← Return to the book

FUTURECENTRAL PRESS · BOOK SAMPLE

AI for Retail Banking

The Guide for Retail Banking Professionals

Chapter 3: AI in Personal Lending and Unsecured Retail Credit

The AI-native unsecured-credit stack, the four-outcome decision engine, and the credit officer’s loop that now holds it to account

Read the endnotes ↓

Chapter Outcomes

Unsecured personal lending is the part of the retail book where AI has gone furthest into production, and it is the chapter the rest of Part II reads against. Five questions are worth settling before the product chapters that follow.

The first is how the AI-native unsecured-credit stack is actually wired in 2026, from the data layer through the scoring models to the decision engine, so that the credit officer can see where in the flow a human judgment still enters and where it has been engineered out.

The second is what the Account Aggregator rails changed for the thin-file and gig borrower, why consented bank-statement data widened the addressable book, and where the alternative-data underwriting still breaks.

The third is how the four-outcome decision engine routes an application, what separates an auto-approval from a referral and a referral from a decline, and why the routing logic, not the score, is where the bank’s risk appetite actually lives.

The fourth is what disciplined override looks like, why automation bias is the quiet failure of a high-performing model, and what the credit officer’s loop has to deliver for the second line to sign the deployment off.

The fifth is how a bank decides to build, buy, or partner for an unsecured stack, what the fintech lending partnerships actually trade, and what the RBI Digital Lending architecture now requires of every one of those arrangements.

Opening Vignette

A salaried applicant in Indore opens a lending app at 9:40 on a Tuesday night, grants consent for an Account Aggregator to share six months of bank statements, and accepts a pre-filled personal-loan offer. The money reaches the borrower’s account before 9:42. No officer read the file, because for this borrower there was no file in the old sense: the bureau pull, the bank-statement analysis, the obligation arithmetic, the policy check, and the fraud screen all ran inside the decision engine in the time it took the app to show a spinner. The case never touched a human, and under the bank’s own policy it was not supposed to.

Two floors of a different building hold the cases that did not clear. Meena Reddy is a manager on an unsecured-credit desk at a private bank, and her queue on Wednesday morning holds the residue the engine refused to decide on its own. There is a gig-economy applicant whose Account Aggregator data shows strong but lumpy inflows the scorecard reads as instability; a salaried borrower the model declined on a thin bureau file that her eye says is a young customer rather than a risky one; and a clean-looking application the fraud model flagged because three others with the same employer and device fingerprint arrived in the same hour. None of these is a number to sign. Each is a judgment the model has explicitly handed up, and the bank answers for how she makes it.

A decade ago Meena’s desk decided everything; now it decides the exceptions, and the exceptions are harder than the average case her predecessors handled. The model takes the clean middle of the distribution, where the answer is obvious and the volume is high, and routes the tails to her, where the answer is contested and the cost of being wrong is concentrated. Her hardest daily problem is not whether to trust the model. It is when to overrule a model that has been right far more often than she has on cases that looked just like this one, and how to record the reason so that the override teaches the model rather than merely contradicting it.

Sitting above Meena’s desk is the architecture this chapter is about. The application that cleared in ninety seconds and the three that landed in her queue all moved through the same five-component stack: a data and feature layer, a set of scoring models, a decision engine that routes each case to one of four outcomes, the override loop Meena works inside, and the monitoring and governance layer that the second line of defence runs over the whole thing. Each component is a place where the bank takes a risk and a place where the regulator now expects a documented control.

The credit officer’s decision this chapter examines: in an unsecured book where the model already decides the clean majority, what is the officer actually for, and what has to be true of the override loop for the second line of defence and the regulator to accept that the bank, not the model, is making the lending decision?

3.1 What the AI-native unsecured stack actually does now

The unsecured personal-loan stack is the most fully automated decision in retail banking, and understanding it means following an application through the components rather than admiring the model at the centre. The score is one part of the flow, and not the part where most of the risk decisions are taken.

In a mature unsecured book the model decides the clean majority of cases outright, and the human decides the minority the model is engineered to escalate. For salaried applicants with a solid bureau record, straight-through processing now covers the bulk of origination at the large Indian retail lenders, with disbursal in minutes for pre-approved customers and in hours for fresh-to-bank applicants. The officer’s work has moved to the escalated tail: declines under dispute, thin and contradictory files, and cases the fraud layer has flagged. This is the inversion of the credit desk that the lending chapters build on, and it is most complete in unsecured personal credit precisely because the ticket is small, the volume is high, and the cost of a manual touch on every case would swamp the margin.

The score answers one question, and the bank takes most of its risk decisions in the questions the score does not answer. A probability of default is a ranking; it does not set the cut-off, the price, the limit, the tenor, or the documentation the case must carry, and it does not decide whether a borderline applicant is auto-declined or sent to a human. Those are policy choices the bank layers on top of the model, and they are where the institution’s actual risk appetite lives. A credit officer who treats the score as the decision has misread the stack; the score feeds the decision engine, and the decision engine, configured by the bank’s credit policy, is what lends.

Speed is a product feature with a risk cost, and the unsecured stack is built around managing that trade-off rather than maximizing speed. The ninety-second loan wins the customer who would have abandoned a slower flow, and it is a genuine commercial advantage in a market where the offer that lands first often wins the relationship. It also compresses the window in which fraud and policy controls operate to milliseconds, which is why a fast book that has not invested equally in real-time fraud and affordability controls accumulates losses that show up two or three quarters later as first-payment defaults. The discipline is to make the customer’s experience fast while keeping the controls intact behind the spinner, and the banks that confuse the two pay for it in the tail.

The stack’s real output is a routed decision with a reason attached, rather than a bare approval, and that reason is what the rest of the architecture depends on. Every case leaves the engine labelled, approved at a price and limit, declined with a reason, or referred to a named human queue, and the reason travels with it into the customer communication, the override record, and the audit trail. A stack that produces decisions without legible reasons cannot generate a compliant adverse-action notice, cannot feed the override loop, and cannot be validated by the second line. The next sections walk the components in order; the through-line is that each one has to leave a reason behind.

3.2 The data layer: bureau, Account Aggregator, and the thin-file borrower

The data and feature layer is the first component of the stack and the one that has changed most since 2021, because the Account Aggregator rails brought a new class of borrower inside the addressable book. The officer who understands the data layer understands where the model’s blind spots are.

The bureau file carries most of the weight in unsecured underwriting, and its one structural gap is the borrower with no formal-credit past. Chapter 2 set out how the bureau record works in both directions; what matters at the unsecured desk is the consequence of its coverage. For an applicant with a few years of cards and loans behind them, the CIBIL, Experian, Equifax, or CRIF file ranks repayment risk better than anything else the bank can buy. The first-time borrower and the gig worker paid into an account but never onto a loan ledger fall outside that file, and a model built on bureau features alone refuses them on the absence of evidence rather than on evidence of risk. In a market adding new-to-credit borrowers in the tens of millions, that reflexive decline is at once a fair-lending exposure and a sizeable book left on the table.

Consented Account Aggregator data is what brings the thin-file applicant inside the unsecured book, and it does so by reading how money actually moves through the account. The mechanics of the rail belong to Chapter 2; the gain that matters at the unsecured desk is specific. Six months of structured statement data let a model judge salary regularity, the balance a household holds between paydays, and the obligations already running against the account, so a borrower whose earnings are real but uneven can be priced rather than refused. This is the rail beneath most of the new thin-file lending now reaching gig and self-employed customers. The control the officer has to keep is consent itself: the statement data is licensed for a single purpose and period, and using it past either is a breach the DPDP architecture will not excuse as a growth tactic.

Alternative and device signals extend the book at the margin, and they are also where proxy discrimination enters the stack if no one is testing for it. Device, location, app-usage, and telco-derived signals can add predictive lift for the thinnest files, and several fintech lenders have built their edge on exactly these features. The risk is that an innocuous-looking feature, the device price tier, the pin-code, the handset model, correlates with a protected or sensitive attribute and lets the model decline along a line the bank would never write into policy. The officer need not build the fairness test, only recognise that an unsecured model running on alternative data without a fairness audit is carrying a fair-lending exposure the second line will eventually have to answer for.

When a sound model lends badly, the cause usually sits in the feature layer rather than in the model. Mismatched identities across the bank’s own silos, stale income fields, mis-parsed statements, and duplicate applications read as separate customers all corrupt the decision before scoring begins, and they do it without warning, because the model returns an assured score on poor inputs. Chapter 2 made this its central argument; at the unsecured desk it surfaces as a cluster of odd declines or a climb in first-payment default. The officer who meets either should read the data layer before questioning the model, since the foundation work of the first two chapters is what finally decides whether the stack lends well.

3.3 The decision engine and the four outcomes

The decision engine is the third component and the heart of the stack, because it converts a score into a lending action under the bank’s policy. The officer who can read the engine can see where the bank has set its appetite, and where it has set the boundary between machine and human.

Every application leaves the engine in one of four states, and the design of those four states is the bank’s credit policy made operational. An application is auto-approved at a set price, limit, and tenor; auto-declined with a recorded reason; referred to a human queue for judgment; or approved conditionally, subject to a further document, a step-up in verification, or a lower limit than requested. The four-outcome design is what lets a bank automate the obvious cases at both ends of the distribution while reserving human attention for the contested middle and the high-stakes tail. A book with only approve and decline states is either turning away business it should price differently or auto-approving cases it should have referred.

The boundaries between the four outcomes, not the score itself, are where the bank sets its risk appetite, and they are policy levers the credit function owns. Moving the auto-decline cut-off, widening the referral band, or tightening the conditions on a conditional approval changes the book’s loss rate, approval rate, and operating cost at once, and those are decisions the credit policy committee makes, not the data science team. The officer’s leverage here is real: the referral band is the dial that decides how much of the book a human sees, and setting it too narrow to save cost is the most common way a bank quietly hands the lending decision entirely to the model. Where the band sits is a position the credit function should be able to defend to the second line and the regulator.

The conditional-approval state is the most underused lever in the unsecured stack, and it is where a thoughtful credit policy recovers business a blunt cut-off would decline. A borderline applicant who would be declined on a single threshold can often be approved at a lower limit, a shorter tenor, or against one additional verification, converting a lost customer into a priced risk. The lenders that use the conditional state well run a materially higher approval rate at the same loss rate than those that route every borderline case to a binary decision. The officer’s contribution is to push the policy toward conditional outcomes where the risk is manageable rather than accepting a decline that protects the loss rate by shrinking the book.

The referral queue is where the model’s honest uncertainty is supposed to land, and a referral queue that is shrinking toward zero is a warning, not an efficiency win. A well-calibrated engine refers the cases it genuinely cannot resolve: the contradictory file, the novel pattern, the borderline affordability case, the fraud flag that needs a human eye. When a bank celebrates driving the referral rate down quarter on quarter, it is often celebrating having widened the auto-decision bands past the point where the model is reliable, which moves losses into the tail and removes the human check exactly where it was most needed. The officer should read a falling referral rate as a question about where the model’s uncertainty went, not as a productivity result.

3.4 Override discipline and the credit officer’s loop

The override loop is the fourth component and the one the regulator cares about most, because it is where the bank demonstrates that a human, not a model, is accountable for the lending decision. The officer’s loop is the chapter’s central subject, and getting it right is what makes the rest of the stack defensible.

Automation bias is the quiet failure of a high-performing model, and it sets in precisely because the model is usually right. When a scorecard is correct ninety-four times in a hundred, the officer who reviews its referrals learns, case by case, to agree with it, and the override rate drifts toward zero not because every case deserved the model’s call but because disagreeing feels like swimming against the evidence. At that point the human in the loop has become a rubber stamp, and the bank has the accountability of a manual process with the judgment of an automated one. The discipline is to keep the override a live act of judgment, which means measuring the override rate, reviewing the cases where the officer agreed with a model that later defaulted, and treating a zero-override desk as a control failure rather than a model success.

An override is only worth anything if it carries a reason the system can learn from, and an unreasoned override is noise the model will eventually overwrite. When an officer overturns a model decline, the value is not the single rescued customer; it is the recorded reason, which tells the model-risk team whether the model has a systematic blind spot, the gig-income pattern it misreads, the thin-file young customer it over-penalizes, that the next model version should fix. An override loop that captures decisions without structured reasons gives the officer a veto but gives the bank no learning, and over time the model drifts back toward the very errors the officers were correcting by hand. The officer’s reason code is the mechanism by which human judgment improves the machine rather than merely arguing with it.

The override loop is also where the adverse-action obligation is met, and a decline the bank cannot explain is a decline the bank cannot lawfully make. Whether the case was auto-declined or declined after referral, the customer is owed a reason that is accurate to the actual basis of the decision, and a model whose declines cannot be rendered into honest reason codes fails this test however well it ranks risk. This is the point where explainability stops being a technical nicety and becomes a lending requirement: the reason in the adverse-action notice, the reason in the override record, and the reason the model actually used have to be the same reason. Chapter 15 develops the explainability machinery; here the officer’s takeaway is that an unexplainable decline is not a defensible one.

The second line will not sign off a stack whose override loop it cannot audit, so the loop has to be built for inspection from the start. Model-risk management, compliance, and internal audit need to see the override rate over time, the distribution of override reasons, the outcomes of overridden cases, and the evidence that officers are exercising judgment rather than deferring. A first line that treats the override record as its own operational data, rather than as the evidence the second line and the regulator will read, is building a loop it will have to rebuild under supervision later. The officer who documents the loop as if it will be audited, because it will be, is doing the work that lets the bank run the model at speed.

3.5 Build, buy, or partner: the bank and the fintech lending stack

No bank assembles the entire unsecured stack alone, and the build-buy-partner decision is where the credit function meets the technology and compliance functions on shared ground. The officer who understands what each arrangement trades can tell which partnerships are sound and which import a risk the bank cannot see.

The fintech partnership trades the bank’s balance sheet and licence for the fintech’s origination and decisioning, and the value of the trade depends entirely on who controls the credit policy. Arrangements with lenders such as Bajaj Finance in its own right, and the digital-credit models of players like KreditBee and Slice, run the spectrum from the bank simply funding a fintech’s book to the bank owning the credit policy while the partner provides the customer interface. The arrangement is sound when the bank sets and can audit the credit policy, the cut-offs, the data use, and the collections conduct, and unsound when the bank funds a book whose decisioning it does not control and cannot inspect. The officer’s question for any partnership is simple: whose credit policy is this, and can the bank prove it owns it?

The RBI Digital Lending architecture has rewritten what a lending partnership is allowed to look like, and several older arrangements do not survive the rewrite. The Digital Lending rules require that the loan be disbursed directly into the borrower’s bank account rather than passing through a lending-service-provider’s pool account, that the Key Fact Statement disclose the all-in cost before the borrower commits, that a cooling-off period let the borrower exit without penalty, and that default-loss-guarantee arrangements with partners stay within the prescribed cap. A partnership built on pass-through disbursal or undisclosed costs is not a commercial preference the bank can keep; it is non-compliant, and the officer evaluating a partner needs to read the structure against these requirements rather than against the partner’s growth deck.

Building in-house buys control and costs speed, and the right answer differs by segment rather than holding for the whole book. A bank with a strong data foundation and a mature model-risk function can build an unsecured stack it fully controls, which is the right call for its core salaried book where the data is rich and the volume justifies the investment. For a thin-file or new-segment book where the bank lacks both the data and the specialized models, a partnership that the bank governs tightly reaches customers the in-house stack cannot, at the cost of a control burden the bank must actually carry. The officer’s contribution is to resist the single-answer instinct: the book is a portfolio of segments, and build, buy, and partner can each be right for a different part of it.

Whatever the arrangement, the accountability does not move with the origination, and a bank that forgets this discovers it under supervision. When a partner originates a loan the bank funds, the borrower is the bank’s borrower, the conduct obligations are the bank’s obligations, and a fair-practices or data-use failure by the partner is the bank’s failure in the regulator’s eyes. The Digital Lending architecture is explicit that the regulated entity owns the compliance outcome regardless of who built the interface, which means the officer’s due diligence on a partner’s decisioning and conduct is the bank’s own risk management, not an optional courtesy. The partnership can outsource the work; it cannot outsource the answer.

Framework: The Retail Credit Underwriting Architecture

The Retail Credit Underwriting Architecture is the five-component map this chapter uses to hold the unsecured stack together, and it is built to be placed over the reader’s own book so the officer can name where each risk is taken and each control sits. The components run in sequence, and a weakness in an early one cannot be repaired by a later one.

Component one is the data and feature layer. This is the bureau record, the Account Aggregator consented data, the alternative and device signals, the KYC and internal customer data, and the feature pipeline that turns them into model inputs. The officer’s control point is data quality and consent: a confident model on bad or improperly-consented data lends badly and unlawfully, and most surprising losses trace to this layer rather than to the model.

Component two is the scoring-model layer. This is the application score, the behavior and bureau-derived models, the affordability assessment, and the fraud models that run alongside them. The officer’s control point is knowing what each model can and cannot see, and reading a model’s score as a ranked input to a decision rather than as the decision itself.

Component three is the decision engine. This is the policy layer that routes each application to one of four outcomes, auto-approve, auto-decline, refer, or conditional approve, at a set price, limit, and tenor. The officer’s control point is the boundaries between the outcomes, because the referral band and the conditional state, not the score, are where the bank’s risk appetite is actually set.

Component four is the override and human-in-the-loop layer. This is the officer’s loop: the referral queue, the override decision, the structured reason code, and the adverse-action communication. The officer’s control point is keeping the override a live act of judgment that captures learnable reasons, because this layer is where the bank demonstrates that a human owns the lending decision.

Component five is the monitoring and governance layer. This is model-risk management, drift and performance monitoring, fairness testing, and the documentation the FREE-AI and Digital Lending architecture requires, run by the second line over the whole stack. The officer’s control point is building the first four components for inspection, because the speed at which the bank can run the stack is set by how much of it the second line can actually govern. Used together, the five components let the officer answer the question the chapter opened with: where, in this book, does a human still decide, and can the bank prove it?

Sidebar: Failure Modes — How a High-Performing Unsecured Stack Goes Wrong

Override decay. The model is usually right, the officer drifts toward agreeing with it, and the override rate falls toward zero. The bank now carries the accountability of a manual process with the judgment of an automated one. The tell is a desk that has not overturned a model decision in months; the fix is to measure override rates and review agreed-with cases that later defaulted.

Thin-file drift. A model tuned on a salaried book is extended to gig and new-to-credit segments without retraining, and it misreads irregular-but-sound income as instability, declining good customers and approving the wrong ones. The tell is a rising first-payment-default rate in a new segment; the fix is segment-specific models and a referral band wide enough to catch what the model has not learned.

Proxy discrimination. An alternative-data feature correlates with a protected or sensitive attribute, and the model declines along a line the bank would never write into policy. The tell is a decline pattern that tracks geography or device tier; the fix is a fairness audit before deployment and reason codes that expose the actual basis of each decline.

Partnership pass-through. A funded fintech book runs on pass-through disbursal, undisclosed cost, or an out-of-cap loss guarantee, and the bank discovers under supervision that it owns a non-compliant arrangement it did not control. The tell is a partnership the bank cannot fully audit; the fix is to read every arrangement against the Digital Lending architecture before the book is funded, not after.

India Cases

Case 1, HDFC Bank’s unsecured retail engine: scale built on a controlled stack. HDFC Bank runs one of the largest unsecured retail books in Indian banking, and the bulk of its personal-loan and pre-approved origination is decisioned through a model-driven stack with disbursal in minutes for eligible customers. The lesson for the credit officer is that the bank’s scale is a function of control, not of automation alone: the pre-approved engine works because the bank owns a clean customer record, a mature scorecard, and a policy engine it can audit, and the human desk concentrates on the escalated tail rather than the clean middle. A bank that wants HDFC-scale unsecured volume without building the underlying data and governance is asking for the volume without paying for the foundation.

Case 2, Axis Bank’s decisioning consolidation: one engine instead of many. Axis Bank’s work to consolidate its retail decisioning onto a unified platform illustrates a problem most large banks face: unsecured products that grew up separately accumulate separate scorecards, separate cut-offs, and separate override practices, so the same customer can be read differently by two desks in the same bank. Consolidating the decisioning onto one engine is what lets a bank apply a consistent policy, audit a single override loop, and present the second line with one stack to govern rather than five. The officer’s lesson is that fragmentation is itself a risk: a book decisioned by inconsistent engines cannot be governed coherently, however good each individual model is.

Case 3, the fintech partnerships, Bajaj Finance, KreditBee, and Slice: where origination and balance sheet meet. The Indian unsecured market runs substantially on arrangements between banks and digital lenders, from Bajaj Finance originating on its own balance sheet to digital-credit players such as KreditBee and Slice building the customer interface and decisioning for books that banks may fund or co-lend. The lesson the credit officer should carry is that these arrangements are sound exactly to the degree that the regulated entity owns and can audit the credit policy and the conduct, and that the Digital Lending architecture has moved several older structures out of bounds. The partnership that survives scrutiny is the one where the bank can prove it owns the lending decision; the one that does not is a funded book the bank does not control.

Global Cases

Case 4, Upstart: thin-file underwriting on alternative data, and the fair-lending test that comes with it. Upstart built a US lending business on underwriting thin-file and near-prime borrowers with models that use education, employment, and other alternative signals alongside the traditional bureau, partnering with banks that fund the loans. Its history is the clearest global illustration of the thin-file opportunity and its fair-lending cost together: the alternative-data model reaches borrowers a bureau-only lender would decline, and it draws sustained regulatory attention to whether those same features produce disparate outcomes along protected lines. The lesson for the Indian officer is that the Account Aggregator and alternative-data book carries the Upstart trade, more inclusion and a standing fairness obligation, and that the obligation does not wait for the regulator to raise it.

Case 5, Klarna: buy-now-pay-later as a credit decision made invisible. Klarna built a global buy-now-pay-later business by embedding a real credit decision inside a checkout that the customer never experiences as a loan application, approving or declining in the moment of purchase on thin data and behavioral signals. The model is instructive precisely because it shows the unsecured decision compressed to its limit: the speed that wins the sale also removes the friction that traditionally slowed a borrower into affordability, and the losses concentrate where the underwriting was thinnest. The lesson for the credit officer is that the faster and more invisible the credit decision becomes, the more the affordability and conduct controls have to be engineered into the flow rather than left to the borrower’s caution, a point the Indian regulator has made central to its own digital-lending architecture.

Regulatory Landscape: Unsecured Retail Credit

Unsecured retail credit is the most heavily regulated AI surface in the retail book, and the two layers below set the floor the credit officer’s stack has to clear. This box is the chapter’s flagship regulatory treatment; Chapters 14 through 16 develop the architecture in depth.

Global Layer. The global frame for AI in consumer credit consolidated around fair-lending and model-risk expectations that the Indian officer should read as the direction of travel. In the United States, the Equal Credit Opportunity Act and Regulation B require an accurate adverse-action notice for every decline, and the Consumer Financial Protection Bureau has stated that a complex or AI-based model is no excuse for vague reasons: the lender must give the actual, specific basis of the decision, which makes explainability a lending requirement rather than a technical preference. The European Union’s AI Act classifies creditworthiness assessment and credit scoring as high-risk, attaching documentation, human-oversight, and data-governance obligations and administrative penalties for the most serious violations. The model-risk discipline articulated in the United States Federal Reserve’s SR 11-7 remains the reference standard for validating, monitoring, and governing the models a credit stack runs on. For the Indian credit officer, the message across all three is consistent: a decline must be explainable, a credit model must be governed, and the speed of the stack is bounded by the institution’s ability to evidence both.

India Layer. The Indian architecture for AI-driven unsecured credit is now dense, and the credit officer’s stack is read against it case by case. The RBI Digital Lending framework, issued as Guidelines in 2022 and consolidated as the RBI (Digital Lending) Directions, 2025, requires that loans be disbursed directly to the borrower’s bank account without pass-through pool accounts, that a Key Fact Statement disclose the all-in annualized cost before the borrower commits, that a cooling-off period allow penalty-free exit, and that default-loss-guarantee arrangements with partners stay within the prescribed cap. The FREE-AI principles of 2025 set the governance expectation for the models inside the stack, naming what the second line of defence must document and oversee. The DPDP Act of 2023 governs the consented data the stack runs on, making purpose-limitation on Account Aggregator and customer data a compliance boundary rather than a growth lever. Credit-information-company reporting discipline governs how the resulting accounts are reported to the bureaus that the next model will read. Four questions reach the second line together here: can every decline be explained, is every model under governance, is every data use consented and purpose-bound, and does every partnership hold up against the Digital Lending structure. Co-lending arrangements, increasingly used to fund the unsecured book, are separately governed by the RBI (Co-Lending Arrangements) Directions, 2025, effective 1 January 2026, which require each lender to retain at least 10% of every co-lent loan, cap the originator’s default-loss guarantee at 5% of the portfolio, mandate a blended interest rate to the borrower, and require both lenders to classify a loan identically, extending the regulated-entity-owns-the-arrangement principle to the co-lending book.

Practitioner’s Lens: The AGM (Retail Credit Risk) at a private sector bank

The AGM running retail credit risk at a private bank sits where the model, the policy, and the regulator meet, and the role is the clearest vantage on what the credit officer’s job becomes when the model decides the clean majority. The AGM does not build the model and does not approve the individual loan; the AGM owns the policy the engine runs and the loop the humans work inside.

The AGM’s first instrument is the decision engine’s configuration, because that is where the book’s risk appetite is actually set. The cut-offs, the referral band, the conditions on the conditional state, and the segments routed to human review are the levers the AGM moves, and a change to any of them ripples through the loss rate, the approval rate, and the operating cost at once. The AGM reads the weekly book as a set of policy outcomes, not model-performance charts: where the engine is auto-approving too widely, where the referral band has been narrowed past prudence to save cost, where a conditional state would recover business a blunt decline is turning away.

The AGM watches the override rate as closely as the loss rate, because a zero-override desk is a control failure the loss rate will not reveal until later. When the override rate on a referral desk drifts toward zero, the AGM reads it as automation bias setting in, not as the model being vindicated, and the response is to review the cases where officers agreed with a model that later defaulted and to retrain the desk to treat the override as a live act of judgment. The AGM knows that a desk which never overrides is not saving the model from error; it is removing the human check exactly where the bank told the regulator the human check would sit.

The AGM’s hardest conversations are with the business heads who read the referral queue as a cost rather than as a control. A product head under volume pressure wants the referral band narrowed and the conditional state simplified, because every referral is an officer’s minute and every condition is a point of friction in the funnel. The AGM’s job is to hold the line that the referral queue is where the model’s honest uncertainty lands, and that narrowing it to hit a volume number moves losses into the tail and removes the accountability the bank promised the second line. This is a credit-risk conversation conducted in the language of funnel conversion, and the AGM has to keep it tied to the loss the narrowing will produce two quarters out.

The AGM is the person who has to make the override loop legible to the second line, and that work is never finished. Model-risk management and compliance need the override rate, the reason-code distribution, the outcomes of overridden cases, and the evidence of genuine judgment, and the AGM is accountable for producing all of it in a form the second line can audit and the regulator can read. The AGM treats the override record as the bank’s evidence rather than the desk’s operational data, because the day a supervisor asks how the bank knows a human is making the lending decision, the answer is the override loop or it is nothing.

The AGM’s twelve-month agenda is dominated by the partnerships and the segments the in-house stack does not reach. The salaried core book runs well on the bank’s own engine; the growth and the risk both sit in the thin-file, gig, and partnership books where the data is newer and the governance is harder. The AGM spends a disproportionate share of the year on whether the bank owns the credit policy in each partnership, whether the new-segment models have been validated and fairness-tested, and whether the Digital Lending structure holds across every funded arrangement, because that is where the next supervisory question, and the next loss, is most likely to come from.

Applied Exercise: The Line’s Underwriting Stack Map

The deliverable for this chapter is a one-page map of the credit officer’s own unsecured stack across the five components, a working document for finding where the bank takes a risk without a matching control. Work the five steps in order; alone it takes about two hours, and with a colleague from model risk, half a day.

Step one, the data layer. Name the sources the unsecured book underwrites on, bureau, Account Aggregator, alternative and device signals, internal data, and for each note the main blind spot and the consent basis. Close with the one data-quality or consent issue most likely to be causing surprising declines or first-payment defaults in the book today.

Step two, the model layer. List the scoring, affordability, and fraud models the book runs, and for each name the segment it was tuned on, the date of last validation, and one thing it cannot see. Flag any model being used on a segment it was not trained for, because that is where the thin-file drift in the Failure Modes sidebar begins.

Step three, the decision engine. Write down where the four outcome boundaries currently sit, the auto-approve and auto-decline cut-offs, the width of the referral band, and the conditions on the conditional state, and note who owns each boundary. Mark the one boundary that has moved in the last year to hit a volume or cost target, and what it did to the tail.

Step four, the override loop. Record the current override rate on each referral desk, whether overrides carry structured reason codes, and whether the reasons feed the next model version. Name the desk closest to a zero-override rate, because that is the desk where automation bias is most advanced and the human check is most nearly gone.

Step five, the governance layer. Note what the second line currently sees of the stack, the override data, the fairness tests, the drift monitoring, the Digital Lending evidence, and what it does not. Close the map with the single component where the bank takes a risk the second line cannot yet inspect, because that is where the officer’s next quarter of work pays off.

Executive Briefing

Once the model clears the clean majority, the real question is what the credit officer is now for, and the answer sits in the override loop rather than the average case. The model has taken the obvious cases at both ends of the distribution and handed the officer the contested tail, which is harder than the average case the desk once handled. The officer’s value is now concentrated in the referral queue and the override decision, and a desk that treats those as residual work rather than as its core function has misread where the bank now takes its unsecured risk.

The override rate is the number to watch this quarter, because a desk drifting toward zero overrides is a control failure the loss rate will not show until later. Automation bias sets in because the model is usually right, and the officer who reviews referrals learns to agree with it until the human check is a rubber stamp. The move this week is to measure the override rate on each desk, review the agreed-with cases that later defaulted, and treat a zero-override desk as a problem to investigate rather than a productivity result to celebrate.

The referral band is a credit-policy lever the credit function owns, and the pressure to narrow it for volume is the most common way a bank quietly hands the lending decision to the model. Every narrowing of the band looks like an efficiency win and lands as a loss two or three quarters out, because the cases that stop being referred are exactly the ones the model could not reliably resolve. The conversation to have with the business heads is that the referral queue is where the model’s honest uncertainty is supposed to land, and that the band’s width is a risk-appetite decision, not a funnel-conversion setting.

Every decline has to be explainable, because a decline the bank cannot put into an honest reason code is a decline the bank cannot lawfully make. The reason in the adverse-action notice, the reason in the override record, and the reason the model actually used must be the same reason, and a model that ranks risk well but cannot be rendered into accurate reasons fails the test whatever its performance. The officer’s near-term work is to confirm that the unsecured stack can explain its declines to the standard the adverse-action obligation requires, and to flag any model that cannot.

The partnerships and the new segments are where the next supervisory question and the next loss are most likely to come from, so that is where the year’s attention belongs. The salaried core book runs well; the thin-file, gig, and fintech-partnership books carry both the growth and the governance risk, and the Digital Lending architecture has moved several older partnership structures out of bounds. The standing question for every funded arrangement is whether the bank owns and can audit the credit policy, because the accountability for the loan never moves to the partner, however the origination is built. Chapter 4 turns to secured lending and the mortgage stack, where the calibration of model against human shifts again at the higher ticket.

Endnotes

1. Meena Reddy is a composite, assembled from publicly reported facts about unsecured-credit operations at Indian retail banks; she stands for no single person or bank, and the case details are illustrative.

2. On straight-through processing and minutes-to-disbursal in Indian unsecured retail lending, see the annual reports and digital-lending disclosures of HDFC Bank, ICICI Bank, and Axis Bank for the FY2022–FY2025 period.

3. On the Account Aggregator consent framework, live from 2021, and its application to cash-flow-based and thin-file underwriting, see Reserve Bank of India and Sahamati materials on the AA ecosystem.

4. On India’s credit information companies (TransUnion CIBIL, Experian, Equifax, CRIF High Mark) and bureau-reporting discipline, see the RBI Master Direction on Credit Information Reporting and the bureaus’ public disclosures.

5. On the RBI Digital Lending framework, issued as Guidelines in September 2022 and consolidated as the RBI (Digital Lending) Directions, 2025, including direct-to-borrower disbursal, the Key Fact Statement, the cooling-off period, and the default-loss-guarantee cap, see the relevant Reserve Bank of India publications.

6. On the default-loss-guarantee (DLG/FLDG) arrangement and its prescribed 5% cap, originally introduced by the Reserve Bank of India guidance of June 2023 and since consolidated into the RBI (Digital Lending) Directions, 2025, see the relevant Reserve Bank of India publications.

7. The FREE-AI committee report (Responsible and Ethical Enablement of Artificial Intelligence, Reserve Bank of India, 2025) sets the governance expectations referenced here for models inside the credit stack.

8. On the DPDP Act of 2023 and its purpose-limitation obligations on consented data, whose implementing Rules were notified in November 2025 with obligations phasing in across 2026 and 2027, see the Government of India and Reserve Bank of India publications.

9. On HDFC Bank’s unsecured retail engine and pre-approved origination, see HDFC Bank annual reports for the FY2022–FY2025 period.

10. On Axis Bank’s retail decisioning consolidation, see Axis Bank annual reports and investor disclosures for the FY2021–FY2025 period.

11. On Bajaj Finance’s balance-sheet origination and the digital-credit models of KreditBee and Slice (the latter following its merger with North East Small Finance Bank), see the respective companies’ public disclosures and press coverage of the period; specific partnership structures are characterized in general terms.

12. On Upstart’s alternative-data underwriting, bank-partnership model, and the associated fair-lending scrutiny, see Upstart public disclosures and US regulatory commentary, including CFPB statements on adverse-action obligations for AI-based credit models.

13. On Klarna’s buy-now-pay-later model and embedded credit decisioning, see Klarna public disclosures and trade-press coverage of the period.

14. On US adverse-action requirements under the Equal Credit Opportunity Act and Regulation B as applied to complex models, see the Consumer Financial Protection Bureau circular on adverse-action notification (2022). On the EU AI Act’s high-risk classification of credit scoring, see Regulation (EU) 2024/1689. On model-risk governance, see US Federal Reserve Supervisory Letter SR 11-7 (2011).

Key Terms

Straight-through processing (STP). Origination in which the decision engine approves and disburses without a human touch, used for the clean majority of unsecured cases so that human attention concentrates on the escalated tail.

Decision engine. The policy layer that converts a model score into a lending action, routing each application to one of four outcomes at a set price, limit, and tenor. It is where the bank’s risk appetite is operationalized.

The four outcomes. Auto-approve, auto-decline, refer to a human queue, and conditional approve. The boundaries between them, not the score, are where the bank sets its appetite.

Referral band. The range of cases the engine sends to a human rather than auto-deciding. A band narrowed to save cost quietly hands the lending decision to the model and moves losses into the tail.

Conditional approval. An outcome that approves a borderline applicant at a lower limit, shorter tenor, or against additional verification, recovering business a blunt cut-off would decline.

Automation bias. The drift of a human reviewer toward agreeing with a usually-correct model until the override becomes a rubber stamp. A zero-override desk is a control failure, not a model success.

Override loop. The component in which an officer reviews a referral, overturns or confirms the model, records a structured reason, and generates the customer communication. It is where the bank evidences that a human owns the decision.

Reason code. The structured, accurate basis of a decline. It must match the reason the model actually used, because the same reason has to serve the adverse-action notice, the override record, and the next model version.

Adverse-action obligation. The requirement to give a declined applicant an accurate, specific reason. A model whose declines cannot be rendered into honest reason codes cannot lawfully make those declines.

Key Fact Statement (KFS). The standardized disclosure of the all-in annualized cost of a loan that the Digital Lending framework requires before the borrower commits.

Default-loss guarantee (DLG/FLDG). A partner’s guarantee to absorb a capped share of a funded book’s losses. The Digital Lending framework caps the arrangement and requires it to be structured transparently.

Lending service provider (LSP). A partner that performs origination or servicing on a regulated lender’s behalf. The accountability for the loan and its conduct remains with the regulated lender, not the LSP.

Reflection Questions

Each question assumes the reader runs or supports an unsecured book, keeps a calendar, and means to apply the architecture rather than discuss it.

Lay the officer’s own unsecured book over the five components of the Retail Credit Underwriting Architecture. At which component does the bank currently take a risk without a matching control, and what would closing that gap cost in the next quarter?

Find the referral desk in the book closest to a zero-override rate. Is that desk handling genuinely easy cases, or has automation bias set in? What review would tell the difference, and who should run it?

Identify one boundary in the decision engine, a cut-off, the referral band, or a conditional-state condition, that has moved in the last year to hit a volume or cost target. What did the move do to the tail, and was the credit function or the business head the one who actually decided it?

Take one unsecured decline the book made this month and trace whether the reason in the adverse-action notice, the reason in the override record, and the reason the model used are the same reason. If they are not, where does the chain break?

Pick one segment the book underwrites on Account Aggregator or alternative data. What fairness test currently runs on it before deployment, and if none does, what is the bank’s exposure and who owns it?

Take one fintech partnership the bank funds or co-lends into. Can the bank prove it owns and can audit the credit policy, and does the structure meet the Digital Lending requirements on disbursal, the Key Fact Statement, and the loss-guarantee cap? Where is the weakest point?

Consider the AGM (Retail Credit Risk) Lens. Of the five components, which does the officer’s own second line see most clearly today and which least, and what would it take to make the override loop fully legible to them?

Further Reading

For the regulatory spine of this chapter, the essential reading is the RBI Digital Lending Directions (2025) and the 2022 Guidelines they consolidated, the FREE-AI committee report of 2025, and the DPDP Act of 2023, read together rather than separately, since the unsecured stack is governed by all three at once. The IIBF Bank Quest journal and IDRBT research are the working banker’s standing references for the operational and technical detail, and the credit officer preparing for IIBF examinations should map these readings to the relevant credit and digital-banking modules.

For the global frame on AI in consumer credit, the Consumer Financial Protection Bureau’s circulars on adverse-action notification for complex models are the clearest statement that explainability is a lending requirement, and the EU AI Act’s high-risk treatment of credit scoring is the direction of travel for documentation and oversight. The model-risk discipline in the US Federal Reserve’s SR 11-7 remains the reference for governing the models a credit stack runs on, and the McKinsey and BCG retail-credit analytics reviews of 2023 through 2025 give the working consultancy view of where the stack is heading.

For the credit-modeling substrate beneath the AI conversation, Naeem Siddiqi’s writing on scorecard development is still where practitioners learn how application and behavior scorecards get built and validated, and it anchors the officer’s sense of what a score does and does not capture. For the affordability and fair-lending dimension, the regulatory commentary on alternative-data underwriting is the reading that connects the thin-file opportunity to its standing obligation.

Chapter 4 turns to mortgages and home-loan underwriting, where the higher ticket, the property valuation, the legal-title check, and the relationship layer shift the calibration of model against human away from the straight-through pattern this chapter described. The officer who has mapped the unsecured stack across the five components will see immediately which components carry forward to the secured book and which are replaced.