From Dealer Offers to Risk Outcomes: What Cyclops Reveals About AI in Credit

As part of my ongoing coverage of how Indian NBFCs are using AI in real, operational settings, I looked at remarks from Asheesh Goel, Chief Executive of Farmer Finance at L&T Finance, during the company’s Investors Digital Day. He outlined the company’s AI underwriting and dealer-offer optimisation system, Cyclops, and shared early results from its rollout.

The goal here isn’t to spotlight one company. It’s to understand how Indian lenders are applying AI at scale, what’s working, and what other financial institutions can learn when they plan their own deployments. India Inc is moving fast on this front, and many of these improvements are replicable across the sector.

Using the transcript and my own prompt, I summarised Cyclops’ performance in the table below (Dealer coverage, STP gains, LTV shifts, GNS outcomes, and more). The numbers aren’t meant to sell a success story. They show something more practical: AI is starting to move core credit, risk, and productivity metrics inside Indian NBFCs, not just peripheral customer-facing tasks.

This kind of transparency helps the wider industry understand what modern AI systems can actually deliver, what timelines look like, and how business teams adapt to machine-assisted underwriting. It also sets a benchmark for other lenders evaluating similar projects.

Cyclops is just one example, but it highlights broader lessons:

1. AI value is clearest when tied to a business bottleneck
   Here, the bottleneck was pricing, underwriting, and dealer engagement — not some generic “digital transformation.”

2. Full-funnel improvement beats isolated metrics
   Offer quality, STP, LTV, ticket size, and GNS all moved in the right direction. That’s what sustainable AI adoption looks like.

3. Execution speed separates leaders from laggards
   A five-month rollout is not typical in BFSI. It shows the impact of product-like execution rather than project-like execution.

4. Risk-adjusted gains are the strongest signal
   A drop in GNS and non-starter rates suggests that the model isn’t just automating decisions; it is refining them.

5. This isn’t unique or proprietary — it’s replicable
   None of the gains require exotic research. They require clean data, disciplined deployment, and a business team that is willing to trust the system.

I am tracking other applications of AI- will cover going ahead.