A predictive dialer places more outbound calls than it has agents available, using a statistical model of answer rate, ring time and call duration to predict when an agent will become free, so that a live human is already on the line by the moment the agent hangs up the previous call.
Everything else that appears on a dialer feature grid - CRM integration, recordings, dashboards - is shared with every other dialer category. The prediction is the only thing that makes it predictive, and the prediction is also the only thing that can hurt you, because its unavoidable by-product is a call that a stranger answers with nobody behind it.
The four dialer types, side by side
"Auto dialer" is the umbrella term for any software that dials without a person pressing keys. Four types sit underneath it, and they differ on exactly one axis: how many lines are open per available agent.
| Lines per agent | Agent hears | Abandoned calls | Compliance burden | Best fit | |
|---|---|---|---|---|---|
| Manual | 1, dialed by hand | Every ring and voicemail | None | Lowest | Low volume, high consent sensitivity |
| Preview | 1, after the agent reads the record | Every ring and voicemail | None | Low | Complex or high-value records |
| Power / progressive | 1 per agent, dialed automatically | Ring and voicemail, but no dialing time | None | Low | Small teams, warm and mid-warm lists |
| Predictive | Several per expected free agent | Connected conversations, mostly | Yes, by design | Highest: rate cap, disclosure message, record-keeping | Large agent pools on low-answer-rate lists |
Two things fall out of the table. Every predictive dialer is an auto dialer, but most auto dialers are not predictive. And the abandoned-call column is the only one that carries a statutory cap, which is why the rest of this article spends more time on pacing and law than on features.
How the pacing algorithm actually works
Strip away the vendor language and a predictive dialer is running a small queueing model, recalculated continuously.
- Measure the answer rate. Over the recent window of the campaign, what fraction of dialed numbers reach a live human? On a cold consumer list this can sit in the low single digits. On a warm inbound list it can be far higher. The dialer tracks this per campaign, sometimes per list segment, sometimes per time of day.
- Measure ring-to-answer time. How long between dial and pickup, distributed, not averaged. This sets how far ahead of an agent freeing up the dialer must start dialing.
- Measure talk time and wrap time. How long agents stay on a connected call, and how long they need after it before they can take the next one.
- Count available agents. Not logged-in agents. Agents who will be free within the ring window.
- Compute the pacing ratio. Divide expected free agents by expected connect probability to get the number of simultaneous lines to open. Naively, a 5 percent answer rate and one agent about to free up implies roughly twenty lines. No dialer actually opens twenty, because that figure only balances the expected number of answers against one agent, and any answer beyond the first is an abandoned call. The binding constraint is therefore the abandonment ceiling, not the answer rate, and the ratio a dialer will really run is whatever keeps abandonment under that ceiling given the size of the agent pool.
- Open the lines, then race. Whichever call is answered first gets the agent.
- Handle the losers. Every other answered call now has a human saying "hello" with no agent behind it. This is the abandoned call, and it is not an edge case. It is a designed output of the algorithm.
- Feed the result back. The outcome of every call updates the answer rate, and the ratio moves.
When the loop is tuned well, agents talk almost continuously. When it is tuned badly, either agents sit idle or strangers pick up the phone to silence.
What an abandoned call is, in law rather than in marketing
Two separate US rules govern this, and they do not say the same thing.
The FCC rule, at 47 CFR 64.1200(a)(7), prohibits abandoning more than three percent of all telemarketing calls that are answered live by a person, measured over a 30-day period for a single calling campaign. If the campaign runs longer, the rate is calculated separately for each successive 30-day period. The rule then defines the term precisely: "A call is 'abandoned' if it is not connected to a live sales representative within two (2) seconds of the called person's completed greeting." The same section separately requires that an unanswered telemarketing call is not disconnected before at least 15 seconds or four rings, and requires the seller to maintain records establishing compliance (47 CFR 64.1200).
The FTC Telemarketing Sales Rule, at 16 CFR 310.4(b)(1)(iv), flatly prohibits "abandoning any outbound telephone call", using the identical two-second definition. Liability is then avoided through a safe harbour at 310.4(b)(4), which requires four things together: technology that ensures abandonment of no more than three percent of all calls answered by a person, measured over the campaign or each successive 30-day period; ringing for at least fifteen seconds or four rings before disconnecting; promptly playing a recorded message naming the seller and its phone number whenever no representative is available within two seconds; and retaining records establishing compliance with all three (16 CFR 310.4).
Read those together and three practical facts fall out.
The three percent is a ceiling on a rate you cannot directly set. You set the pacing ratio. The abandonment rate is an emergent property of that ratio meeting a list whose answer behaviour you measured in the past and are assuming will hold. It drifts.
The two seconds is measured from the end of the called person's greeting, not from pickup. Answering machine detection sits inside that window. So does any network delay. The budget is smaller than it sounds.
The record-keeping is not optional. Both rules require you to be able to prove the rate. That means campaign-level call logs, retained and queryable.
On exposure, be careful about which figure applies. The statute behind the FCC rules provides a private right of action to recover actual monetary loss or $500 in damages for each violation, whichever is greater, with trebling available for willful or knowing violations (47 U.S.C. 227). Whether that private right of action reaches the abandoned-call rule specifically is not settled: 47 CFR 64.1200(a)(7) was adopted under the telemarketing-practices provisions at 227(c) and 227(d), and courts have divided on whether a private plaintiff can sue over it at all, as against the well-established private claims for autodialed or prerecorded calls. Treat the $500 figure as the ceiling of the general TCPA regime rather than a settled per-abandoned-call price, and take advice on your own programme. Enforcement by the FCC and FTC, and state attorneys general, does not depend on that question. The FTC also publishes a plain-language compliance guide to the Telemarketing Sales Rule that is worth reading before any outbound programme (FTC Telemarketing Sales Rule guidance).
None of this makes predictive dialing illegal. It makes it a regulated, measured, auditable activity with a compliance function attached.
Answering machine detection is a guess with a timer
Vendors present AMD as a checkbox. It is a probabilistic audio classifier operating under a deadline, and its parameters tell you more about real dialer behaviour than any feature grid.
Twilio, whose Programmable Voice API is a common building block for outbound dialers and is what Inflowave's own dialer runs on, documents the mechanics openly. AMD runs in one of two modes. Enable returns a verdict as soon as it identifies the called party, with possible results of machine_start, human, fax or unknown. DetectMessageEnd waits for the end of a voicemail greeting instead, returning machine_end_beep, machine_end_silence, machine_end_other, human, fax or unknown. Four tuning parameters control the trade-off: MachineDetectionTimeout, allowed between 3 and 59 seconds with a default of 30; MachineDetectionSpeechThreshold, between 1000 and 6000 milliseconds, default 2400; MachineDetectionSpeechEndThreshold, between 500 and 5000, default 1200; and MachineDetectionSilenceTimeout, between 2000 and 10000, default 5000. Twilio notes plainly that the engine may interpret a very short two-second voicemail greeting as a human picking up, and that when the timeout is reached, AnsweredBy returns unknown (Twilio answering machine detection).
Sit with the default numbers. The speech threshold alone is 2400 milliseconds. The regulatory budget between a completed greeting and a connected representative is 2000 milliseconds. Detection that waits long enough to be confident has already spent the compliance budget; detection fast enough to stay inside it is guessing on less audio. That tension is not a vendor defect. It is the shape of the problem, and every predictive dialer on your shortlist is somewhere on that curve whether or not its marketing says so.
The arithmetic of when predictive pays
The honest case for predictive dialing is idle time, and it scales with agent count for a reason worth spelling out rather than asserting.
Take a cold list with a 5 percent answer rate. A power dialer means an agent works through roughly nineteen unanswered or machine-answered calls to reach one conversation. Predictive dialing collapses that wait by opening lines ahead of the agent.
The catch is variance. The number of those open lines that get answered is a random count, not a fixed one. With one agent and twenty lines open, the expected number of answers is one, but two or more answers happen a meaningful share of the time, and every extra answer is an abandoned call. Increase the agent pool and the same relative variance shrinks against a larger expected value, so a pool can absorb the overshoot that a single agent cannot. That is the whole mechanism: pooling, not software quality.
Two consequences follow, and neither requires a threshold number nobody has published:
- In a small pool, the only ways to hold abandonment near zero are to pace close to 1:1, at which point the dialer behaves like a power dialer with extra connect latency, or to accept abandoned calls and meter them. There is no third option, because the abandonment rate falls out of the ratio and the pool size together.
- In a large pool, aggressive pacing and a low abandonment rate stop being in tension, which is why predictive dialing is standard in contact centres and unusual in small sales teams.
If you want a threshold for your own operation, it is measurable rather than universal: run your list, record connects per hour and abandonment per hour at the pacing ratios your vendor allows, and find the point where the pacing ratio you can actually sustain stops being close to 1:1. That number depends on your answer rate, your talk time and your agent count, so a published figure from someone else's campaign would not transfer anyway.
Who is actually selling you what
Pricing below was checked against each vendor's own live pricing page on 15 September 2026. Pricing in this category moves, so re-check before you buy.
Convoso is a pure-play outbound predictive dialer vendor, built for high-volume outbound call centres. It does not publish pricing. Its own pricing page explains that it tailors quotes to seats, features and integrations rather than offering a one-size-fits-all model, states that it "primarily offers annual plans to deliver better value and ensure long-term success", and adds that "We do offer 20% off annual plans currently" (Convoso pricing, checked 15 September 2026). If you are a five-seat agency, that is a signal about who the product is designed for, not a criticism of the product.
Five9 publishes per-seat pricing, which is rarer here and genuinely useful. Its Digital tier is listed at $119 monthly per seat and its Core tier at $159 monthly per seat, both billed monthly; the Plus, Pro and Enterprise tiers are quote-only. Read the footnote on the same page before you do the arithmetic: "Prices above are per concurrent user and usage-based pricing may apply, with a minimum of 50 seats." That minimum puts the real entry point at roughly $5,950 per month on Digital, not $119 (Five9 pricing, checked 15 September 2026). Five9 is a full contact centre platform, including inbound routing, workforce management and quality management. If you run a contact centre, it is a serious answer. If you run a small agency, you are buying a workforce management suite to get a dial button.
Genesys and NICE are in the same tier and the same shape. They are excellent at what they do and they are not scoped for a ten-person agency.
PhoneBurner is the interesting one, because it is a power dialer that positions explicitly against predictive dialing. Its published plans are Standard at $165 per user per month billed monthly, or $140 per user per month billed annually; Professional at $195 billed monthly, or $165 billed annually; and Premium at $215 billed monthly, or $183 billed annually (PhoneBurner pricing, checked 15 September 2026). Every tier lists "Delay-free connections" as a call quality line item, and that phrase is the whole pitch: a power dialer has no connection delay because there is nothing to predict. PhoneBurner is also better than Inflowave at one thing worth naming, which is that call recording retention is a published, tiered commitment on the same page, at 30 days, 90 days and unlimited across the three plans.
GoHighLevel is where many agencies land, because it is a CRM first. Its published agency plans, all billed monthly, are Starter at $97 per month, Unlimited at $297 and Agency Pro at $497, each also offered annually at ten times the monthly figure, with an Enterprise tier by quote, plus a long list of per-sub-account add-ons (GoHighLevel pricing, checked 15 September 2026). Its dialing is a power dialer, not a predictive one.
So the real split is not "predictive versus power". It is:
- Contact centre platforms (Convoso, Five9, Genesys, NICE) that sell pacing, seats and workforce management, mostly by quote.
- Per-seat sales dialers (PhoneBurner, Kixie) that sell one rep dialing faster.
- CRMs with a dialer inside (GoHighLevel, Inflowave) where the dial button is a feature of the contact record rather than a separate phone system.
Where Inflowave sits
Inflowave is in the third group and ships a power dialer, not a predictive one: one call at a time per rep, from a queue built out of lead filters rather than a CSV, with no pacing ratio and therefore no abandoned-call rate to meter. Each call writes a single row joined to the lead, the client and the workspace, carrying the Twilio call SID, duration, the AMD verdict and the outcome tag, so the result lands next to that person's DMs, SMS and email rather than in a separate phone system. Recordings and transcripts attach to the same record.
The boundary is deliberate and worth stating plainly. If your operation genuinely needs twenty seats paced against a cold consumer list, buy a contact centre platform and staff the compliance function that goes with it. Plan limits also matter when you compare: connected social accounts and client workspaces are 1 on Influencer at $27 per month, 1 on Business at $149, 10 on Pro at $297 and 50 on Max at $497, all billed monthly, with team seats at 1, 1, 5 and uncapped on Max (Inflowave pricing, checked 15 September 2026).
If you want calls placed and qualified without a rep on the line at all, that is a different product again, an AI voice agent, and it sits alongside a dialer rather than replacing it. If you are still designing the outbound motion itself, start with what cold calling actually involves before choosing a dialer at all.
The short version
A predictive dialer is a pacing algorithm with a phone attached. Its job is to keep agents talking, and its unavoidable by-product is a call that a human answers with nobody behind it. US regulators cap that by-product at three percent of live-answered calls, define the failure as anything longer than two seconds after the greeting, and require you to keep the records proving it. The benefit comes from pooling agents, so it grows with headcount while the compliance obligation does not shrink with it.
For most small teams, the useful question is not "predictive or power". It is whether the dialer knows who it is calling.
Frequently asked questions
What is a predictive dialer?
A predictive dialer is outbound calling software that dials more numbers than it has agents free, using measured answer rate, ring time and call duration to predict when an agent will become available. The aim is that a live human is already on the line when an agent finishes their previous call. It is used mainly in high-volume outbound call centres, where a large pool of agents makes the prediction statistically reliable.
What is the difference between a predictive dialer and an auto dialer?
Auto dialer is the umbrella term for any software that dials numbers without a person pressing the keys. Predictive dialing is one type within that family, defined by dialing several lines per available agent and modelling agent availability. Other types include preview dialers, which show the record before dialing, and power dialers, which dial one number at a time per agent. So every predictive dialer is an auto dialer, but most auto dialers are not predictive.
Are predictive dialers illegal?
No, but they are regulated. In the US, FCC rules at 47 CFR 64.1200(a)(7) prohibit abandoning more than three percent of telemarketing calls answered live by a person, measured over each 30-day period of a campaign, and define an abandoned call as one not connected to a live representative within two seconds of the called person's completed greeting. The FTC's Telemarketing Sales Rule imposes an equivalent safe harbour. Operating a predictive dialer legally means measuring and recording that rate.
Is an autodialer illegal?
Automatic dialing is not banned outright, but specific uses are restricted. US rules restrict calls placed with an automatic telephone dialing system or a prerecorded voice to certain numbers without prior express consent, require ringing for at least fifteen seconds or four rings before disconnecting an unanswered telemarketing call, and require honouring do-not-call requests. Rules also vary by state and by country, so the legality depends on who you are calling, with what consent, and from where.
What are the key differences between a predictive dialer and a manual dialer?
Manual dialing means a person reads a number and dials it, which guarantees intent and consent checks but spends most of the hour on dialing, ringing and voicemail. A predictive dialer removes that wait by dialing ahead of the agent, at the cost of a percentage of calls answered by someone with no agent behind them. Manual dialing has no abandonment rate to manage and no pacing to tune; predictive dialing has both, plus record-keeping obligations.
What are the benefits of using predictive dialer software?
The main benefit is agent utilisation. In a large team working a low-answer-rate list, agents can spend most of the hour in conversation rather than listening to ring tones and voicemail greetings. Secondary benefits include campaign-level pacing controls, automatic retry scheduling and centralised reporting on connect rates by list and time of day. The benefit comes from pooling agents against a shared queue, so it grows with agent count.
Is there a cheap predictive dialer available?
Cheap is uncommon in this category, and the published headline figures are not the entry price. As checked on 15 September 2026, Five9 publishes $119 monthly per seat for its Digital tier and $159 monthly per seat for Core, billed monthly, but the same page states a minimum of 50 seats, so the actual floor is around $5,950 per month rather than $119. Convoso does not publish pricing at all and states that it tailors quotes and primarily offers annual plans, currently at 20 percent off. Power dialers are the cheaper adjacent option: PhoneBurner publishes $140 per user per month billed annually on its Standard plan, or $165 per user per month billed monthly, and CRM-based dialers bundle calling into a platform fee rather than charging per seat.
