Case study · ElevenLabs
How we cut voice bot latency by 75% for a leading US insurance provider in four weeks, and tripled the call volume it handles with no added headcount.
One of the leading insurance providers in the United States, serving hundreds of thousands of policyholders across multiple states, used an ElevenLabs voice agent to call its payers about premium reminders, overdue payments, and policy updates. Payers weren't staying on those calls.
The bot spoke slowly, paused for more than three seconds before every reply, and sounded flat. Payers talked over it, hung up mid-sentence, or asked "Am I talking to a machine?" in the first few seconds. Nearly 4 in 10 calls ended before the bot got to the reason for the call. Fewer payers committed to payments, and the operations team spent time handling complaints about the calls.
The client asked us to make the bot fast and natural enough that payers would stay on the line.
We audited the whole call pipeline: more than 500 call recordings, the latency logs, and every setting in the agent. We found four problems, and each one added delay or stiffness.
Most of the delay came from the LLM. The agent ran a heavier model than a payment reminder call needs, and a long system prompt added processing time to every turn. The voice was set up for narration, with stability so high that the bot's intonation went flat. The turn-taking settings made the bot wait too long after a payer stopped talking. And the scripts read like policy documents, full of long sentences and formal phrasing that no human agent would say on a call.
We fixed one layer at a time over four weeks and tested each change on real call scenarios before starting the next.
Week 1
01We moved the agent to a faster LLM built for real-time conversation and cut the system prompt by more than 60%. We turned on streaming, so the bot starts speaking while the rest of its reply is still generating. That removed most of the dead air.
Week 2
02We picked a warmer voice for customer calls and lowered the stability setting so the bot's tone rises and falls the way a person's does. We also switched to ElevenLabs' low-latency voice model for real-time agents, which generates audio faster at the same quality.
Week 3
03We shortened the silence detection and turn timeout settings, so the bot answers soon after a payer finishes and stops talking when a payer interrupts. We also rewrote every call script in short spoken sentences with small acknowledgments like "Got it" and "Sure, one moment," and split long policy details into pieces a payer can follow by ear.
Week 4
04We ran more than 300 test calls covering payment reminders, objections, wrong numbers, and callback requests. We adjusted the settings based on those calls, then rolled the new agent out to the full calling program.
Here is how the calls compared before and after the four weeks.
Average response latency
Early call drop-offs
Call completion rate
Baseline →
+52%
Successful payment commitments
Baseline →
+34%
"Talking to a robot" complaints
Baseline →
Down more than 80%
Call volume handled
1x →
3x, with no added headcount
Before, our payers knew within two seconds they were talking to a bot, and most of them hung up. Now they stay on the call and talk to it like they would to one of our agents. Payment commitments are up, complaints are down, and we've tripled our call volume.
They reviewed the model, the voice settings, and our scripts, and they fixed the problems in each. We had a working agent in four weeks.