The Wrong Question
For the better part of two decades, the lending industry has been asking the wrong question. The debate has been framed as scoring versus judgement, algorithm versus credit officer, automation versus discretion. That framing misses the point entirely.
The more useful question is simpler and more uncomfortable: is the borrower in front of you actually scoreable? That question has been avoided because it does not suit the vendor pitch, the investor presentation, or the enthusiasm of institutions eager to cut turnaround time and headcount. But it is the question that the data is now forcing us to answer.
Every wave of credit automation has arrived with more promise than preparation. Scorecards in the 1980s and 1990s. Machine learning models in the 2010s. Automated credit decisioning engines through the last decade. And now, AI. Each wave delivered real value in the right contexts. Each wave also produced damage when institutions moved faster than the evidence warranted. The pattern is not a coincidence.
What Scoreability Actually Means
Parameterisation works where input variance is low. Salaried income arrives on a fixed date, in a predictable amount, from a verifiable source. Mortgage repayments follow a schedule. Supply chain invoices against confirmed purchase orders have clear cash flow visibility. These are narrow distributions. Clean data. Predictable borrower behaviour over time. Scoring works here not because the algorithms are clever but because the underlying data and boundaries are well-behaved.
The self-employed or marginal credit borrower- a SME owner, rural borrower, gig worker, thin-file first-time credit seeker. Income is irregular, is a different problem. Multiple income streams do not consolidate neatly into any single data source. Cash flows are seasonal, lumpy, and sometimes informal. The variance in these populations is structurally high. No algorithm, however sophisticated, converts a noisy signal into a reliable prediction without the error rates beyond tolerance. Scoreability is a property of the borrower in a specific context, not of the model.
This distinction matters enormously in the Indian lending context. The country’s formal credit infrastructure was built around the salaried urban borrower, especially after 1991. The real growth opportunity, including the financial inclusion imperative, lies overwhelmingly with borrowers outside that profile. Applying scorecards designed for one population to another is not innovation. It is a category error. The noise in lending fintech’s over the last decade has proven this as has the continued ‘tradtionalism’ of larger banks in commercial lending, including SME.
How the Models Evolved
The history of credit scoring is a history of successive model generations, each solving the previous generation’s limitations while introducing new ones of its own.
The first generation was built on linear discriminant analysis and logistic regression. LR-based scorecards dominated from the 1970s onward and remain the regulatory default in many jurisdictions today. Their advantage is interpretability: the model’s decisions can be read, explained, and justified. A credit officer or regulator can understand why a loan was declined. This matters in regulated environments where adverse action notices are a legal requirement. There are limitations though. LR assumes that variables have fixed, relatively stable functional relationship independent effects on creditworthiness, and that the relationship is linear. Non-linearity and interactions can be incorporated, but they must usually be designed in advance rather than discovered freely from the data. In complex borrower populations with interacting risk factors, those assumptions fail consistently.
The second generation introduced ensemble methods like Random Forest, XGBoost, LightGBM, gradient boosting machines. These models do not assume linearity or variable independence. They discover interaction effects across hundreds of features that no analyst would have thought to specify. Research across multiple datasets consistently shows 10 to 25 percent accuracy improvement over LR on the same credit data. XGBoost and LightGBM became industry workhorses through the 2010s, particularly in digital lending and fintech underwriting.
But the real leap was not algorithmic. It was the shift from static to adaptive. A scorecard trained once and deployed for three years degrades as borrower behaviour shifts. The meaningful change was retraining frequency. Rolling retrain cycles kept models current with recent portfolio experience. This solved the lag problem that plagued first-generation models.
It also introduced a new problem. Adaptive models can drift silently. The model keeps scoring. The outputs look statistically normal. But what it is actually measuring has quietly shifted as the borrower population or macro environment changes. A multi-year study of credit scoring models found that during economic stress, income and employment status became dominant predictors, displacing credit score itself in importance. The model’s internal logic changed without anyone noticing. Institutions monitoring only accuracy metrics would have missed it entirely. This is concept drift, and it is an unresolved governance problem in deployed ML scoring.
The regulatory response was XAI (Explainable Artificial Intelligence). SHAP (SHapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations) emerged as post-hoc interpretation layers. You run the XGBoost model, then run SHAP over it to reconstruct which features drove each decision. This became the compliance workaround: high-performing opaque model, interpretability layer on top. It is a functional solution to a real regulatory requirement. But it is not a solution to model reliability. Recent research has documented that SHAP values are themselves unstable under class imbalance and concept drift, the precise characteristics of stressed credit portfolios. Beyond technicality, the moot question remains – does this approach give enough regulatory confidence?
The Fintech Experiment
The last decade was a live, large-scale test of pushing parameterisation into thin-file, small-ticket, quick-credit lending. The results are now visible in the data.
India’s digital lending ecosystem grew rapidly through the 2010s and into the post-COVID period. Fintech NBFCs and lending apps scaled aggressively into segments that traditional banks had avoided- thin-file borrowers, gig workers, blue collar, New to Credit (NTC) and in Tier 2, Tier 3 and deeper geographies. The experimentation was around alternative data as a substitute of formal credit history or predictable cash flows like salary. Models trained on UPI transaction patterns, GST filings, utility payments, bank statement analysis. This was further stretched to SMS, Social Media and other behavioural data sources.
The models have worked well during the growth phase. What the models did not have was stress-cycle data. They had never seen their borrower populations under genuine macro pressure. What compounds the issue is the shape-shifting nature of the stress itself. No two recessions are similar.
When conditions shifted in 2023 and 2024, the evidence arrived quickly. Small-ticket personal loans below Rs 10,000 showed a 44 percent jump in delinquencies among loans originated in a single six-month window between December 2023 and June 2024. Unsecured retail lending came to account for over half of new NPAs in the retail portfolio in the first half of FY25. The RBI flagged high delinquency in sub-Rs 50,000 loans as a category requiring close supervisory monitoring. The models were disbursing. They were not assessing. That is not a model limitation. That is a design choice about what the model was trained to optimise.
Several specific regulatory interventions punctuated this period. In October 2024, the RBI barred four NBFCs from sanctioning new loans, ostensibly citing excessive interest spreads and inadequate monitoring. I infer much of this gap was possibly driven by failure to adequately assess income, repayment capacity and shape-shifting risks despite having technology-driven underwriting systems in place.
The broader pattern across the sector was consistent: models trained on growth-phase data with no stress-cycle observations, optimised for speed and approval rates, deployed into borrower segments with structurally high variance. The portfolio deterioration that followed was not sudden. It was a slow, sector-wide convergence on the same outcome, visible only in retrospect when the RBI tightened risk weights on unsecured lending in November 2023 and the growth machine slowed.
The experiment produced a finding, not a failure. Parameterisation in small-ticket lending is conditionally valid. The conditions are a. sufficient training data including at stress cycle b. data infrastructure that reflects real-time exposure, and c. models designed to assess repayment capacity rather than to approve disbursals. My biggest lament with ‘fintech’ models has been on first point. Where does a new lender get a large enough set of stress-cycled data?
Where It Actually Works
The fintech experiment did not prove that scoring fails. It proved that scoring fails when deployed outside its valid range. There are contexts where parameterisation performs well and where the case for automation is strong.
Pre-approved loans and cross-sell limit management are the clearest examples. Here the bank already owns longitudinal relationship data like account behaviour, transaction patterns, repayment history across products. The problem being solved is not fresh underwriting of an unknown borrower. It is limit adjustment for a known one. The model is working with clean, consistent, long-duration data on a borrower whose behaviour it has already observed. That is a fundamentally different statistical problem, and scoring handles it well.
Portfolio analytics is another area of genuine strength. Early warning systems, vintage analysis, cohort-level delinquency prediction and concentration risk monitoring, are applications where ML models add value because the data is structured, historical, and well-labelled. The model is not making a binary approve-or-decline decision about an individual borrower. It is identifying patterns across a portfolio that human analysts would miss.
Consumer lending at scale has also benefited meaningfully from scoring especially in Credit cards, Salaried Personal Loans, Home Loans for documented borrowers. The borrower profile is relatively homogeneous, the data sources are standardised across the industry, and bureaus have accumulated decades of repayment history including multiple stress cycles. Models trained on this data have seen recessions, job losses, and rate shocks.
The relative success of surrogate products like Average Bank Balance and GST multiples also shows the differentiated impact of long-duration, structured data.
The distinction the industry collapses repeatedly is between scoring as a portfolio deepening tool and scoring as an acquisition underwriting tool. They are not the same instrument. Deploying the former in the latter role with ‘tweaks’, discounts the large context difference between the two. This is where the problems begin.
AI Changes the Equation. But Not in the Way You Think
The arrival of Large Language Models (LLMs) and AI-driven reasoning systems genuinely changes something in credit decisioning. The change is real and significant.
ML models adapted to changing conditions through retraining frequency. AI can do something qualitatively different: it can reason over unstructured context. A management interview transcript. A news article about a borrower’s sector facing tariff changes. A stock exchange filing. A field officer’s visit report. These are not inputs that any scoring model currently processes. AI can synthesise them into a credit-relevant narrative. That is a genuine capability expansion, not a marginal improvement.
But the problems this creates are also genuine, and they are different in character from the problems of earlier model generations.
The first problem is hallucinated confidence. An LLM produces coherent, authoritative-sounding credit narratives even when the underlying data is thin. A scorecard with insufficient data returns a low confidence score or declines to score. An AI system may still make assumptions, extrapolations to return a well-structured analysis. In a credit context, that is dangerous precisely because it sounds like judgement.
The second problem is inherited bias. AI systems are trained on historical data that overrepresents English-language, large-enterprise, urban financial contexts. Borrower types like a Tier-3 MSME borrower, a freelancing professional, a cooperative member, a rural agri-processor, are systematically under-represented in training distributions. The model will reason, but it will reason from the wrong priors. And unlike a scorecard where the bias is visible in the feature weights, AI bias is embedded in the model’s reasoning patterns and is much harder to surface.
The black box problem does not get better with AI. It gets worse. SHAP and LIME provided imperfect but functional windows into ML model decisions. The interpretability of transformer-based models is an open research problem. Regulators globally are already flagging this. The BIS and IAIS have both noted that AI model opacity creates governance challenges that existing model risk management frameworks were not designed to handle. RBI’s own articulations point towards these risks.
The realistic take is that AI stands today as a capable aid to the credit manager, not a replacement for credit judgement. It can surface information faster. It can process more context than any human analyst. But it cannot yet be trusted as the sole decisioner in complex, high-variance credit situations. Not because the technology lacks ambition or isn’t evolving, but because the governance infrastructure to deploy it safely does not yet exist at most institutions. More importantly, the speed at which AI is evolving, governance itself can’t catch up. Not until the dust settles.
History is worth noting here. The wave of automated credit decisioning engines that excited institutions through the 2010s delivered real efficiency gains. It also produced portfolios that deteriorated when conditions changed because the governance came after the deployment. AI is arriving with the same institutional excitement and similar unpreparedness.
Future Refuses to Become a Feature
There is a structural limit to credit parameterisation that better algorithms and more data do not solve. All current models, including AI, are backward-looking by design. Every predictive model learns principally from observed history. Forward-looking information can be added, but its reliability, comparability and predictive weight are much harder to establish.
A credit scoring model ingests what has happened. Bureau data, transaction history, repayment patterns, bank statement flows. These are records of past behaviour. They predict future behaviour by assuming continuity. In stable environments, that assumption holds. In dynamic ones, it breaks down at exactly the moment when accurate credit assessment matters most.
A garment exporter who has automated cutting and stitching has a fundamentally different competitive position than one operating entirely manually. That difference affects export capacity, margin structure, and resilience to labour cost increases. It does not appear in any bureau score or transaction pattern. A mid-sized logistics company whose anchor client is moving its distribution hub, is invisible to any scoring model until the receivables stop arriving. The AI adoption level of a borrower’s sector, the quality of the management team, the intangibles of loyalty and reputation, a regulatory shift in a key market and calibrated impact of geo-political unpredictability aren’t baked into any credit scoring architecture.
This is not a data availability problem. It is a model design challenge. The models were not built to ingest these signals because converting them into structured, trainable features is genuinely hard. You can count for error rates, but the tolerances are often breached. Until that problem is reliably solved, the most consequential credit judgements will still need the guardrail of human-in-the-loop.
Real Decision Is Where Not to Automate
The credit officer of today is not making the same judgements as the credit officer of 2000. The tools and data available are substantially more powerful but the judgements that matter most are still the ones that require reasoning about ‘touch and feel’ of credit, anticipation of what will happen at entity level. These subjective interpretations fall a bit outside the pattern matching against what has happened. That reasoning requires domain knowledge, sector understanding, and the kind of contextual intelligence that takes years to build and cannot be retrained on a rolling quarterly cycle. AI may reach there but there is still time.
But human judgement in credit is not the romantic alternative to algorithms. It has its own flaws of bias, inconsistency, and mood swings. However, it is the necessary bridge while institutions build the governance, data discipline, and model oversight that responsible AI-driven decisioning actually requires. That bridge does not disappear because the technology improves. It shifts in character as AI evolves.
Institutions that got into difficulty in the lending cycle of the past decade were not uniformly reckless. Many were running models that were technically sound and internally consistent. What they lacked was a governance layer that asked the right questions. Were the models being run on borrower populations they were designed for? Was there stress-cycle data in the training set? Was the model optimised for the right objective? Was someone watching for drift? These are not technology questions but one of governance and transparency.
The AI wave arriving now will be faster, more capable, and more opaque than anything that preceded it. The vendors will be more convincing. The investor presentations will be more compelling. The peer pressure to adopt will be harder to resist. The consequences of deploying unprepared will scale with the technology.
The governance response isn’t static either. Its evolving with localised Small Language Models (SLMs), domain-specific RAG, mandatory citations and confidence thresholds. These controls will not remove judgement risk, but they can make AI use narrower, more traceable and more governable.
Scepticism about AI is not the safeguard. Discipline about deployment is. That means understanding which borrowers are actually scoreable within the loan, ticket, economic cycle bounds, before choosing a model. It means building governance before scaling. It means keeping human judgement load-bearing in the credit decisions that matter most, not as a fallback for when the model fails, but as a designed and valued part of the process. It is also means realising that the role and weight of human in credit will change and the skills required will need an upgrade as well
Credit is not a technology problem that still needs humans. It is a judgement problem that technology can help solve. That distinction matters more than ever.
Technical Glossary
- Logistic Regression (LR) A traditional statistical method widely used in credit scorecards. It studies borrower information, such as income, repayment history and debt levels, to estimate the likelihood of default.
- XGBoost and LightGBM
Machine-learning techniques that analyse large numbers of variables and identify complex patterns in the data. They often perform better than traditional scorecards, but their decisions can be harder to explain. - Random Forest
A machine-learning method that uses many decision trees instead of relying on a single one. Each tree reaches its own conclusion, and the model combines these results to produce a final prediction. - SHAP and LIME
SHAP, or SHapley Additive exPlanations, and LIME, or Local Interpretable Model-agnostic Explanations, are tools that help explain why a machine-learning model reached a particular decision. They show which inputs influenced the result and in what direction. They make the output easier to interpret, but they do not prove that the model is accurate, fair or reliable - Large and Small Language Models (LLMs and SLMs)
LLMs are trained on vast amounts of text and can perform a wide range of tasks, while SLMs are smaller models designed for narrower, more specific uses. SLMs generally require less computing power and can be easier to control, test and deploy for defined credit processes. - Concept Drift
A situation in which the patterns learned by a model change over time. For example, income may become a more important predictor of default during an economic slowdown. Unless the model is regularly monitored, it may continue producing scores even though its original assumptions are no longer valid.
References:
- Ayari et al. (2026): ML-powered credit scoring-systematic literature review, Springer Artificial Intelligence Review
- Chang et al. (2022): XGBoost and LightGBM vs logistic regression in P2P lending- Quantitative Finance and Economics
- Ballegeer et al. (2025): Stability of SHAP and LIME explanations in cost-sensitive credit scoring- European Journal of Operational Research
- BIS FSI Insights No. 63 (2024): AI regulation, model risk, and governance challenges in financial services
- BIS FSI Occasional Paper No. 24 (2025): How regulators can address AI explainability in financial services
- CRIF High Mark and DLAI (2025): FinTech Barometer, Vol. II- delinquency trends in small-ticket personal loans
- FACE (2025): FinTech Personal Loans: April 2018–March 2025 – digital lending growth and portfolio delinquency trends
- RBI Financial Stability Report, December 2024: Unsecured retail-credit growth, household leverage, and emerging areas of stress
- RBI action against Navi Finserv and three other NBFCs, October 2024: Excessive pricing and weaknesses in household-income and repayment-capacity assessment
- RBI circular on consumer credit, November 2023: Higher risk weights and tighter oversight of unsecured consumer lending
