The CIBIL score is useful. It is also backward-looking, self-referential, and designed for a borrower who has already has taken loans before.A borrower who has never taken a loan from a regulated institution has no CIBIL score to show. A borrower who took one consumer loan five years ago and repaid it has a thin file that is technically active but informationally limited. A gig worker with Rs 40,000 monthly income has excellent cash flow discipline and no credit score at all.

World Bank research shows that combining bureau data with alternative data improves predictive accuracy by up to 25% for thin-file borrowers. That improvement is not marginal. For a lender running 50,000 new-to-credit originations per month, a 25% improvement in default prediction accuracy translates directly into lower provisioning, lower write-offs, and a more competitive rate offered to the borrower.

The bureau data lag adds a second problem. While RBI's 2025 mandate has proposed moving bureau reporting to weekly frequency from fortnightly currently. It still means a lender cannot see a borrower's current position.
Account Aggregator data, pulled in real time with consent, solves this. A borrower who drew down unsecured personal loans from three lenders last week will not show it in the bureau file for up to seven days. AA will show it the moment the transactions clear.
Every lender that has tried to use a bureau score to underwrite a gig worker or a first-time two-wheeler buyer has run into the same wall. The score is missing, thin, or out of date. The borrower is creditworthy. But the risk model says NO lending. This problem is growing larger, as the average age of the borrower drops. And the demographics ensures that growth is happening in the GenZ Segment, who is entering the workforce. This customer is earning but not does not have the credit score to get loans.

The CIBIL score is not the problem. The problem is building a credit model that can say yes when the bureau file says nothing.

Availability of Alternative data in India has really increased in India, as the economy gets formalised. Of course, each data set has its own consent requirement, its own regulatory status, and its own predictive value depending on the borrower segment. Additionally, with the implementation of DPDP getting borrower consent and maintaining the consent has become very important. Each data source carries its own consent architecture under the DPDP Act and they are not interchangeable.

Account Aggregator cash flows are the most powerful signal available. Twelve to twenty-four months of bank statement data - multiple bank accounts in one go, salary credits, EMI debits, recurring expenses, cash withdrawals tell a credit model what no single income document can. AA helps us understand how a borrower manages money under varying conditions. Seasonal income dips, one-time large expenses, rent payment regularity, and average balance trends are all visible. The consent constraint is firm: AA data pulled for credit assessment is purpose-limited under the DPDP Act. It cannot be used for cross-sell, top-up pre-approval, or collections communication without fresh, specific consent.

UPI and NACH transaction behaviour serve as proxy income and reliability signals. A borrower who maintains a NACH mandate for investments, pays utilities via UPI consistently, and receives stable inflows from a single merchant ID is demonstrating repayment behaviour without having formally repaid a loan. AI models assess thin-file borrowers using UPI transaction behaviour, NACH mandate performance, GST filing regularity, and Account Aggregator bank statement analysis as primary inputs. This data is what powers Paytm/Phonepe’s lending business, a daily view into the merchants cash flows.

GST filing consistency is the primary alternative signal for MSME underwriting. A business that files GST returns on time, shows stable turnover, and has consistent input credit claims is demonstrating financial discipline across three independent axes simultaneously.

Voice interaction data is Bajaj Finance's most distinctive contribution to this list — and the one that no textbook on alternative data in lending has addressed directly.

Bajaj Finance's Q2 FY26 investor presentation highlighted AI-driven analytics that extract "meaningful data variables" from audio recordings to enhance underwriting and customer engagement. By Q4 FY26, this had reached 52 million voice interactions converted to structured text. A data layer built from every customer call, every Personal Discussion note, every collection conversation. The insight being mined is not what the borrower says. It is the pattern across what borrowers in similar situations have said, and what happened to those loans.

This is not a standard alternative data input. Bureau data, AA flows, and GST records are structured by design. Voice interaction data is unstructured, proprietary, and requires an LLM layer to extract variables that a credit model can use. Bajaj Finance has 203 dedicated AI personnel and a 115 million customer base generating call data at scale. The resulting underwriting signal is not replicable by a lender that started this process last quarter.

Trust signals and geolocation data are the most contested category. Companies are training models on location details, third-party app usage, SMS data, payment transaction behaviour, and metadata to enhance underwriting for new-to-credit customers. Fraud syndicate detection stems from analysing patterns such as multiple Aadhaars linked to a single mobile number.

This is where the RBI’s FREE-AI framework comes to play. AI models used for credit must be explainable, auditable, and free of discriminatory bias. A model using geolocation as a proxy for creditworthiness must document that the variable is not acting as a demographic proxy for caste, income class, or geography of historical credit exclusion.

Every credit team building an AI-first underwriting model faces a version of the same question: logistic regression scorecard or ML ensemble model?

I don’t think this is an either/or decision. It could be both at different stages of the decisioning architecture, with clearly separated accountability.

The traditional scorecard is a logistic regression model translated into a points-based system. Its primary virtue is interpretability: a credit officer can explain precisely why a borrower scored 680 rather than 720 . The FOIR ratio contributed this many points, the enquiry count reduced it by that many and so on. This explainability satisfies RBI's Fair Practices Code requirement that declined borrowers should have definitive reasons behind the decline. It also satisfies the operational need for a credit team to calibrate, override, and audit the model without a data scientist in the room.

ML ensemble models with gradient boosting, random forests, XGBoost typically outperform logistic regression on prediction accuracy, particularly for thin-file and new-to-credit borrowers where variable relationships are non-linear. A borrower with irregular income but very low expense volatility is a different risk profile than the scorecard's income-to-obligation ratio captures. The ensemble model learns this. The scorecard cannot.

In practice, risk teams can adopt a layered strategy: use ML models for challenger modelling and performance benchmarking, deploy interpretable models for production scoring with periodic reviews and model audits, and introduce governance frameworks around ML retraining, monitoring, and fairness testing.

This is the classic champion-challenger architecture. The champion model runs on 100% of live originations and makes all binding decisions. The challenger runs on 10% to 20% in shadow mode its decisions are logged, never enacted, and compared to the champion's outcomes after 60 to 90 days of portfolio ageing. When the challenger outperforms the champion on accuracy, fairness metrics, and operational stability, it gets promoted. (The governance requirement is that "outperforms" must be defined and documented before the challenger goes live, not negotiated after it does.) This process is continuous there is always a champion, always a challenger, and always a documented reason why the incumbent is still winning.

SHAP values (Shapley Additive Explanations ) are the mechanism that makes ensemble models explainable enough for production use. They quantify each variable's marginal contribution to a specific credit decision. A lender can show that a borrower was declined as it was affected by cash flow volatility in the last three months and two outstanding loans opened in the past 60 days. That explanation satisfies the Fair Practices Code. In short, SHAP is the auditor standing next to your model ensuring it is giving the correct outputs and explaining the same to the regulator. It is essential to deploy and monitor the same to ensure consistent outputs.

The choice is not scorecard or ML model. The choice is whether you have built the architecture to know, at any point, which of your models is better and whether you can prove it to RBI.

In the last edition, I did mention Bajaj’s FINAI strategy, it will be a recurring theme in this series. Simply because its is by far the most documented journey of AI implementation at scale. Usage of AI in Underwriting and credit is a major part of the FINAI strategy. The data foundation comes first. By Q4 FY26, Bajaj Finance had completed 52 million voice-to-data conversions and 2.3 million text-to-data conversions, generating structured data from customer interactions for AI-driven underwriting insights. This is not simply a large number. It represents a structured data asset built from conversations that previously produced nothing more than call logs. The variables being extracted e.g. sentiment patterns, stated financial situations, product understanding, intent signals feed into both underwriting enrichment and portfolio early warning systems.

The underwriting efficiency targets are explicitly quantified. The underwriter productivity improvement target is 30% by FY27, driven by AI-generated summaries of Personal Discussion notes. A credit officer reviewing a PD note summary instead of writing one spends that time on credit judgement rather than documentation which is the correct allocation of human capacity in a hybrid AI-human underwriting process.

The ECL model governance deserves a separate mention. Bajaj Finance conducts an annual model refresh incorporating recent portfolio performance and forward macro outlook. The Q4 FY25 refresh resulted in an additional Rs 359 crore provision. Yes, it is a significant number, but the consequence of a model that had not been updated frequently enough. The lesson drawn internally was institutionalised fast: the FINAI data layer, with feature marts, embeddings, and continuous data annotation across voice, text, and video, is designed so that the ECL model and the underwriting model are both trained on current portfolio performance rather than just annual snapshots. When this is fully live in FY27, model drift will surface faster and corrections will be smaller.

Bajaj Finance's AI governance framework, completed in Q4 FY26, directly aligns with RBI's FREE-AI principles. The fraud detection layer at origination deploys Vision AI, Voice AI, and network anomaly detection inline catching application fraud at the point of submission rather than after disbursement. These are not separate compliance systems. They are embedded in the underwriting workflow as design features.

Bajaj Finance's credit stack is notable not for the AI it deploys but for the data architecture it has spent three years building to support that deployment.

L&T Finance reported that its Net Non Starters in Two-Wheeler Finance dropped from an indexed 100 to 11 in 14 months. That is the number that matters, and it deserves deeper examination.

Net Non Starters are borrowers who default before making their first payment, it is the most unambiguous credit quality metric in the origination function. A borrower who never repays was never a viable credit risk. Moving this metric from 100 to 11 means the model is reliably distinguishing genuine credit demand from fraudulent or unviable applications at the point of origination, before disbursement.

L&T Finance’s Project Cyclops is a three-dimensional credit engine. It evaluates bureau data for historical credit behaviour, liabilities data via Account Aggregators for liquidity insights, and trust signals. This is statistically validated data from payment gateways, geolocation intelligence, and banking transactions reflecting income and consumption patterns. For customers with robust credit histories, bureau data suffices. For those with limited or no credit history, Cyclops relies heavily on these trust signals.

The architecture is deliberately multi-source because no single data source is sufficient for the segments L&T Finance wants to reach. The "statistically validated" qualifier on trust signals is the critical phrase. Signals are tested for genuine predictive power before being included in the production model. They are not simply ingested because they are available. That validation step separates a compliant alternative data model from a governance liability.

Cyclops is now live in Two-Wheeler Finance, Farm Equipment Finance, SME Finance, and Personal Loans, with Home Loans and Rural Group Loans scheduled for FY27. The NTC improvement of 34% in approval rates was achieved while simultaneously reducing NNS meaning the model is approving more borrowers and approving better borrowers at the same time. This is the result that pure scorecard optimisation cannot produce, because the scorecard's training data does not include the borrowers it historically refused.

L&T Finance has built the post-origination stack to match. Project Nostradamus handles portfolio monitoring. Project Helios is the AI Underwriting Co-Pilot live in SME Finance. Project Orion is the automated conversational assistant for portfolio intelligence. The same underlying data model runs through all four — which means the origination decision, the portfolio monitoring signal, and the collections trigger share a common data architecture. The feedback loop from collections into the credit model is closed by design.

L&T Finance has not built an underwriting model. It has built an underwriting architecture Cyclops for origination, Nostradamus for portfolio management, Helios for SME co-piloting, Orion for conversational portfolio intelligence. The same data runs through all four. That is the point.

L&T Finance and Poonawalla Fincorp have built parallel fraud layers. Poonawalla's 57-project AI programme covers fraud detection across multiple workflows. L&T Finance's data architecture flags inconsistencies in trust signals at application stage rather than at portfolio review.

The regulatory framing matters here. RBI's FREE-AI principles require explainability in fraud decisions. A borrower whose application is rejected on a fraud flag must, in principle, be informed in plain language why. That is harder than it sounds. A black-box neural network that flags fraud but cannot say why is not deployable under FREE-AI. The fraud models being built into Indian origination stacks today are, by design, interpretable. The model says fraud and points to the specific signal. Signature anomaly score X. Geolocation mismatch Y. Document tampering confidence Z.

Inline fraud detection is the only origination AI use case where the bar is set by the regulator, not the lender. FREE-AI makes interpretability a deployment condition.
(Will deep dive into Fraud in the next edition)

From Bajaj Finance's data-first approach, L&T Finance's Cyclops architecture, and the FREE-AI governance requirements, a working picture of AI-first credit underwriting in India emerges. It has four layers.

Data ingestion with consent mapping. Every data source — bureau, AA, GST, voice interaction data, trust signals — has a corresponding consent record specifying the purpose for which it was collected. The data layer and the consent layer are not separate systems. Consent records are queryable alongside the credit decision, so an audit can reconstruct exactly what data was available to the model, under what consent, at the time of every decision.

A champion model with documented explainability. The production credit model — whether scorecard or SHAP-augmented ensemble — generates a reason code for every decision. The reason code is expressed in plain language. The data science team owns accuracy. The risk team owns fairness testing. The compliance team owns the audit trail. Three accountability domains, separated by design — because the incentive to deploy a more accurate but less explainable model is real, and someone needs to have the authority to say no.

A live challenger infrastructure. A challenger model runs on every origination cohort in shadow mode. Its decisions are logged but not enacted. After 60 to 90 days of performance data, it is evaluated against the champion on accuracy, fairness, and operational stability. Promotion from challenger to champion requires documented sign-off from risk, compliance, and credit governance. No single team promotes a model unilaterally.

A collections feedback loop. The outcome of every loan : repayment performance, delinquency timing, recovery rate feeds back into the model's next training cycle. L&T Finance's connected architecture between Cyclops, Nostradamus, and the collections stack operationalises this. Bajaj Finance's continuous voice-to-data conversion means the feedback includes not just repayment outcomes but the qualitative signals from customer interactions through the loan lifecycle. A model that never learns from its own portfolio is not an AI credit model. It is a static scorecard with better marketing.

The credit function of an AI-first lender is a governance architecture. The model is the output. The governance is the system that ensures the output is fair, explainable, and improving.

The CIBIL score will not disappear. Bureau data remains the most reliable signal for borrowers who have formal credit history — which is why Cyclops uses it as one of three dimensions rather than replacing it.

The structural change is that bureau data alone is no longer sufficient to underwrite the Indian credit market at the scale it needs to reach. India's alternative lending market is projected to grow from $26.74 billion in 2024 to $52.30 billion by 2029. That growth is almost entirely in segments the bureau score cannot serve: gig workers, MSME proprietors, agricultural borrowers, first-time credit users.

Bajaj Finance has spent three years building the data architecture that makes AI-first credit underwriting possible at scale. L&T Finance has the Cyclops NNS numbers to prove the model works. The lenders who build the multi-dimensional credit stack now, have to do it with consent architecture, bias testing, and champion-challenger governance. They will own those segments when the market doubles.

The lenders who wait for the bureau file to fill in will still be waiting.

Sources:

Bajaj Finance Q2, Q3, Q4 FY26 investor presentations and earnings call transcripts;
L&T Finance — Project Cyclops documentation and Q4 FY26 results;
RBI FREE-AI Committee Report (August 2025);
PwC/Dvara Research/FACE — Principles of RTAI in Digital Lending (March 2026);
World Bank research on alternative data and thin-file borrowers;
Business Standard — Alternative Data in Indian Lending (October 2025);
RBI Utkarsh 2029 (March 2026)