Abandoned call rate and AMD: how detection affects the 3% limit

What the abandoned call rules say

Two numbers matter for US outbound calling under the Telemarketing Sales Rule and FCC rules:

• The 3% limit. Abandoned calls must not exceed 3% of answered calls, per campaign, measured over 30 days.
• The 2-second rule. A call is abandoned if the person is not connected to a live agent within 2 seconds of their completed greeting.

The UK's Ofcom has similar rules for silent and abandoned calls. Other countries have their own versions.

The second rule is the one most teams underestimate. "Abandoned" does not only mean the dialer hung up. A call where the person says "Hello?" and then waits in silence for four seconds before an agent arrives can count too. That is where answering machine detection comes in.

How the 2-second clock interacts with AMD

Look at a typical answered call on a predictive dialer with AMD on:

1. The call is answered.
2. The person says "Hello?" and stops. Their greeting is complete.
3. AMD is still listening, because it has not decided yet.
4. AMD decides "human" and the dialer sends the call to a free agent.
5. The agent hears the person and starts talking.

The 2-second window starts at step 2. Everything after it, including the rest of the AMD analysis and the hand-off to the agent, uses up that window. So the question is not just "how long does AMD take?" but "how long does AMD keep listening after the person has stopped talking?"

Take a hypothetical example. A person answers and says "Hello?" within the first second. If AMD keeps listening until the 5-second mark, that person has waited around 4 seconds in silence before an agent can even be connected. If AMD decides at the 2-second mark, the wait is around 1 second plus the hand-off. Same person, same agent, very different exposure.

Three ways AMD drives abandoned and silent calls

1. AMD time

Asterisk's stock AMD(), which VICIdial uses by default, decides from silence and word timing. Its shipped defaults allow a total analysis time of 5000 ms, with an initial silence allowance of 2500 ms. When the answer is ambiguous, AMD can return NOTSURE after using all of that time. A quiet "Hello?" followed by waiting is exactly the pattern that sends timing-based AMD to its limits, and it is also exactly how many real people answer an unknown number.

2. False machine verdicts

When AMD decides a live person is a machine, VICIdial hangs up, or plays your voicemail message to them through "AMD send to VM exten". To the person, that is either a silent call that ends or a recording that makes no sense. Either way, a live person answered and was never connected to an agent.

Here is the catch: the dialer does not know it happened. It logs the call as an answering machine (AA in VICIdial), not as a drop. Your reported drop rate can look fine while live people are being hung up on. We cover this in more detail in our post on VICIdial dead air and ghost calls.

3. Missed machines

The opposite error matters too. Every voicemail greeting AMD misses goes to an agent, who spends time listening to it and dispositioning it. While they do, they are not free. A predictive dialer that has dialed for the expected number of free agents now has fewer than it expected, and the next live answer waits in the queue or is dropped. Poor machine detection quietly raises the drop rate even when it never hangs up on a person.

How to measure your abandoned call rate in VICIdial

Today's rate

The Real-Time Main Report shows, per campaign, the calls dropped and answered so far today and the dropped percentage. This is the number most supervisors watch, and it is what VICIdial's adaptive dial methods steer toward.

The 30-day rate

The rule is measured over 30 days per campaign, so a single good day proves nothing. Run your status reports for each campaign over a rolling 30-day date range and calculate:

abandoned call rate = DROP calls / calls answered by a person

How the denominator is defined is a question for counsel, so agree the definition once and use it consistently.

To make the limit concrete, take a hypothetical campaign with 20,000 calls answered by people over 30 days. At 3%, it can afford at most 600 abandoned calls in that period. If its reported drops are 450, it looks comfortable at 2.25%. But if AMD also hung up on, say, 300 live people that were logged as machines, the real number of people who answered and never reached an agent is 750, which is 3.75%. The example numbers are invented for illustration; the point is that the reported rate only counts what the dialer knows about.

The hidden part

Neither report shows live people your AMD marked as machines. To estimate those:

1. Pull a random sample of calls with the AA status from one campaign, ideally a few hundred.
2. Listen to each recording and note whether a person answered.
3. The share of live people in that sample, multiplied by your AA volume, estimates how many people AMD hung up on.

If you use nomoVM, the dashboard keeps the audio it heard for each call, so you can play back MACHINE verdicts directly and export the log as CSV. Our guide to AMD accuracy explains why you should measure both error types.

Practical steps to stay under 3%

1. Set a drop limit with a margin

   On an adaptive dial method, VICIdial's Drop Percentage Limit caps how aggressively the dialer dials based on its drop rate. Set it below 3% so normal variation does not push you over. Exactly how much margin is a business and legal decision.

2. Watch the dial level, not just the limit

   Adaptive dialing reacts to drops after they happen. If a campaign keeps bumping against the limit, lower the maximum dial level, especially at the start and end of shifts when agent counts change quickly.

3. Shorten AMD time

   Every second AMD listens after a person's greeting comes out of the 2-second window. Review your amd.conf or AMD() arguments, and check how often calls end as NOTSURE.

4. Stop hanging up on people

   Sample your AA calls as described above. If a meaningful share are live people, tuning timing values only moves the errors around. Detection that listens to what the answer sounds like is the real fix.

5. Catch more machines

   Fewer voicemails reaching agents means more free agents for the next live answer, which lowers drops at the same dial level.

6. Handle dropped calls deliberately

   VICIdial lets you choose what happens to a call that waits too long for an agent, including how many seconds it waits and whether a message plays. Ask counsel what your dropped-call handling must include.

7. Check the trunks

   Slow answer signalling or one-way audio adds silence before AMD even starts. If agents report "dead" connections, look at the carrier side as well as AMD.

How nomoVM helps

nomoVM replaces timing rules with a model that listens to how the call is answered:

• Faster decisions. Most calls are decided about 2 seconds after answer. Only unclear calls listen up to 3.2 seconds. There is no long NOTSURE wait.
• Fewer people hung up on. In our measured test of 400 real outbound calls checked by a person, 94.9% of live people were sent to agents, and 88% of calls sent to agents were live people. We publish both numbers so you can compare.
• Safe when something fails. If nomoVM cannot judge a call, it returns FALLBACK before using any audio and your dialplan runs Asterisk's AMD() as before.
• Same VICIdial logic. It sets the same AMDSTATUS that VD_amd.agi reads, so your drop limit, dial method and voicemail settings stay as they are.

It runs on the dialer you already have: one script, one new extension and one campaign setting, in about 10 minutes, with no new servers and no carrier change. Plans start free at 30,000 calls a month; see pricing.

Want to see the effect on your own numbers? Start free, switch one campaign using the VICIdial AMD setup guide, and compare its drop rate and agent-connected calls with the 30 days before. The dead air calls page covers the related problem of silence on connected calls.