← All notes

How Cars24 runs voice AI across 300+ retail hubs

On-ground intelligence + assisted-sales agents on the ElevenLabs stack. 22,000 hours analyzed, 35% conversion lift, 3M minutes of AI calls.

Originally written for the ElevenLabs blog. Republished here with the parts I care most about.

Cars24 runs more than 300 retail hubs across India. Millions of buyers walk in for test drives, inspections, and financing. Most are first-time car owners. Each test drive runs roughly 60 minutes, varies by SKU, and gets conducted in whatever language the customer is most comfortable in.

That’s a lot of conversations to be blind to.

We built two systems on top of the ElevenLabs stack to stop being blind: on-ground intelligence for what happens inside hubs, and AI voice agents for the assisted-sales funnel that runs in parallel.

On-ground intelligence with Scribe

The blind spot

Thousands of test drives a week. Mechanical issues raised mid-drive, loan objections at the negotiation table, SOP misses by salespeople — all of it disappearing into the air. We had no structured view of what actually happened on a drive, only the outcome (sold / didn’t sell). That made regional decision-making guesswork.

What we shipped

Hardware-free: open-mic recording on a tablet, audio streamed to ElevenLabs Scribe for transcription + diarization, then through an LLM pipeline that labels signal — customer vs. salesperson, which SKU, which hub, which objection.

Output is structured. We can filter for “every drive last month where the customer raised a brake issue”, or “every conversation where the loan offer was the deal-breaker”, or “which hubs are skipping the SOP intro on EVs.”

What it changed

  • 22,000+ hours of audio analyzed
  • +35% conversion uplift
  • +20% CSAT
  • −10% hub cost

The big shift isn’t the numbers. It’s that regional managers stopped arguing from anecdote and started arguing from transcripts.

Assisted sales with real-time agents

Why phone, not chat

Buying a car in India is a high-trust, high-consideration purchase — especially for the ~70% of our buyers who are first-time owners. Most conversions happen on a call. We run 1,300+ contact agents and roughly half of all sales are touched by an assisted call. Wait time and call quality directly move the funnel.

What we shipped

ElevenAgents, with:

  • Scribe Realtime V2 for live transcription
  • TTS Flash v2.5 for low-latency, natural-sounding responses
  • An orchestration layer tuned for buying guidance and objection handling — built around the playbook our top human agents already use

The goal: 30% of all calls handled by AI agents, with the bar being “indistinguishable from a strong human agent on the same script.”

What it changed

  • 3M+ minutes of AI-supported calls
  • 45% of assisted sales now touched by an agent
  • −50% calling cost

Customers don’t notice. That’s the whole point.

Where this is going

We’re building toward an AI-native operating model where voice is the primary surface for both customer understanding (what they say in the hub) and sales execution (what we say back). Inspection, sales, financing, post-sale — voice ends up in all of it.

The mission stays the same: better drives, better lives.


Co-authored on the ElevenLabs side by Tauseef Khan and Ruta Bhatt. Cars24 side: me, plus the AIRO and CLIO teams.