Every AI underwriting pitch sounds roughly the same. Faster decisions. Broader data signals. Thin-file borrowers were approved where bureaus would have drawn a blank. Better risk separation, lower cost, fewer humans in the loop. Some of this is accurate.
What rarely makes it into the pitch is the other half of the story. Credit cycles are indifferent to training data. Models optimised for benign conditions fail in ways that are both predictable and underestimated. Edge cases sit permanently outside any model's distribution. Human judgment has not been made redundant. It has just been inconveniently moved further from the decision.
Lenders who have deployed AI at scale understand both sides. Those who deployed it carelessly got an education on the second side, usually during a collection cycle that the model did not see coming.
Three areas have seen real, measurable change, and they are worth acknowledging clearly before getting to the limits.
Speed at Scale
A well-trained model processes thousands of applications simultaneously, applying consistent logic without fatigue, variance, or the mood of whichever analyst is on shift. Cost per decision falls, turnaround compresses, and what was a headcount problem becomes a compute problem. For retail and MSME portfolios at scale, that shift compounds meaningfully over time.
Pattern Recognition Across Variables
Scorecard-based underwriting works with a short list: bureau score, income, existing obligations, repayment history. AI models surface non-linear relationships across thousands of variables that no scorecard would find. Twelve months of cash flow patterns carry more predictive signal than a bureau score taken on a single date, and a well-trained model extracts that signal consistently across every application.
Thin-file Underwriting
This is where AI earns its most consequential contribution in markets like India. Salaried workers in the informal economy, first-time borrowers, small traders with no formal accounting — a significant share of creditworthy borrowers are simply invisible to bureau-based underwriting. Alternative data changes this.
UPI transaction patterns, GST filing history, utility payment regularity, and supply chain flows each construct a credit picture for borrowers who have none on paper. AI is what makes this data actionable at volume. Without it, thin-file underwriting stays a manual exercise that cannot be run profitably at scale.
Every capability has a boundary. In credit underwriting, AI's boundaries tend to appear at the worst possible moment.
The Credit Cycle
No AI vendor's deck addresses this cleanly: models are trained on historical data, and credit cycles are structural and recurring. A model trained during low defaults learns the correlations of that environment. When rates rise, liquidity tightens, or a sector shock arrives, those correlations may stop holding. Nothing in the model's architecture flags this. It keeps approving.
Pavitra Pradip Walvekar, a fintech operator who has underwritten credit through multiple market conditions and built lending infrastructure from the inside, makes the point without qualification: AI does not see regime shifts coming. It sees its training data and extrapolates from it. A lender treating model output as a reliable compass during a market transition is substituting pattern recognition for judgment. Those are not the same thing.
Model Overconfidence in Regime Shifts
In steady states, well-trained models are often more accurate than human analysts. That is genuinely true and worth acknowledging. What is also true is that those same models become confidently inaccurate when the regime shifts. Confidence does not waver. Precision appears unchanged. Scores arrive with the same apparent authority they always have, while the underlying assumptions have already broken down quietly in the background.
Disciplined lenders do not ask whether the model performs well in normal conditions. It does. They ask what happens when the delinquency curve moves in a direction the model has not seen, and whether the lender has the infrastructure to detect that movement before the portfolio damage compounds into something structural.
Edge Cases and Human Judgment
Every credit portfolio carries borrowers who sit outside the distribution the model was trained on. A business owner whose revenue is real but whose accounting is informal. A borrower with a thin bureau file and a strong six-month transaction history. A loan purpose that does not map cleanly to the model's categories. These cases are too frequent to ignore and too varied to train on reliably.
Human judgment is a permanent layer of the underwriting stack, not a fallback for when the model breaks down. A lender who has removed human review entirely has not automated underwriting. They have automated approvals and eliminated the mechanism that catches what the model structurally cannot see.
Lenders who have deployed AI well do not use it as a replacement for underwriting judgment. They use it as a triage engine that sorts decisions by complexity and routes each to the right review layer.
Clear approvals and clear declines are automated. A borrower with strong bureau data, consistent cash flows, and a loan purpose inside the model's distribution gets a decision in seconds. Models are highly consistent in these cases, and the cost of an occasional error is manageable within the portfolio economics.
Cases at the margin, or those carrying signals the model weights poorly, go to human review. Human attention is finite and should concentrate where it generates the most value. An experienced analyst reviewing a hundred complex cases contributes more to portfolio quality than reviewing a thousand routine ones alongside a model that has already decided.
Circuit breakers are pre-defined triggers that pause or constrain automated decisioning when portfolio indicators move outside expected bounds: a delinquency rate crossing a threshold in a segment, an approval rate diverging from historical norms, a loan category that has started performing differently from expectations. Most lenders underinvest in this layer. It is also the one that matters most during a deteriorating cycle.
When a circuit breaker fires, the correct response is to stop. Examine whether the training conditions still hold. Decide whether the model needs to be retrained, recalibrated, or temporarily set aside in favour of tighter manual review. This governance layer is what separates lenders who use AI responsibly from those who discover its limits inside a deteriorating book.
AI has made credit underwriting faster, broader, and more accessible to borrowers who were previously invisible. These are genuine improvements, and lenders who have not incorporated AI into their underwriting stack are operating at a structural disadvantage in high-volume segments.
Credit risk itself is unchanged. Borrowers default when circumstances shift. Portfolios deteriorate when macro conditions move. Models fail when the environment they were trained on stops resembling the one they are operating in. None of these dynamics has been solved. They have been deferred, sometimes elegantly, sometimes dangerously.
Pradip Walvekar's position, shaped by operating through multiple credit cycles, is that the discipline gap is wider than the technology gap. Most lenders have access to capable AI tools. Fewer have built the governance infrastructure around those tools to know when to trust them and when to override them. That gap is where durable books are built or lost.