No Human in the Loop: Cars24's First AI Loan Workflow

Most AI projects in financial services are actually AI-assisted. A human reviews, approves, or signs off somewhere.

What went live this week at Loans24 is different: a reference check workflow that runs from call initiation to loan origination system (LOS) status update with no human involved at any step.

Reference checks sound simple, but they are where most fulfillment friction hides. When someone applies for a car loan, the team calls the references listed on the application to verify details: name, relationship, address, employment.

It sounds simple.

In practice, reference checks are one of the more friction-heavy steps in fulfillment. Miss a step or get a mismatch, and the underwriting queue stalls.

The architecture runs in three layers. First, an AI calling agent dials the reference automatically when their number is added to the loan file. No human schedules the call or dials it.

Second, every call, whether made by AI or by a human agent, is transcribed and analyzed by an LLM that extracts key data points and tags each with a confidence score.

Third, a rules engine matches those tagged outputs against what is already in the LOS.

The matching layer is where most of the engineering difficulty lived. Spoken addresses do not match typed addresses reliably. Someone says "DLF Phase Two Gurgaon" and the system has "DLF Phase 2, Gurugram" on file.

Phonetic matching, fuzzy string handling, and LLM confidence scoring each exist as solved problems individually. Combining them in a production lending context took real iteration.

25% of all reference calls are now fully AI-dialed and AI-analyzed. 100% of all calls, whether AI-dialed or agent-dialed, are transcribed and analyzed.

Every status update flows from the calling service directly into the LOS in real time.

The 75% that still involves human callers is not a capability gap. It is a deliberate choice.

The team is benchmarking AI-extracted outputs against approvals from the Risk Control Unit. Once confidence data is strong enough across a meaningful sample, that remaining volume moves to full AI.

Loan decisions have real consequences, which is why auto-approvals are not yet live. Getting a reference check wrong does not just slow down a file.

It can wrongly block credit for someone who needed it, or let through an application that should have been flagged.

Building confidence through parallel running, where AI analysis runs alongside human decisions and both are tracked, is the right approach in this context.

This is not caution born of fear. It is caution because the deployment environment demands it.

Every AI-extracted output is confidence-scored and logged. When auto-approval goes live, it will be backed by thousands of calibration data points, not by a belief that the model performs well in theory.

For borrowers, this collapses the wait from days to minutes. A reference check that used to wait for an available agent now happens within minutes of the application reaching that stage.

In auto lending, where the buying decision is often time-sensitive, that delta matters.

A customer who applied Friday evening should not still be waiting on a reference call Monday morning because the weekend queue was long.

This is the first workflow in the Loans24 stack that is fully automated end to end. It sets a template. The approach is repeatable.

Identify a high-volume, rule-bound step that depends on human execution. Build AI to handle it. Run in parallel until confidence is established. Then cut over.

Reference checks were the right first candidate because the task is well-defined, success criteria are measurable, and the cost of any individual error is contained.

The harder automations, the ones closer to final credit decisions, will follow the same methodology but require longer calibration windows. That is the right tradeoff.

Most AI in lending is cosmetic. This one is operational. There is a lot of AI in lending that is cosmetic: chatbots at the top of funnel, dashboards that surface insights a human then acts on, recommendations that still require someone to decide.

The distinction that matters is whether AI owns a workflow end to end and is accountable for the output. This reference check system does not recommend an action. It takes one.

As this model extends across more steps in the fulfillment time, the economics of loan processing change materially: faster cycle times, fewer manual touchpoints, lower error rates on data extraction.

The number of borrowers processed quickly without a human in the loop expands. That matters most in markets where loan volumes are growing faster than the ability to hire and train agents at scale.