MAYUR VIRKAR explains how AI reads one’s business before their banker does and how one’s data becomes their collateral
Imagine a fabrication contractor in Bicholim who needs `40 lakh of working capital. Three years of GST returns. Two years of e-invoices. A current account with steady receipts. No property to pledge that a credit committee will accept. Her file goes up, sits, and comes back six weeks later asking for collateral she does not have.
Nothing in that file was missing. What was missing was a cheap, fast way to read it. That reading is the job Indian lending is now handing to software, and it changes what every Goan business owner should be doing with their own paperwork, starting well before they next walk into a branch.
What the model actually reads
A cash-flow underwriting model does not look for a flat to attach. It opens the GST returns and the bank statement side by side and asks whether invoiced sales actually arrived as receipts, and how late. It learns the seasonality every Goan business lives with, the monsoon trough and the pre-season rush, so a thin July reads as a monsoon and not as distress.
It flags what fails to reconcile: an invoice with no matching credit, a single customer who is one cancelled order away from taking the business down with them. What comes out is a repayment capacity estimate and a reason code, which a credit officer accepts, questions or overrides.
This is not a forecast. The RBI’s FREE-AI committee, reporting in August 2025, named these alternate signals, GST filings, utility payments, transaction behaviour, as credible inputs for underwriting exactly the thin-file borrower who cannot produce collateral. The same committee found only about one in five of the regulated entities it surveyed using or building AI at all, and that fifth driven almost entirely by large banks and NBFCs. The technique is proven. The distribution is not.
Why the gap is wider here
As on 30 June 2026, Goa’s banks held `1,37,333 crore in deposits against `47,194 crore in advances. Credit-deposit ratio: 34.36%. The State Level Bankers’ Committee has been careful with that number, and so should we be. Its own RBI regional director told the August meeting that structural issues and the actual usage of credit in the state are not reflected in the ratio, and MSME lending drew praise rather than criticism, at `4,523 crore in one quarter. But the shape underneath is familiar to anyone who has raised working capital here. Collateral-rich borrowers get funded. Collateral-light, cash-flow-healthy ones wait.
What you can do before you next ask for money
Start treating your GST filing as credit data, not tax compliance. A model built on returns and receipts can only assess the record that exists; gaps and irregular filing are not neutral to it, they read as absence of evidence.
Push receipts through the bank account. Cash trade is invisible to any system reading statements, which means a profitable business can score worse than a smaller one that banks everything.
Reconcile your own invoices to your own receipts, quarterly, before a lender does it for you. Whatever does not tie will be flagged, and you would rather have the explanation ready than be asked for it cold.
Know your customer concentration and be able to say it out loud. Concentration risk is one of the first things these models compute, and one of the few a business can change.
The lender side is moving too
The Unified Lending Interface, the RBI’s consent-based pipe carrying land records, satellite imagery and property search to any lender that connects, had 64 lenders on it by December 2025, up from 36 a year earlier. Nothing moves through it without the borrower’s permission, which makes consent a thing worth understanding rather than clicking. And in June 2026 the RBI put out draft guidance on model risk management that, if notified as drafted, would require board-approved oversight and independent validation of every model that shapes a lending decision. That is what makes an algorithmic decline explainable when you ask why.
The contractor in Bicholim does not need a new bank. She needs her existing one to read what she has already generated, and she needs that record to be worth reading. The first half is being built. The second half is entirely hers.