The numbers Brief Episode map Takeaways Charts Quotes Clips Glossary Go next Transcript

Episode 147 · Fintech · 59 min

A second to decide, two years to know

Scienaptic's argument is that Indian lending will not be won by approving more people but by touching fewer of them — decline the worst 10–15% outright, approve the clean 20–30% outright, and let machine learning grind down the grey area in between. The catch is that credit only tells you it was wrong two years too late, which is why the company now sells an AI whose layers you can peel open like an onion.

Credit & lending AI & machine learning Data & digitisation

Joy

Business leader, India, Scienaptic Systems · with Vishal Krishna

58:45

Said in this episode

▶ 30:48

40%

Applications decided with no human

Straight-through processing declines the clearly high-risk 10–15% and approves the clean 20–30%, leaving only the grey area for underwriters.

▶ 21:42

40–50

Lending clients in India

As stated on air; the India business accelerated after COVID, when lenders started wanting more sophisticated algorithms.

▶ 22:11

~100

US credit-union clients

The US book is larger than India's; credit unions are the American analogue of cooperatives, but far better organised and able to pay more.

▶ 20:06

1 point

Loss reduction that flips an NBFC

For lenders taking higher risk at higher rates, saving one percentage point of losses can decide whether the book is profitable.

▶ 5:35

1–2 sec

Time to a risk decision

The stated target for combining alternative and traditional data into a customised amount, tenure and rate for one borrower.

▶ 37:47

5 of top 10

Microfinance NBFCs on the platform

Named on air as public information, with CreditAccess Grameen and Samasta cited as clients; the transcript garbles both names, so treat the exact list as approximate.

The brief

The argument in sixty seconds

The claim here is that Indian lending finally has the three things a one-second credit decision needs — dependable data, the algorithms to make sense of it, and a platform that applies them identically every time without hundreds of underwriters — and that the binding constraint has moved from data scarcity to triangulation. GST filings behind an API and an OTP, the MCA portal, account aggregators, Udyam and fraud registries mean a kirana owner with no salary slip is no longer invisible; the reason two lenders quote you different rates on the same day is simply that they weight the same data differently. The economics do the rest of the arguing: nobody can spend an hour underwriting a ₹1 lakh loan, and sending someone to recover ₹10,000 costs more than the ₹10,000, so small-ticket unsecured lenders — mid-size NBFCs, fintechs, small finance banks, not the PSUs — are where a decision engine earns its licence, and a single percentage point of losses saved can flip a book from unprofitable to profitable. The near future is automation at both ends: decline the clearly bad, approve the clearly good, leave roughly 40% of applications untouched by a human, and shrink the grey middle iteratively. What holds it back is not accuracy but auditability — lending is a lagged business where you discover two years later that the model was wrong, so risk officers, auditors and the RBI want explainability and scorecards rather than a black box. The stakes sit in tier three and beyond: a farmer nobody can size, worth a ₹50,000 loan or a ₹50 lakh one, whose UPI transactions are at last filling in the ledger.

Worth your time if you are

Chief risk officers choosing between rules and models

NBFC and fintech operators underwriting small unsecured tickets

Enterprise software founders selling into regulated lenders

Anyone lending to borrowers with no salary slip

Episode map

Where the conversation travels

Every block is a chapter, coloured by what it's about. Click any of it to jump straight to that minute on YouTube.

Credit & lendingAI & machine learningData & digitisationSaaS & enterpriseRegulation & policySales, GTM & growth

01Cold open: science, synapse, ten years in 0:00 Vishal sets up a conversation about how Indian banks buy technology, and the guest explains the company name as science grafted onto synaptic — a decade-old business born in the US, with the US still its largest market, India second and four or five others in play. 02Finding the peak of a loan 3:46 Lend too little and the interest is trivial, lend too much and repayment collapses, so the whole product is finding the peak — the amount, tenure and rate a specific borrower can carry — in a second or two, using alternative data for the millions whose income is real but undocumented. 03Data, algorithms, platform 7:07 The problem breaks into three pieces — where the dependable data comes from, what turns it into intelligence, and the platform that does it the same way every time — because a rival will decide a small-ticket loan in five minutes and no NBFC can afford manual underwriting on ₹50,000. 04The API stack that replaced the PDF 9:30 Rewinding to a 2006 cover story where small businesses were first treated as retail borrowers, the guest walks through what changed: GST filings pulled through an API and an OTP, the MCA portal, account aggregators, Udyam and fraud registries — so much data that the new problem is sorting it. 05Why every lender quotes a different rate 13:23 Nobody uses all the available data and nobody synthesises it the same way, which is why offers differ borrower to borrower — and against the over-leverage worry, the answer is a spectrum from risk-averse PSUs to growth-first fintechs, with the RBI watching each month. 06Where a risk engine actually sells 17:08 PSUs are content with semi-annual processes and big private banks take years to approve a new vendor, so the sweet spot is small finance banks and mid-size NBFCs and fintechs taking real risk, for whom saving one point of losses can decide whether the book makes money. 07Small, unsecured, and impossible to touch 20:40 An hour of human time destroys the margin on a ₹1 lakh loan and recovering ₹10,000 costs more than ₹10,000, so unsecured small-ticket lending is where automated decisioning is worth most — a segment that took off in India after COVID and now carries 40 to 50 clients. 08Priced per decision, licence or SaaS 23:45 Some regulated lenders will never let data leave the building so they licence it, others hand over operations entirely; either way pricing is an annual block of applications plus overage, counting rejections as well as approvals — and cooperatives only work if ticket sizes or volumes rise, hence the idea of selling to a cluster as one client. 09What a business rule engine must do now 26:03 Frictionless ingestion, models behind a front end a non-coder can change the day Tamil Nadu floods, simulations of five hypotheses against historical data at speed, and models that keep learning from whatever passes through — security and scale assumed. 10Grow up, lose less, automate the middle 28:59 The consultant's 2x2 — growth on one axis, rupees lost per hundred lent on the other — plus a third dimension, straight-through processing that declines 10–15%, approves 20–30% and leaves the grey area for humans while machine learning keeps shrinking it. 11Four jobs around a single loan 31:20 Pre-approval off current-account flows that knows a kirana is short of cash before Diwali, core underwriting, early-warning signals that flag red-amber-green months before a default so loans can be restructured or an EMI resequenced, and cross-sell — which is exactly where borrowers start finding it intrusive and the RBI starts clamping down. 12The tier-three blind spot 36:11 A farmer in Tumkur may repay better than a salaried borrower but has no documented evidence, so lenders give him nothing; the guest names microfinance clients including CreditAccess Grameen and Samasta, says five of the top ten are on the platform, and argues even 10% of rural India dwarfs tier one and two combined. 13UPI's exhaust, ULI's promise 40:17 India is further ahead digitally and far more inclusive than the mature markets even if less regulated, and with UPI turning sparse rural bank accounts into dense transaction ledgers — plus account aggregators, unified lending rails and centralised data portals — the addressable segment keeps widening. 14Explainable AI, because lending lags 44:33 Risk officers are cautiously optimistic and sometimes plainly scared, because a black box in a lagged business means finding out two years later that the loans were bad — so large tickets keep a human, small tickets get AI made auditable, and the RBI now pushes scorecards over binary policy rules. 15A drop in the five trillion 48:24 Client-facing teams are regional while the machine-learning bench sits centralised in Bangalore serving other countries too, and the company claims to be growing faster than the lending market because lenders must now be not only bigger but sharper — yet at any plausible scale it remains a rounding error on a $5 trillion economy. 16McKinsey, Jack Welch and career planning 52:10 The leadership coda: consulting teaching the art of learning fast and knowing when not to speak, reading leadership books to audit his own gaps, deliberately taking a thousand-person people job and then a P&L in Singapore, and the advice to plan a career the way you'd plan a decade.

Takeaways

Ideas to carry out of this hour

A one-second decision needs three separate things

Turning an unknown borrower into an accurate repayment estimate in a second or two is not one problem but three: sourcing dependable data on someone who may have no salary slip and no bureau history, building the models that make intelligence out of it, and running a platform that applies the answer identically every time. The guest's point is that a lender needs all three — most vendors bring one — and that automation is not a preference but arithmetic, because a competitor decides a small-ticket loan in five minutes and no NBFC earns enough on a ₹50,000 loan to pay an underwriter to think about it.

▶ 7:22 ▶ 8:11 ▶ 8:58

The scarcity was data; the problem is now triangulation

A decade ago underwriting meant one bureau, a salary slip and a human reading a downloaded PDF. Now GST filings come through an API with an OTP — two to three years of returns, filed on time or not, top line and bottom line — alongside the MCA portal, account aggregators that surface every bank account on one consent, Udyam and fraud registries. There is now too much data, which is exactly why offers differ: not everyone uses all the sources and nobody weights them the same way, so two lenders can reach the same borrower on the same day with different rates and both be defensible.

▶ 11:20 ▶ 12:38 ▶ 13:37

The value of a risk engine is highest where the loan is smallest

Gold and property lending can absorb a human decision, because you can seize the collateral. Unsecured small-ticket lending cannot: spending more than an hour on a ₹1 lakh application eats the spread, and sending someone to recover ₹10,000 on a consumer-durable loan costs more than the ₹10,000 itself. That inverts the usual enterprise logic about selling upmarket — PSUs are content with semi-annual processes, large private banks take years to approve a vendor, and the buyers who move are the mid-size NBFCs and fintechs for whom one percentage point of losses can be the difference between a profitable book and an unprofitable one.

▶ 20:06 ▶ 20:40 ▶ 21:11

Automate both ends and iterate on the grey middle

The stated shape of straight-through processing is blunt: of every hundred applicants, decline the clearly high-risk 10–15% outright, approve the clean 20–30% outright, and roughly 40% of the book is decided without anyone lifting a finger. The remaining grey area is where underwriters should spend their time, and machine learning's job is to keep narrowing it. Behind it sits a consultant's 2x2 — how fast are you growing against how many rupees you lose per hundred lent — with the honest admission that approving everyone would send risk through the roof.

▶ 29:45 ▶ 30:30 ▶ 31:04

The most valuable model is the one that runs before the default

Underwriting is only one of four jobs. Pre-approval reads current- and savings-account flows to know that a kirana store is short of cash before Diwali and flush after it, so a bank offers an inventory loan instead of cold-calling a thousand people. Early-warning signals score the existing portfolio monthly and flag borrowers whose pressure is building two months before they miss, when restructuring or resequencing the EMI to land right after salary still works — by collections it is often too late. The same intimacy is what makes borrowers uneasy and the regulator attentive.

▶ 32:39 ▶ 34:40 ▶ 35:11

Explainability is the toll for using AI in credit

Lending is a lagged business: you trust the model, keep lending, and only two years later discover what was wrong with the loans you wrote — by which time the money is out of the door. That, not accuracy, is why risk officers are cautiously optimistic and sometimes scared, and why auditors and regulators ask how anyone knows the model is right. The practical settlement is a split: ₹1 crore tickets still justify a human who can be paid to look at five files a month, small tickets get AI that has been made auditable so the layers can be peeled back after the fact, and the RBI is pushing lenders from binary policy rules towards weighted scorecards.

▶ 45:50 ▶ 46:06 ▶ 47:53

Tier three is a documentation problem, not a credit problem

A farmer near Tumkur with one microfinance loan and a passbook may be wealthier and better able to repay than a salaried borrower, but the lender has documented evidence only on the salaried one. He could be worth a ₹50,000 loan or a ₹50 lakh loan, and because nobody can tell, he gets nothing. What is closing the gap is exhaust rather than paperwork: UPI has turned accounts that saw two cash-era transactions a month into dense ledgers that reveal ticket size, volume and seasonality — and five of the top ten microfinance NBFCs are already running the platform against exactly this population.

▶ 36:45 ▶ 37:18 ▶ 42:17

Plan the career the way you would plan a decade

The leadership coda is unfashionably deliberate. Consulting taught the art of learning an industry in three months and, just as usefully, when to stay quiet in a room of CXOs twice your age. Reading Jack Welch and other leadership books was a gap analysis — what do they have that I don't — which produced two conscious detours: a thousand-person people-management job that felt like changing diapers daily, then a P&L in Singapore. The advice that follows is to check a few metrics every few weeks against where you said you'd be, rather than chasing a 20% salary jump.

▶ 53:45 ▶ 55:48 ▶ 56:45

The numbers, drawn

What the episode measures

Every figure below was said on air — timestamps included, caveats kept.

Conversation share

portion of the hour spent on each theme

Credit & lending26%

AI & machine learning20%

Data & digitisation16%

SaaS & enterprise12%

Regulation & policy9%

Sales, GTM & growth7%

Computed from the chapter map of this episode.

Where the automation actually lands

% of applications

Auto-declined15%

Auto-approved25%

Grey area — human review60%

As described on air: 10–15% declined and 20–30% approved straight through, adding to roughly 40% of applications decided without anyone lifting a finger. Midpoints shown; the grey middle is what machine learning is meant to keep shrinking.▶ 30:48

Two books, two markets

clients

As stated in conversation — 'about 40, 50 clients' in India (midpoint shown) and 'close to 100' credit-union clients in the US, where the business is described as larger than India's.▶ 22:11

Worth keeping

Lines that stay

He could be worth a ₹50,000 loan or a ₹50 lakh loan. You don't know — so you're not going to give him anything. That's where the blind spot is.

— Joy ▶ 37:18

I buy a cell phone from the store and get a loan from somebody, and tomorrow sending somebody to my house to recover that ₹10,000 actually costs more than the ₹10,000.

— Joy ▶ 21:11

Lending is a lag business. You trust the model, and after two years you realise there were a lot of issues in the loans you made — it's too late, you've given the money out.

— Joy ▶ 46:06

If I get the SBI ledger for you now, I almost know the ticket size, the volume, the seasonality. For a company like us, that's salivating.

— Joy ▶ 42:17

I'm glad you said machine learning — in this entire conversation you've not said AI.

— Vishal Krishna ▶ 44:47

Clips that travel

Short on time? Start here

Credit teams still asking borrowers to upload statements

The API stack that replaced the PDF

The concrete inventory of India's alternative-data rails — GST via API and OTP, MCA, account aggregator, Udyam — and why abundance replaced scarcity as the problem.

9:30 → 13:23 · 4 min ▶ Watch clip

Enterprise software founders selling into regulated lenders

Why a risk engine sells to the small lender first

The unsentimental segmentation of Indian lenders, and the underwriting arithmetic that makes small unsecured tickets the only place automation pays for itself.

18:59 → 23:45 · 5 min ▶ Watch clip

Chief risk officers balancing approval rate against defaults

Automate the ends, work only the grey area

The growth-versus-loss 2x2 and the straight-through split that takes 40% of applications off human desks.

28:59 → 31:20 · 2 min ▶ Watch clip

Lenders and impact investors eyeing tier-three India

The ₹50,000 or ₹50 lakh farmer

The blind spot stated plainly, the microfinance clients working inside it, and the arithmetic that makes rural India larger than tier one and two combined.

36:11 → 40:17 · 4 min ▶ Watch clip

Anyone deploying models under a regulator's eye

Why lenders are still scared of the black box

Lending's two-year feedback lag, the auditor's question, explainable AI as the settlement, and the RBI's push from policy rules to scorecards.

44:33 → 48:24 · 4 min ▶ Watch clip

Glossary

The jargon, unpacked

Business rule engine (BRE)

The system that ingests a borrower's data, runs rules and models against it and returns a lending decision — with a front end business users can change without writing code.

NBFC

A non-banking financial company: lends from its own book under an RBI licence but cannot take deposits, and typically borrows from banks to lend onward at higher risk and higher rates.

Alternative data

Everything outside salary slips and credit-bureau records used to judge repayment capacity — GST filings, POS terminal sales, ledger apps, utility and fraud registries — the only way to size a borrower with no formal income history.

Account aggregator

India's consent-based rail that lets a borrower expose selected bank accounts to a lender through an API, replacing downloaded PDF statements and printouts.

Straight-through processing

Deciding an application end to end with no human intervention — here, auto-declining the clearly high-risk and auto-approving the clearly clean, about 40% of applications between them.

Early warning signal

A monthly model over an existing loan book that flags borrowers whose repayment risk is building before they actually miss, so the loan can be restructured while intervention still works.

Explainable AI

Model output that can be opened up after the fact — peeling back the layers to show what drove a decision — so auditors and regulators can accept a system whose maths no one can write as an equation.

Unified Lending Interface (ULI)

The proposed public rail, described on air alongside the account aggregator, meant to carry bank, GST, insurance and financial-statement data to lenders the way UPI carries payments.

Connections

If this resonated, go here next

EP 68 · Fintech

Full transcript

The whole conversation, searchable

Auto-generated captions, lightly cleaned. Click a timestamp to open that moment on YouTube.