By Abhishek Gupta, Chief Technology Officer, Stashfin
Speed and safety stopped being a trade-off once risk evaluation became software.
For most of the history of consumer lending, speed and safety pulled in opposite directions. An instant approval meant taking on more risk. A careful one meant days of document checks and manual review. Lenders picked a point between the two and lived with it.
That trade-off was never a law of nature; it came from how the work was done. When risk assessment depends on people reading documents, every extra check adds hours. When it runs as software, an extra check adds milliseconds. So the real question isn’t whether AI can make credit faster without making it riskier. It’s how a lender has to be built for that to be true.
Fast doesn’t mean Fewer Checks
The usual worry is that “instant” means corners were cut. In a well-built system it means the opposite: more signals get evaluated, not fewer. One application can be checked against bureau history, banking behaviour, device and identity signals and known fraud patterns, all at once, before a human reviewer would even have opened the file.
There’s a less obvious point too: being slow is a risk in itself. Creditworthy borrowers have options. A lender that takes three days loses its best applicants to whoever answers first, and is left with the ones who had nowhere else to go. Done well, speed improves the quality of the loan book.
AI as the Rails, Not a Feature
At Stashfin, AI isn’t an add-on. It runs underneath underwriting, fraud detection, customer support, engineering and collections. That’s how loan applications have grown 25x while the company has stayed at roughly 200 people, serving more than 50 million registered users.
Part of this is the infrastructure we own. We run an in-house fleet of NVIDIA Blackwell GPUs that serves a range of open-weight models, from compact 3-billion-parameter models up to 120 billion parameters, each matched to the job it does. Together they process close to 3 billion tokens a day. Keeping models in-house means sensitive customer data stays inside our environment, and lets us tune models to Indian credit behaviour instead of relying on general-purpose systems.
For prime and super-prime customers, the credit decision takes sub seconds, and money can reach their account in under five minutes.
The Checks Customers Never See
Fast lending attracts fraudsters as surely as it attracts good customers, so a lot of our AI work is in places customers never notice. Silent mobile verification confirms in the background that the phone number belongs to the device applying, without adding steps. Device, identity and behaviour signals are checked against patterns of synthetic identities and mule accounts. Unusual activity is flagged in real time, not found weeks later in a portfolio review.
Explainable by Design
In regulated lending, a decision nobody can explain is a liability. As an RBI-regulated NBFC, we draw a clear line. Models make individual credit decisions; people own everything around them. Our credit and risk teams set credit policy, approve models before they go live, watch them for drift and bias, and handle exceptions. Every decision can be traced back to reasons we can explain to the customer and to the regulator.
Collections: Empathy at Scale
Managing risk doesn’t end at approval. It also covers how we treat customers when repayment gets hard. Our aim is to stop financial stress from turning into default.
A small language model built for collections decides who to contact, when, and on which channel. The result is fewer, better-timed and more respectful conversations, in the customer’s own language. AI handles routine reminders and prioritisation, while trained human teams handle sensitive hardship cases. It’s AI and people working together, not AI replacing people.
The festive season is the real stress test. During our usual Diwali-to-Christmas quarter, traffic on our platform doubles, disbursals to new borrowers rise more than 40%, and average ticket sizes grow 15%. Over the same period, AI-led engagement pushed collections and repayments up. The platform absorbed the surge without giving up portfolio quality.
Next: The Interface Disappears
The next step for AI in financial services isn’t another chatbot. It’s fewer screens. Whether a customer is borrowing, paying over UPI or buying digital gold, they shouldn’t have to work through menus. Speak it. Tap it. Pay it. Borrow it. Save it. The system understands what the customer wants, remembers context and removes friction. Over 99% of customer queries are already resolved without a human agent, with people available whenever a customer needs one.
The real test of AI in lending is whether the business holds up commercially. Stashfin has crossed ₹20,000 crore in lifetime disbursals and has been profitable for four years in a row. Built properly, speed and risk management aren’t a balance to be struck. They come from the same architecture.