The Real Bottleneck in Digital Lending Isn't Technology

5 MINS

I led digital lending products at Axis Bank for almost three years. We doubled the small business loan book, launched native personal loans on UPI rails, and shipped two new variants in partnership with risk and policy. People assume that the hard part of digital lending is the tech — the API, the credit model, the funnel. It isn't. The hard part is everything that happens before the engineer ever opens a ticket.

What actually slows a lending feature down

Engineering velocity in fintech is rarely the constraint. Here's where the time actually goes:

Risk policy alignment — what's the underwriting box this product is allowed to live inside, and who signs that off?

Channel partner consensus — banks rarely ship lending features alone; partners (UPI players, fintechs, consumer apps) all have an opinion.

Compliance and audit — the regulator's questions you can predict, and the ones you can't, both have to be answered in writing.

Marketing and segmentation — who is even allowed to see this offer, and what's the comms story when they do? In my experience, the engineering build was 25% of the calendar. The other 75% was alignment.

The "Known to Bank" lesson

When we conceptualised Personal Loans for Known to Bank customers using an alternate credit model, the riskiest assumption was not the model. It was: *will the risk team trust an alternate signal enough to put their name on a sanction?*

The answer was, eventually, yes — but only because we did the unglamorous work first. We co-designed the credit model with risk, not for risk. We let them poke holes. We built a parallel-shadow comparison so they could see how the alternate signal performed against the existing one, on a real cohort, before a single rupee was disbursed.

The engineering went from 12 weeks to 3 once the alignment was done. Most teams never get there because they don't budget for the alignment work as work.

The partnership trap

Partner integrations look like a shortcut. They aren't. Each new partner adds a tax on every future change. By the time we'd built partner-specific user journeys for a few major fintechs, our roadmap was effectively held hostage to whichever partner had the loudest VP that quarter.

The lesson I took into payments and into Cisco: build the integration as a product, not as a deal. If the integration has a clear lifecycle, an owner, and a deprecation path, it scales. If it's bespoke, it becomes the next quarter's debt.

What stayed with me

Three things I carry from those years that still apply to AI products today:

Compliance is a feature, not a tax. If your product looks compliant on the surface but is hard to audit underneath, you'll pay for it the first time someone asks for a trail.

The alternate signal beats the obvious one. In credit, in growth, in AI evaluation — the most valuable signals are the ones nobody else is using yet.

The roadmap is a negotiation, not a document. Every line on it represents an alignment that was won. The doc is just the receipt. Lending taught me that. AI didn't change it.

Why I Bet On Building an Integrations Platform Before Features

Using LLMs to Evaluate Agents Is Harder Than Building the Bot

Shivani skipped presentations and built real AI products.

Shivani Kolala was part of the March 2026 cohort at Curious PM, alongside 17 other talented participants.