What Is a Contact Center Predictive Dialer?

What Is a Contact Center Predictive Dialer?

A contact center predictive dialer is the outbound calling engine that places more calls than agents available, using rolling statistics on answer rate, call duration, and abandon rate to keep agents busy. It is the most aggressive of the four classic dial modes — preview, progressive, power, predictive — and the one most likely to violate abandon-rate regulation if pacing math is off. In a 2026 AI contact center, predictive dialers pipe live audio into an LLM voice agent on LiveKit or Pipecat, and the dialer’s pacing assumptions become a quality signal the bot inherits.

Why It Matters in Production LLM and Agent Systems

Predictive dialers were designed around human-agent throughput, not LLM-agent throughput. A human agent has predictable wrap time; an LLM agent might be infinitely scalable but takes 800 ms to load context per call. If the predictive model assumes “agent” = “infinite capacity”, the dialer over-dials, calls connect to a still-warming bot, and the first two seconds of audio are missed. The greeting is corrupted, the customer hangs up, and the dialer logs a successful connect. Resolution rate drops without a clear cause.

The pain is felt across roles. A voice engineer chases reports of “the bot starts talking before the customer hears it” and discovers the dialer is running at 2.4× pacing. A compliance officer is asked whether the campaign honoured the 3% TCPA abandon-rate limit, and finds abandons spike when the LLM provider has high TTFA. A product owner sees ConversationResolution drop on Tuesdays because the dialer recalibrates pacing weekly and Tuesday traffic is bursty.

In 2026 stacks, the only way to keep predictive pacing safe with an LLM agent is to expose dialer state on every voice trace — dialer.mode, dialer.attempt_number, dialer.answer_type, dialer.pacing_ratio — and slice eval results by it. Without those attributes, dialer-induced quality drops look like model regressions.

How FutureAGI Handles Predictive Dialers

FutureAGI’s approach is to surface dialer state as first-class span attributes on every outbound voice trace. When traceAI-livekit or traceAI-pipecat is wired, every span carries the dialer mode and the answer-type. ConversationResolution runs per call and is sliced by dialer.mode = predictive versus other modes, exposing pacing-related quality drops. ASRAccuracy catches transcript errors that correlate with first-second audio loss. The Agent Command Center’s routing-policy can branch on dialer.answer_type so voicemail connects route to a TTS-only flow and live-person connects route to the full LLM agent.

A concrete example: a sales-development team runs an outbound campaign with a Pipecat voice agent dialed by a Five9 predictive dialer at 2.0× pacing. Their FutureAGI dashboard shows ConversationResolution of 0.74 on predictive-mode connects and 0.86 on progressive-mode connects. The team narrows pacing to 1.4×, wires answer-machine detection into the routing policy, and reroutes voicemails to a leave-a-message TTS branch. Resolution on predictive recovers to 0.84 and the abandon rate drops below the regulated 3% ceiling. Without dialer-aware tracing, the team would have blamed the model and migrated to a more expensive provider.

How to Measure or Detect It

Predictive-dialer-driven AI calls need a focused signal set:

• ConversationResolution by dialer.mode: predictive vs. progressive vs. preview can differ by 15+ points on identical scripts.
• First-2-seconds audio drop rate: percentage of calls where transcription starts late; correlates with pacing aggression.
• Abandon rate by hour: regulatory ceilings (TCPA 3%, OFCOM 3%) are hourly, not daily; track per hour.
• ASRAccuracy by attempt number: third and fourth dial attempts often have worse audio; track separately.
• Pacing-ratio drift: dialer recalibrates over time; alert when pacing deviates by more than 20%.

Minimal Python:

from fi.evals import ConversationResolution

evaluator = ConversationResolution()
result = evaluator.evaluate(
    input="Outbound goal: schedule discovery call",
    output=conversation_transcript,
)
print(result.score, result.reason)

Common Mistakes

• Treating the LLM agent as infinite-capacity in pacing math. Bot warm-up time is real; configure the dialer for it.
• Tuning pacing weekly, evaluating quality monthly. Pacing changes leak into eval scores; align the cadences.
• Same script across all dial modes. Predictive’s first second sounds different to the customer than preview’s; scripts must adapt.
• No dialer.mode attribute on traces. You cannot slice resolution by mode if the field never reaches the LLM span.
• Ignoring abandon-rate regulation. TCPA, OFCOM, and ACMA enforce hourly ceilings; one bad hour is a violation.