← All articles
Product UpdatesNov 16, 2026 12 min read

The Fine Print in 'Abandonment Rate' Every Owner Should Read Once

The 3% abandonment cap explained in plain English: what counts, what does not, and how dialer misclassification quietly writes abandonments your compliance report never sees.

The Fine Print in "Abandonment Rate" Every Owner Should Read Once

Abandonment rate is the number on every predictive dialer compliance dashboard, and almost every owner I talk to has read the rule as a one-liner: keep it under 3%. That is roughly right and dangerously incomplete. The fine print is about what counts as an abandonment, and the fine print is where dialer misclassification lives. A classifier that drops live humans as machines is not just a revenue leak. It is writing abandonment events that your denominator logic may be hiding from you.

I run answering machine detection for outbound shops on ViciDial, Asterisk and FreeSWITCH, and compliance questions come to me second, always after someone's carrier or legal counsel asked for numbers the shop could not cleanly produce. Here is the rule in plain English, what actually counts, and the two ways misclassification corrupts the metric.

The rule, in plain English

Under the FCC's rules implementing the Telemarketing Sales Rule, when a predictive dialer reaches a live person, that person must be connected to a sales representative within two seconds of the human's completed greeting. If the call is dropped instead, it is an abandonment. Abandonments may not exceed 3% of calls answered by live persons, measured per campaign per calendar day.

The FCC's consumer material on unwanted robocalls gives the plain-language version: https://www.fcc.gov/consumers/guides/stop-unwanted-robocalls-and-texts

Four words in that paragraph carry the entire risk for dialer shops.

"Answered by live persons"

The denominator is live answers, not dials, not connects, not all answers. Machines and carriers announcements do not count in the denominator. Only humans. Which means the number that governs your legal exposure is produced by your AMD. The compliance layer sits directly on top of the classification layer, and the classification layer is the one component that shipped free, unaudited, with the dialer.

"Within two seconds of the completed greeting"

The clock does not start at answer. It starts when the human finishes saying hello. That grace period is exactly the window answer detection operates in, and it is why detection latency has a compliance dimension, not just a productivity one. A verdict that arrives in about an eighth of a second leaves the dialer plenty of runway to make the two-second window. A verdict that arrives late eats the runway and turns good calls into abandons. We covered the latency mechanics in latency versus accuracy in AMD.

"Per campaign per day"

A 2.9% average across the account can hide a 6% Tuesday on one campaign. The cap is measured where you dial, not where you aggregate. If your reporting only shows account-level abandonment, you are looking at an average of numbers that are each individually capable of being a violation.

What counts and what does not

Event Counts as abandonment? Why
Live human answers, no agent available, call drops Yes The textbook abandonment
Live human answers, slow transfer, human hangs up after 3+ seconds of silence Yes Missed the two-second window
Human answers, AMD says machine, call drops Yes, functionally A dropped live answer looks identical from outside
Answering machine reached, message left or dropped No Not a live person
Carrier announcement or tone No Not a live person
Google Call Screen or iOS screening bot answers Gray zone Bot answered, not a person; misrouting it to agents wastes the call
Busy, no-answer, congested No Never answered

The third row is the one to sit with. If your AMD misreads a human as a machine, the call drops, and from the outside, from the regulator's or plaintiff counsel's chair, that event is indistinguishable from a dialer that reached a human and abandoned them. Your internal report may log it as a machine drop, which removes it from both your numerator and your denominator. That is the corruption, and it cuts both ways.

The two corruptions

Corruption one: hidden abandonments

A misread human logged as a machine never enters your abandonment count. Your dashboard says 1.8%. The true number, counting the humans you dropped as machines, is higher, potentially much higher at the top of the misclassification range. Stock dialer AMD misclassifies an estimated 10 to 20% of live humans. On a campaign where machines legitimately dominate the answers, a small absolute number of misread humans can move the true rate past the cap while the reported rate stays comfortably clean. You cannot defend yourself against a number you do not know you have.

Corruption two: a denominator you cannot defend

Flip the failure mode. If your AMD overcounts humans, classifying machines as live answers, your denominator inflates and your reported abandonment rate looks better than reality. That is not a gift. In a dispute, the plaintiff's expert re-scores your logs with real classification, the denominator shrinks, the rate rises, and your dashboard's historical numbers stop matching the recalculated ones. An inflated denominator is an audit finding waiting to be discovered, and it is discovered at the worst possible moment.

Either corruption exists because the classification layer is unauditable or wrong. The fix is the same in both directions: verdicts that are correct, timestamped, and stored per call.

What a defensible position looks like

Compliance defense is fundamentally a record-keeping question: when a dispute arrives, can you show, call by call, why each person was or was not connected, and can an outsider verify the classification that decision rested on? We built the full argument in how AMD logs become your compliance defense. The short version of the standard:

  1. Every answered call carries a stored verdict: human, machine, screening bot, carrier announcement.
  2. Every verdict carries a timestamp and the decision it drove: connect, drop, message.
  3. The verdicts come from a system that demonstrably works, with accuracy you can show on your own traffic, not a vendor brochure number.

With those three in place, the abandonment rate stops being a dashboard assertion and becomes a derivable fact. Anyone re-scoring your logs gets the same number you reported. That property, more than any threshold, is what keeps a routine inquiry from becoming a case. The regulatory environment around outbound calling keeps tightening, most recently around cell phones and consent, which we covered in the TCPA cell phone ruling and what it means for your dialer, and tighter scrutiny makes derivable records worth more every quarter.

The uncomfortable audit you should run once

Pull yesterday's machine-classified calls. Sample 50 at random. Listen to the first six seconds of each. Count the human greetings. That count, times your daily machine volume, is the number of abandonment-adjacent events your reports are not seeing. If the answer is zero, you have a clean classification layer and real peace of mind. If the answer is five, you have a compliance exposure wearing a productivity costume.

The stakes scale with you, too. Growth multiplies dials, answers and misclassifications together, which is why compliance risk grows exponentially rather than linearly with volume, an argument we made in why growth makes compliance risk exponential. The shop that audits this at 20 seats inherits a very different exposure at 120.

The two-second window, in practice

The grace period deserves its own walkthrough, because it is where dialer engineering and the rule meet, and where most shops discover their configuration has quietly been at odds with their obligations.

What the dialer is doing during those seconds

A human answers. Audio starts flowing toward the dialer. The AMD layer is analyzing the opening audio: energy, timing, the shape of the greeting. Meanwhile the pacing engine is holding a pool of agents and deciding which one to attach. The rule says the human must hear a representative within two seconds of finishing their greeting. Every millisecond the verdict takes is a millisecond subtracted from the attach window.

Now consider what stock timing-based AMD actually does in that window. It waits. It needs enough audio to make a timing decision, often a second or more of it, and if the greeting is short or unusual, it waits longer. A verdict that lands in about an eighth of a second, as GPU-backed detection does, hands the dialer nearly the entire legal window to work with. A verdict that lands at 1.5 seconds does not just risk the sale, it risks the rule.

The screening bot wrinkle

Google Call Screen and iOS Call Screening add a wrinkle the rulebook did not anticipate: the call is answered by a synthesized voice, which is not a live person, so a strict reading puts it outside the abandonment definition entirely. But the dialer has to know that is what happened. A classifier with no screening-bot category either parks the bot on an agent, burning the connect, or drops it ambiguously, polluting your logs with events an examiner cannot categorize. Distinct verdicts for screening bots are a compliance feature, not a nicety, and we broke the detection side down in Google Call Screen detection.

The measurement discipline

Because the cap is per campaign per day, the operational requirement is boring and non-negotiable: compute it per campaign per day, every day, and alert on the campaigns, not the account. The shops that sleep well run a standing report with the abandonment figure derived from stored per-call verdicts, so the number on the wall matches the number the logs produce. The shops that get surprised run account-level averages over a classification layer nobody has tested.

When the examiner asks: a walkthrough

Picture the scenario the fine print exists for. A complaint lands, or a carrier flags your traffic, and someone asks for documentation. Here is the difference between the two postures.

The unprepared posture

Counsel asks for abandonment figures by campaign by day. Someone opens the dialer admin and finds account-level averages. Counsel asks how the dialer decided which calls were live humans. The answer is a settings screen last touched years ago. Counsel asks to verify the classification on a sample of calls. There is no stored verdict audio, no per-call decision log, just a disposition code. Every question surfaces a gap, and the gaps read as negligence, which is the word that turns inquiries into penalties.

The prepared posture

Same questions. The answer is a per-call log: timestamp, verdict, verdict confidence, the routing decision each verdict drove, and the connect latency for every human-classified call. The examiner samples 50, re-scores them, and gets your number. The abandonment report matches the underlying records across every campaign and day. Nothing to explain, so nothing escalates. The difference between the two postures is not staff, budget, or counsel. It is whether the classification layer wrote down what it did while it was doing it.

This is the same argument we made in how AMD logs become your compliance defense, and it is worth repeating in the abandonment context specifically, because abandonment is the figure regulators and carriers actually pull first. It is the headline number of outbound compliance. Build the record before anyone asks for it.

The questions compliance-minded owners ask

"Does voicemail count toward the 3% at all?"

No. The denominator is live persons, and the rule's purpose is protecting humans from silence, so machine answers are excluded from both sides of the fraction. But notice what that exclusion costs you procedurally: the entire metric hinges on correctly separating humans from machines, call by call, and the burden of proof for that separation sits on your logs. An examiner who doubts your classifier doubts your denominator, and a doubted denominator unravels the whole calculation. The exclusion is where the compliance value of accurate detection is concentrated.

"What about wrong numbers and DNC scrubs?"

A wrong-number answer is still a live person answering. If the dialer drops it, it counts like any other abandonment, regardless of how categorically undeserving the number was. The clean handles are upstream: DNC and litigator scrubbing before dialing, and dedupe against your opt-outs, so the call never happens. Once the phone rings and a human speaks, the rule treats all humans equally. There is no intent discount for numbers you wish you had not dialed, which is why list hygiene is a compliance control, not just a courtesy.

"Our abandonment dashboard says under 2%. Why worry?"

Because that number is only as good as the classification feeding it, and you have presumably never audited that classification, because almost nobody does. The dashboard does not measure abandonments. It measures how often your dialer labeled an event an abandonment, given a classifier you have not tested since install day. Those are the same number only if the classifier is right. The sample-and-listen audit from the previous section is how you find out, and it is deliberately cheap: an afternoon, 50 calls, a pair of headphones, and the nerve to count what you hear.

"If we fix detection, does our reported rate go up?"

Possibly, and that is a feature. If your stock AMD has been overcounting humans, machines logged as live answers inflated your denominator and flattered your rate. Accurate detection shrinks the denominator to truth and the reported number may rise toward honesty. That momentary honesty feels like a regression and is the opposite: it is your compliance posture aligning with reality before an examiner or plaintiff aligns it for you, at a time of their choosing, with their math. The only rate worth having is one you can defend call by call, and every other number is a liability with good posture.

"Where do screening bots actually land in the books?"

Treat them as their own category with their own disposition, not as humans and not as machines. Google Call Screen and iOS Call Screening answers are synthesized voices, neither a live person for the abandonment denominator nor a machine for message-leaving purposes. Logging them distinctly keeps both sides of your fraction clean and gives marketing the real number on how much screening traffic each campaign draws, which is becoming a list-quality signal in its own right.

Practical closing

The 3% number is not the hard part. The hard part is that both your numerator and your denominator are downstream of a classifier nobody has tested since installation day. Test it once, honestly, against your own traffic. Detection that returns verdicts in about an eighth of a second, identifies screening bots distinctly, and logs every decision costs less than one consultation hour with regulatory counsel: free Sandbox tier at 50,000 detections a month, paid plans starting at $79 a month with 500,000 included.

Read your fine print before someone else reads it for you.