begins its opening about 0.6 seconds after the call is answered, or sooner if it hears a short "hello", rather than waiting for the carrier verdict.
2. Carrier detection keeps running. If it comes back with "machine" after Milo has started, Milo stops.
3. Milo listens for voicemail too. If the AI recognises a voicemail greeting that carrier detection missed, it ends the call immediately without saying anything more.
4. The player is retried later. Voicemails and no-answers go back in the queue for the next attempt, within the calling hours and attempt limits you set. See calling hours and contact limits.

The opening itself does the rest. In engagement campaigns, Milo says it's an AI and asks for the person by first name before naming the brand, and it's worth writing your own scripts the same way. If that sentence lands on a voicemail, all the mailbox hears is an AI assistant asking for someone, with no brand or offer attached.

Call screening: the third kind of answer

Phones increasingly answer for their owners. iOS Live Voicemail and Google Pixel Call Screen pick up and transcribe, and newer iPhones can ask unknown callers for their name and reason before the phone rings. To a detector built for "person or machine", these are neither.

They also distort your numbers. Hiya noted in its January 2024 release that these features "show calls screened by voicemail as answered calls", which inflates answer rates. If your answer rate jumped and your conversations didn't, this is often why.

What an AI caller should do with a screener is simple to say and worth writing into your script: give a short, neutral reason and the caller's name, then wait. Don't deliver the pitch to an assistant. Something like "This is Milo, an AI assistant, calling for James about his account" gives the person enough to decide, without leaving your offer in a transcript. The cold calling script templates all use an opening that works this way.

Leave a message, or hang up?

Pew Research found 67% of US adults don't answer unknown numbers but do check any voicemail left, while 14% ignore voicemails too. So a message isn't wasted. Whether you should leave one depends on what it says:

If you do leave one, keep it under 20 seconds, name the business, say why you're calling in one line, and give a number that answers. With Dial Milo, the campaign number answers callbacks: a player who rings back is recognised and the conversation carries on. See when someone calls your campaign number back.

Worked example: what misclassification costs

Illustrative figures. Say a campaign makes 1,000 dial attempts. 450 aren't answered, 250 reach voicemail and 300 reach a person.

• If detection mislabels 10% of voicemails as human and nothing else is listening, 25 mailboxes get your opening line. If your opening names a brand or an offer, that's 25 recordings in places you didn't choose.
• If detection mislabels 3% of people as machines and hangs up, 9 real people get dead air. They're the ones most likely to report the number.
• With a second listener, the first group mostly gets a few seconds of a neutral opening before the AI hangs up. With an immediate start, the second group hears a voice instead of silence.

The cost of the extra seconds is small. Calls that don't connect have no minute charge, and a voicemail that Milo hangs up on lasts a few seconds, so it costs little. In the pilot above, that is how 121 attempts came to $11.53. You can price any destination on the rates page.

How to test voicemail detection yourself

1. 1

   Collect five kinds of greeting

   Ask colleagues to set different voicemail greetings: the carrier's default, a short personal one ("Hi, it's Sam"), a long business greeting, a full mailbox, and an iPhone with screening for unknown callers on.

2. 2

   Add them to a test list

   Put those numbers in a small campaign, with the real script, and run it during calling hours.

3. 3

   Listen to every recording

   Check what happened in each case: did the AI speak into the greeting, how long for, and what exactly did the mailbox hear?

4. 4

   Then answer some yourself

   Pick up and say "hello?" and nothing else, then pick up and stay silent for a second. Neither should be treated as a machine.

Ten minutes of this tells you more than any accuracy claim. For how voicemail handling fits with retries, pacing and budgets across a whole campaign, see AI outbound calling.