AMD: Answering Machine Detection Trade-offs Every Dialer Operator Should Know

What AMD Actually Does

AMD (also called answering machine detection or AMD/CPA — Call Progress Analysis) analyzes the audio of an answered call in the first 1.5–3.0 seconds after connection. It listens for patterns: the length of the initial speech burst, silence periods, the presence of a beep tone, and the cadence of automated greeting phrases.

Based on that analysis, the system classifies the call as one of three outcomes:

• Human answer (live person): Connect to an agent immediately.
• Machine answer (voicemail detected): Either hang up or play a pre-recorded voicemail drop message.
• Uncertain / undetermined: Apply a fallback rule (usually treated as human, sometimes as machine).

The detection window is the entire problem. AMD needs 1.5–3.0 seconds of audio to make its determination. During that window, a live human who picked up the phone is sitting in silence, hearing nothing. That silence is the price of AMD.

The False Positive Problem

A false positive in AMD means the system classified a live human as a voicemail system. The agent never connects. The contact hears silence, possibly followed by a hang-up or a voicemail drop message. From their perspective, someone called them and immediately hung up or left a nonsensical voicemail.

False positive rates on modern AMD implementations range from 2–8% depending on the contact list, the carrier path, and the platform's detection algorithms. On a campaign that connects 500 humans per day, a 5% false positive rate is 25 people per day who were identified as voicemail boxes and never received an agent.

On a 30-day campaign, that is 750 contacts your team dialed, paid the SIP cost for, and then silently alienated without ever attempting the conversation. On high-value lists, the false positive cost is not the connection cost — it is the relationship damage.

The Detection Delay Problem

Every AMD implementation adds latency to the connection event. Even at best-case performance (1.5 seconds), there is a perceptible pause between the contact picking up the phone and hearing a voice on the other end.

A 1.5-second pause does not sound long when described in a vendor spec. On a phone call, it is the difference between a natural greeting and a call that sounds like a robocall connecting. Studies of call center connect rates consistently show that calls with more than 1.0 seconds of silence after the contact answers have meaningfully higher immediate hang-up rates than calls where an agent voice is present within 0.5 seconds.

AMD delay is an inherent trade-off. You cannot accurately detect a machine in 0.3 seconds. If you want detection, you accept the delay and the corresponding increase in immediate hang-ups from live contacts who assume it is an automated system.

When AMD Helps and When It Hurts

AMD is net positive in specific scenarios:

High voicemail rate lists. If your list is reaching voicemail 55–65% of the time — common on cold B2C lists during business hours — AMD prevents agents from sitting through 60-second voicemail greetings waiting to speak. The agent capacity saved is real.

Automated voicemail drop campaigns. If your strategy is to leave a pre-recorded message on voicemail and move on, AMD enables that at scale. The quality of the voicemail drop and the call-back rate it generates are a different question, but AMD is the correct tool for the job.

AMD is net negative in:

High-value B2B lists. The false positive rate on executive contacts — who often have PA-screened calls, have mobile voicemail systems that answer quickly with an unusual greeting, or have a slow pickup that AMD initially reads as a machine — is higher than on consumer lists. The cost of a false positive on a CFO prospect is not recoverable.

Consumer morning campaigns. Mobile voicemail systems have become increasingly varied in their greeting patterns. Modern carrier voicemail sounds less uniform than traditional PSTN voicemail. AMD tuned for PSTN patterns performs worse on mobile-heavy lists.

Regulatory environments with strict silence-on-answer rules. Some jurisdictions treat excessive silence on connection as a form of abandoned call regardless of whether it was caused by AMD. This is a legal and compliance question, not a technical one — but it is worth confirming your AMD configuration is within bounds for the jurisdictions you are dialing.

Tuning AMD Sensitivity

Most AMD implementations offer a sensitivity parameter: conservative (lower false positive rate, higher false negative rate — more live calls treated as unknown) vs aggressive (higher false positive rate, lower false negative rate — more machines correctly identified).

For B2B campaigns: run conservative sensitivity. The cost of a false positive is higher than the cost of an agent briefly hearing a voicemail greeting before hanging up.

For high-volume B2C campaigns on lists with known high voicemail rates: test your specific list with the default setting first. Measure false positives by reviewing call recordings for cases where AMD triggered a machine classification but the audio shows a live human. Adjust accordingly.

The Voicemail Drop Calculation

If you are running AMD specifically to drop voicemail messages, the economics depend on your call-back rate. A voicemail drop that generates a 3% callback on 500 voicemail reaches per day produces 15 inbound calls. If your agents can close 20% of those callbacks, you are generating 3 closes per day from voicemail drops.

Whether that justifies the AMD implementation, the voicemail drop production cost, and the false positive write-off depends entirely on your deal value and close rate. For campaigns where the answer is yes, the auto-dialer network supports AMD integration on outbound campaigns.

What STIR/SHAKEN Has to Do with AMD Performance

Calls that arrive without proper caller ID attestation are increasingly labeled "Spam Likely" by mobile carriers. A contact who sees "Spam Likely" on their screen and lets it go to voicemail creates an AMD classification event where the voicemail answers almost immediately. AMD may correctly classify this as a machine — it is — but the underlying problem is caller ID attestation, not the contact's availability.

On UnlimCall's US and Canada network, STIR/SHAKEN signing is handled at the network level. Improving your attestation status reduces the rate of contacts letting calls go to voicemail before answering, which in turn reduces AMD voicemail detection events and improves live contact rates.