How Cars24 turned a million conversation minutes into recovered leads

Cars24 runs one of the largest AI-native used-car marketplaces in India, with operations extending into the UAE and Australia. OpenAI's case study on the deployment, published July 16, lays out what the company built on the API, and the numbers are worth sitting with: 1M+ conversation minutes handled by AI agents every month, a 50% increase in customer support resolution rates, an 80% cut in turnaround time across key service workflows, and 12% of previously lost seller leads recovered through automated re-engagement.

There's no new model or product attached to this — it's a deployment story. Those are where most of the real economics of production AI actually show up.

The problem: conversion that spans weeks, not sessions

Buying or selling a car in India isn't a checkout flow. It runs through calls, document checks, financing back-and-forth, and follow-ups that stretch over days or weeks, most of it happening off the app. Scaling that by hiring is a linear cost curve against a business trying to grow faster than linear. That's the specific constraint Cars24 built against, which is why they started agents in the middle and bottom of the funnel — where conversation drives conversion — rather than in acquisition.

What the agents do

A buyer calling in gets asked about budget, family size, commute, and preferred car type. The agent recommends cars from the catalog, books a test drive, and opens the financing conversation. Before the visit it confirms, checks whether preferences shifted, and collects financing details; after the visit it checks intent — proceed, rebook, or look elsewhere; after purchase it handles feedback, warranty, and returns.

The seller side mirrors this: an agent collects vehicle details, schedules inspection, sends reminders, reschedules missed appointments, and gathers competitive intel when a seller has shopped elsewhere. The detail that stands out: leads that used to go cold after 10 days without a human following up are now re-engaged automatically, requalified, and routed back into the funnel once Cars24 can meet the seller's price. That workflow — not a better ad or a smarter recommender — is where the 12% recovery number comes from.

Codex as an operating layer, not a coding tool

The second half of the case study is about Codex, and it's the more interesting part for anyone running an engineering org. Cars24 tags Codex into bug reports, uses it to draft and refine Linear tickets, and has it summarize GitHub activity to cut down on standups. They reoriented their project-management workflow around Linear specifically to give Codex a cleaner surface to work from — the right sequencing: shape the process so the agent has clean inputs and outputs, rather than bolting an agent onto a messy one.

Adoption jumped past engineering: finance uses Codex to pull numbers for investor reporting and to review purchase orders above a threshold, flagging anomalies and auto-approving clean ones. About 600 employees across engineering, finance, legal, marketing, and operations are on ChatGPT Enterprise and Codex, with 85–90% daily active usage — a high number for an internal rollout at that headcount.

My read

What makes this worth reading isn't the model or the API surface — neither changed. It's that Cars24 treated conversation infrastructure as the highest-leverage place to spend agent effort, because their funnel is conversation-shaped: multi-session, high-touch, and previously bottlenecked on headcount rather than demand. The 12% lead-recovery number is the tell. It didn't come from a smarter recommendation model; it came from an agent doing the unglamorous, high-volume work of following up on leads nobody had time to chase. That's usually where agentic deployments pay off first — not the flashy customer-facing demo, but the workflow that was previously too tedious to staff.