What AI actually changes in a parallel dialer (and what it doesn't)

Almost every dialer now has AI on the pricing page, and almost all of it is attached to things that happen after a conversation starts: summaries, sentiment, next-best-action, coaching. Those are real features. None of them affect how many conversations an agent has, which on a parallel dialer is the entire question.

Two things do, and both happen in the two seconds before the agent hears anything.

What does AI do in a parallel dialer?

A parallel dialer places several calls at once and has to answer two questions per batch, fast and without a human: is this a person or a machine, and how many lines should the next batch use. Both are classification problems on noisy signals, which is what machine learning is genuinely good at.

1. Answering machine detection. Decide, from the first fragment of audio, whether a live person picked up. Get it wrong one way and a prospect is hung up on; wrong the other and the agent listens to a voicemail greeting.
2. Line pacing. Decide how many numbers the next batch should contain, from the connect rate the campaign is producing right now rather than a number someone typed into a settings box last quarter.

How much is better answering machine detection worth?

Answering machine detection fails in two directions and they cost very different amounts. A false negative — a machine classified as a person — costs the agent about twelve seconds of a voicemail greeting. A false positive — a person classified as a machine — costs an entire conversation, because the dialer hangs up on someone who said hello.

Take the same agent as the rest of this cluster: 5% connect rate, four lines, 256 dials an hour, so about 12.8 people answer per hour. Apply a false-positive rate to that.

Moving from 8% to 1% false positives recovers 0.89 conversations per agent-hour — 143 a month for one SDR. That is roughly what adding a fifth line would produce, except this version does not raise the abandonment rate at all. On a ten-person team it is 1,430 conversations a month that were already paid for and thrown away.

Why line pacing is the part that actually needs a model

The compliant line count depends on the connect rate, and the connect rate moves — by list, by hour of day, by day of week, by how recently the data was sourced. A fixed setting is correct on exactly the day it was chosen.

abandon rate = (lines × c − (1 − (1 − c)^lines)) ÷ (lines × c)

at c = 0.05, 4 lines -> 7.3%   (breach)
at c = 0.02, 4 lines -> 3.0%   (compliant)

Those two rows are the same setting on the same dialer a fortnight apart, after the list improved. A pacing model watches the connect rate the campaign is producing, computes the line count that holds abandonment under target, and changes it — down, when things are going well. That is a control loop rather than intelligence, but it is the feature that keeps a parallel dialer inside the safe harbour, and almost nothing that calls itself an AI dialer does it.

How do you test the classifier yourself?

Vendors quote detection accuracy; nobody audits it. The standard advice — listen to a hundred calls the system marked as voicemail and count the humans — is well meant and statistically useless, and it is worth showing why, because the correct sample size is roughly four times larger.

It is also worth knowing why the false-positive rate is so rarely published. A blended accuracy figure is dominated by the easy cases — dead air, network tones, the long silence before a voicemail beep — which every classifier gets right, so blended accuracy sits in the mid-nineties for products that are not close in quality. Reporting the two error directions separately is what would separate them, and it is the one presentation that makes a weaker model look weak.

You are estimating a proportion of a few per cent. The 95% confidence interval on a proportion is ±1.96 × √(p(1−p)/n), so at an observed 5% false-positive rate a hundred calls gives you ±4.3 percentage points. That interval spans 0.7% to 9.3% — it contains both the best classifier on the market and the worst, which means the test cannot distinguish them and you have learned nothing.

n = (1.96 ÷ margin)² × p × (1 − p)

for p = 5%, margin = ±2 points:
n = (1.96 ÷ 0.02)² × 0.05 × 0.95 = 456 calls

Four hundred and fifty calls is an afternoon of listening, once, before a year-long contract. Sample from a single campaign and a single list rather than across the account, because detection accuracy varies with line quality and mixing sources hides the variance you are trying to measure.

What AI does not fix

Three things get attributed to AI in this category that it has no effect on:

• The connect rate itself. No model makes someone answer. Connect rate is a function of list quality, number reputation and time of day — see connect rate vs contact rate for why the two get confused.
• The abandonment ceiling. 3% is a regulatory limit, not a technical one. A better model reaches it more efficiently; it does not move it.
• Number reputation. Volume is what gets a number flagged, and a parallel dialer increases volume by design. A model that writes better scripts does not change what the carrier's analytics see.

The useful way to read a vendor's AI claims, then, is to sort them by where in the call they act. Anything before the greeting changes throughput and compliance. Anything after it changes the quality of a conversation the dialer had already won — worth having, but not what a parallel dialer is for, and not what should decide which one you buy.