Intel / article

Webinar Recap How Cars24 Automates 3 Million Minutes Of Sales Calls With Voice Ai

Source: elevenlabs.io
Author: ElevenLabs editorial sitemap
Date: Published Oct 6, 2026

Context

Cars24, a used-car marketplace, split longer sales conversations among specialized AI agents after encountering problems with one handling everything. In ElevenLabs' webinar recap, smaller models reportedly lost track of longer calls, while larger models added delay. Combining every workflow in one set of instructions also made changes disruptive.

The reported design assigns a smaller model to establish the customer's intent, a larger model to explore buying needs, and a loan-focused agent when financing comes up. This gives different stages different processing capacity and separates their instructions, making each workflow easier to revise independently. The recap describes that division.

Key takeaways · AI-distilled

Cars24 moved off a single agent because, on 7 to 13 minute discovery calls, mini models lost after 3 to 4 minutes, larger models added latency, and one big prompt meant a loan-flow edit could break the buying flow.
According to Cars24, assembling its own STT, and TTS pipeline kept latency above one second even with the same models; the ElevenLabs platform cut latency 30 to 40%, which it says produced a similar uplift in conversions.
Cars24 caps prompts at 4,000 to 5,000 tokens for calls under two minutes, keeps answers to the most common questions in the prompt itself, and reserves tool calls and for the remaining 20% to avoid lookup latency spikes.
New agents roll out at 5%, 10%, 20%, 50% and 100% of traffic with two-day holds per stage. Cars24 runs evals on every call; a failed eval triggers manual review, and random samples are audited daily.
To cut early drop-offs from network lag or premature interruptions, Cars24 pre-generates and caches the opening message before the call connects and makes that first message uninterruptible.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
RAG — Retrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.

Read the source elevenlabs.io