Every predictive dialer demo shows the same slide: agents talking almost continuously, connect rates that look transformative, idle time near zero.
That demo is not dishonest. It is a real system on a good list, and it sets an expectation that the first week of production quietly demolishes — because the dialer is only one of the inputs, and usually not the dominant one.
What a predictive dialer actually does
It places more calls than you have agents, on the bet that most will not be answered by a human. When one is, an agent should be free to take it.
Everything good and everything painful about predictive dialing follows from that bet. Bet too conservatively and agents sit idle. Bet too aggressively and someone says hello to nobody — an abandoned call, which is both a bad experience and a regulated number.
The dialer is a controller trying to hold that balance while the thing it is controlling keeps moving.
What determines your numbers (mostly not the dialer)
Your list. By a distance the largest factor. Age, source, accuracy and whether these people have any reason to expect a call. A dialer cannot make a bad list good; it can only work through it faster.
Time of day and day of week. The same list produces materially different contact rates at 10am Tuesday and 4pm Friday.
Number presentation. Calls arriving from an out-of-region number get answered less. Local presence is one of the larger single levers on connect rate, and it is not a dialer setting.
Number reputation. Carriers and handsets label numbers. Rotating a small pool hard is how you burn it, and once burned your connect rate drops for reasons no dialer configuration will fix.
Agent count. Prediction gets more accurate with more agents, because it is a statistical bet. With eight agents, variance dominates. With eighty, the maths works. Small floors should expect worse pacing than the brochure regardless of platform.
Then the dialer. Pacing algorithm and detection tuning. Real, and worth getting right — but it is the last factor, not the first.
Metrics that are worth watching
Contact rate — connects per attempt. Mostly a property of your list and your number presentation.
Agent talk time — the fraction of a paid hour actually spent in conversation. This is the number the dialer most directly moves, and the honest one to judge it by.
Abandon rate — must stay under the regulated threshold, and is probably measuring two different things right now.
Attempts per connect — the composite. If this rises while everything else looks stable, something upstream changed: list quality, labelling, or a carrier route.
Cost per conversation — not cost per minute. Minutes are an input; conversations are what you are buying.
Things that will surprise you
Prediction needs a warm-up. The algorithm is estimating agent availability from recent behaviour. After a shift change, a campaign switch or a break, it is estimating from thin data and will pace badly for a few minutes. This is normal and is not a fault.
One global ratio is always wrong. Different campaigns have different answer profiles and different handle times. A single pacing setting across all of them is a compromise that suits none. Pace per campaign.
Answering-machine detection is a trade-off, not a feature. Tuned to catch every machine, it will drop some humans. Tuned to never drop a human, it will pass machines to agents. There is no setting that does both, and any vendor who implies otherwise has not shown you their false-positive rate.
Idle time is not free to eliminate. Driving agent idle to zero means dialing aggressively enough that abandons rise. The right target is not zero.
Your best hour is not your average hour. Plan capacity against peak, not mean. The fifteen minutes when everything happens at once is what your customers experience.
What good looks like
Not a number I can give you — anyone who quotes a contact rate without seeing your list is quoting their best case.
What good looks like is diagnosable: you can say why this week differs from last, split abandons from detection errors, see pacing per campaign rather than in aggregate, and watch attempts-per-connect closely enough to notice a carrier degrading before it costs a month.
A dialer you can reason about will beat a marginally better algorithm you cannot. That is the whole recommendation.
If your numbers are worse than expected
Work in this order — it is roughly the order of effect size:
- The list. Where did it come from, how old, how accurate?
- Number presentation. Are you calling into markets from out-of-region numbers?
- Number reputation. Are your numbers being labelled? Have you checked, rather than assumed?
- Detection tuning. Pull fifty calls dropped as machines and listen to them. Count the humans.
- Pacing. Per campaign, against measured agent-availability curves.
- The dialer itself. Last, and least often the answer.
Almost everyone starts at six and works upward. It is the most expensive possible order.