Risk isn’t only who you lend to. It’s what you offer.

A bureau score answers one question: how likely is this person to repay? It answers in three digits, and decades of credit infrastructure have been built around that number.

But no lender lends to a person in the abstract. It lends a specific amount, at a specific rate, over a specific tenure, in a specific product. Each of those moves the odds. The borrower who comfortably repays ₹40,000 over twelve months may not repay ₹1,00,000 over three. A higher rate can tip a budget that was only just holding. A bullet loan asks something different of a month’s cashflow than an EMI or a credit line.

The score is silent on all of it, so policy fills the gap: a flat ticket, a multiple of income, one rate card for everyone who passes. Numbers a credit committee agreed on and nobody has revisited.

This is not an oversight. It is a measurement problem, and a genuinely hard one.

A lender’s own book contains almost no variation in the terms of a loan that is independent of risk — because the lender’s own policy chose them. Safer applicants got more, for longer, at a better rate. So if you fit a model of default against amount, rate or tenure on your historical book, you do not learn how risk moves with the offer. You recover the policy you already had, with a confidence interval around it.

Mentat is an AI underwriting model for Indian lenders, built by TARA AI Labs. It returns a default probability not once per applicant, but for each offer you could make them — every amount, rate, tenure and product construct, predicted together. The output is a risk surface, not a score. Your policy ceiling stays exactly where it is: you choose the offer that sits inside it, and price the loan to the risk it actually carries.

It works off the applicant’s repayment sequence rather than the score that summarises it. A score is months of behaviour — who paid, when, how late — compressed into three digits. That compression is good at ranking borrowers against each other. It throws away precisely the information that tells you how a borrower’s risk moves when the offer changes.

Mentat learns from credit bureau data — 33 million loan accounts and 1.5 billion repayment events — never from a lender’s book. A lender’s data is used only to score that lender’s applicants.

One variable at a time: here only the amount moves, on a bullet loan, for five illustrative borrower profiles — synthetic, drawn with the same function as the Mentat page, not real applicants. Each row is one profile’s default probability as the amount rises. Note the shape: for most borrowers it runs close to flat across a wide range and then bends, sharply, once the repayment crosses what their cashflow carries. The bend sits somewhere different for every applicant, and the rate, the tenure and the product each move it again. Anyone showing you a smooth upward slope is showing you a model, not a borrower.

Take the Intent-strong profile in the chart above — illustrative, not a real borrower. Bureau 745, which clears almost any policy in the market. Income ₹38,000 a month.

A policy engine offers ₹1,14,000 — three times income. At that amount, this profile’s default probability is 15%, because the repayment lands well past what the cashflow carries.

Mentat offers ₹75,000, where the same profile’s default probability is 7.5%.

Same profile. Same bureau report. Same day. Twice the default risk, for half as much again. No score can express that, because the risk is not a property of the borrower alone. It is a property of the offer.

It runs the other way too. The Capacity-strong profile scores 685, below the 700 cutoff, so a score-based policy declines it. Its cashflow carries more than the score suggests: Mentat funds ₹95,000 at 7.8%, inside the same ceiling.

The same holds when the rate or the tenure moves. Between two offers, the borrower does not change. The risk does.

On 400,523 historical leads from a listed lender, scored against that lender’s own live policy as the control:

Those are gains at the same policy ceiling, on the same leads. The lender’s risk appetite did not move. The offers did: the amount and the rate, priced to each borrower’s risk.

That is a backtest on their historical book, not a realised result on yours. The model reports 0.84 AUC on held-out Indian retail credit data — an internal evaluation, not independently audited. We publish the population, period, outcome definition and baseline in a methodology note, and we would rather you read it than take the number.

It does not replace your policy. The ceiling, the exclusions and the pricing floor stay yours. Mentat tells you where each applicant and each offer sits beneath them.

It does not claim that risk simply rises with the amount. The curve is mostly flat, then it kinks, and finding the kink is the point.

And it does not pretend the measurement problem away. Mentat is calibrated within bands of each variable, and we will tell you exactly where the data supports the surface and where it is extrapolating. On a first engagement we run your own declined and approved files and show you, case by case, what a different offer would have changed — because that is the only evidence that should convince you.

Credit in India is moving to the moment of purchase: a credit line on UPI, pay-later at checkout, a loan approved in a minute. Each needs a risk decision before the payment clears, and each is a different offer — a different amount, tenure and product. A single score was never built to price that.

Hundreds of millions of Indians are assessed by a bureau that already holds their repayment history but reads only part of it. Most are not turned away. They are handed an offer with no relationship to what they could carry, and counted as a credit risk when it breaks.

Measure risk on the whole offer, and a lender no longer has to choose between saying no and lending too much. It can say yes to the right loan.