From Bureau to Behaviour: How Lending Got Smarter

There’s a certain smugness in the world of lending.

You’ll hear it over conference tables and VC calls:
“We have bureau integration.”
“We collect six months of bank statements.”
“Our credit model is rock solid.”

Congratulations. You’ve just unlocked 2016.

Because while lenders continue to obsess over PDFs and bureau scores like desi dads doing shaadi matchmaking - “Beta, bank PDF clean hai, CIBIL 720 hai, toh rishta pakka samjho.”
as if a credit score and a salary slip are all you need to judge repayment behavior. Bharat has moved on.

It pays in UPI, saves in mutual funds, side-hustles on Meesho, and files GST returns from a phone that hasn’t seen a laptop since the Jio launch.

And here’s the kicker: if you rely only on traditional underwriting, you’re probably rejecting some of the best borrowers in India.

• Bureau Score? ✅

• Bank statement with salary credit? ✅

• Aadhaar + PAN? ✅

That should be enough, right?
Only, it’s not.

Because behind that clean PDF, you missed:

• The loan apps they borrowed from three nights ago, too fresh for the bureau, quietly uninstalled before your checks began

• The betting app quietly installed at 2 a.m. every Saturday night

• The ₹25,000 monthly SIP running quietly in the background, showing they clearly have the capacity to repay a ₹5,000 loan

• Intent
  Did they apply for 6 loans in the last 48 hours across every “instant loan” app on the Play Store?
  That’s not credit planning. That’s credit spraying.
  

• Context
  Is ₹15,000 monthly income risky? Not if you’re self-employed in Tier-3, living rent-free and earning in cash.
  Without context, your model’s just reacting to digits, not realities.
  

• Behavior
  Did they uninstall your app minutes after disbursal? Are they rotating between 3–4 lending apps every week?
  That’s behavior screaming short-termism—or worse, fraud.
  

Stability
Are they using a rooted device? Do they change phones every month? Or are they running a 5-year-old phone with no screen lock?
Digital stability isn’t just IT hygiene—it’s a window into borrower seriousness.

Let’s say your borrower is a gig worker who earns ₹25,000 a month:

• They use cash.

• They’ve never taken a formal loan.

• Their account balance rarely crosses ₹500.

Your bank statement model screams “high risk.”
But zoom out, and you’d see:

• Their SMS inbox shows 12 months of consistent BNPL and gold loan repayments

• They’ve never missed an Amazon Pay Later due date

• They opt for annual mobile recharges instead of monthly top-ups

• They applied at 10:30 a.m. on a weekday—not at 2:47 a.m. while panic-scrolling

That’s not a high-risk borrower. That’s your ideal borrower.
But you missed them—because your system was trained to trust balance over behavior.

That’s where alternate data comes in.
And it doesn’t just fill the gaps. It flips the whole underwriting game.

Let’s say a borrower uploads a clean bank statement.

On paper, the borrower looks great.

CIBIL score? 735.
Salary? ₹28,000 credited regularly.
Bank statement? Clean. No visible EMIs. No bounces.
PAN and Aadhaar? Verified.

Traditional policy says: Approve.

But then you dig deeper and suddenly, the story begins to crack.

It starts with the SMS trail.

You notice UPI alerts from banks that were never declared.
In one of those accounts, an EMI for ₹1,850 to a little-known NBFC which was recently taken
Another shows a ₹2,300 deduction to a BNPL app that doesn’t report to bureaus yet.
None of this showed up in the PDF the borrower uploaded. Because they only gave you one of their four active accounts.

And in that moment, you realise: the bank statement wasn’t wrong.
It just wasn’t the whole picture.

Then you look at the installed apps.

The phone recently had two instant loan apps, both installed 72 hours before application.

You also notice a crypto trading app and a fantasy gaming platform.
Not saying crypto or gaming is a crime. But when paired with emergency loans? That’s a pattern.

The red flags don’t end there.

The phone number is prepaid, barely 90 days old, and flagged on Truecaller by over a dozen users.
It’s not even saved under the borrower’s name.

You dig a little deeper—turns out there are 3 instant loan apps opened on the foreground while applying to your loan

The borrower wasn’t shopping for credit.
They were carpet bombing.

And then there’s the timing.

Application came in at 3:43 a.m. on a Thursday.
The borrower rushed through onboarding in under 45 seconds.
Didn’t review any T&Cs.
Skipped the optional liveness selfie.
Uploaded a cropped-out Aadhaar image taken from a file manager, not the camera.

You don’t need an AI model to read the room, this wasn’t a thoughtful application.
It was a panic push.

Just to confirm, you pull Account Aggregator data.

The uploaded bank statement showed a ₹42,000 closing balance. Looked solid.

But AA reveals the truth: the money came in 24 hours before download, and ₹39,000 exited the account two days later.

What looked like savings was just a staged snapshot.

Now here’s the twist.

This borrower says they need a ₹7,000 loan for an emergency.

But you see a consistent ₹25,000 monthly SIP to a mutual fund.
That’s not a borrower in distress. That’s a borrower optimising the system.

Meanwhile, another applicant—CIBIL 620, ₹19,000 income—has a small ₹2,000 SIP, pays their phone EMI via UPI every 5th of the month, and has had the same number for six years.
No flash. Just quiet, stable behavior.

So what do you really know?

The first borrower passed every traditional test.
And would’ve been approved, until alternate data stepped in and said: Not this one.

Because underwriting isn’t just about what’s in the file.
It’s about everything they tried to leave out.

This is what alternate data does, it doesn’t just fill the gaps.
It rewrites the story.

Where traditional underwriting sees balance and score, alternate data sees intent, behavior, context, and reality.

And that’s exactly what you need if you’re building for Bharat.

Let’s say you’ve got all this beautiful alternate data now.

You know what’s on the borrower’s phone.
You’ve seen their UPI flows from five different banks.
You’ve caught their EMIs hiding in the SMS inbox.
You even know what time they apply for loans.

But here’s the twist: if you’re still processing all this with a dusty rule engine that says,

“If credit score > 700 and income > ₹20K, approve the loan” —
you’ve just brought a bullock cart to a rocket launch.

The Rule Engine Problem

Rule engines were great—10 years ago.
Simple IF-THEN logic, easy to write, easy to test.
But then came the “optimizations.”

You know how it goes:

✍️ Rule 1: Reject if salary < ₹15K
✍️ Rule 2: Reject if more than 2 active EMIs
✍️ Rule 3: Reject if applied past 10 p.m.
✍️ Rule 38: Reject if phone is rooted
✍️ Rule 92: Reject if 4 loans in last 90 days
✍️ Rule 173: Reject if savings account balance < ₹2,000
✍️ Rule 198: Approve if postpaid number and KYC verified—even if rule 92 failed
✍️ Rule 199: Wait what?

Soon you’re juggling 200 rules that contradict each other more than Indian relatives during a wedding budget discussion.

And every time the credit head wants to test a new idea, your tech team dies a little inside.

Traditional rule engines work like checklists.

If A, B, and C are true → Approve.
If X or Y are present → Reject.

Simple. Clear. Logical.
And totally out of depth in 2025.

Because real-life borrowers aren’t IF-THEN statements.
They’re patterns. Behaviors. Contradictions.

That’s where machine learning comes in.

Instead of relying on hard-coded rules, ML learns from actual outcomes.
It studies thousands of past borrowers and begins to understand which combinations of signals led to on-time repayments—and which ones quietly exploded six months later.

Not just “CIBIL 700 = good.”
But “Borrowers like this repaid well, borrowers like that didn’t.”

It doesn’t just connect the dots—it learns which dots even matter.

Take this real-life example:

A borrower comes in with:

• CIBIL score: 630

• Monthly income: ₹22,000

• No visible EMIs in the bank statement

• PAN, Aadhaar? Verified

• Applied at 9:13 a.m. on a Monday

• Postpaid number, 6 years old

• UPI spends of ₹500/day

• SIP of ₹1,500 running for over a year

Now a traditional engine stops at that CIBIL score and goes,
“Ah, 630? Nope.”

But an ML model sees the full story:
The SIP shows financial discipline.
The UPI pattern is low-volatility and consistent.
The long-tenure number adds stability.
The Monday morning timing suggests thoughtfulness, not panic.
Put together, this borrower is a quiet rockstar—not a red flag.

And when approved?

Repays perfectly.

Because intent doesn’t live in a score.
It lives in behavior. In rhythm. In the digital fingerprints borrowers leave behind.

And here’s where things get even more interesting.

ML doesn’t just classify. It calibrates.

It doesn’t only say yes or no—it learns to say maybe... at the right price.

Because not every borrower is a clean approval or a hard rejection.
Some fall into the orange zone—too risky for 18%, too stable to discard.

Let’s say a borrower applies with:

• Two BNPL loans taken this week

• A crypto trading app active on their phone

• A ₹25,000 SIP quietly running every month

• And a timestamp of 3:43 a.m. on their application

Would you lend?
A rule engine might say: “Too messy. Reject.”
But a trained ML model might say: “I’ll lend—but at 36%, not 18%.”

That’s risk-based pricing in action.

It’s not about squeezing the borrower.
It’s about cushioning your book.

Because in Bharat, borrowers aren’t binary:

• Some won’t have thick bureau files

• Some won’t upload six months of neat salary slips

• But many will repay, if you underwrite with context

Modern credit systems don’t push everyone out of the funnel.
They find a smarter way to keep more people in it, without blowing up later.

Not every borrower needs a “No.”
Some just need a “Yes, but priced right.”

That’s not just underwriting.
That’s underwriting with insight.

Let’s face it, India has changed.

The borrower isn’t who they used to be. They may not have a thick CIBIL file. They might not come with six salary slips or a CA-stamped bank statement. But they leave behind a digital trail—one that tells you everything you need to know, if you know where to look.

Alternate data gives you vision.
Machine learning gives you judgment.
Together, they give you a shot at building a book that doesn’t blow up in six months.

And if you’re still underwriting based on hardcoded rules and patched-up bank PDFs?

That’s not old-school.
That’s just… lazy.

Because in this market, risk doesn’t show up wearing a red t-shirt that says “Default incoming.” It shows up in your approval list wearing a clean bank statement and a fake smile.

And the good borrowers? They’re out there. Hidden behind prepaid numbers and no bureau history. You just need a better flashlight.

So yeah—if you’re still underwriting the same way you did five years ago?

You’re not just behind the curve. You’re underwriting for a country that doesn’t exist anymore.