The most dangerous credit model is not the one that crashes or produces an absurd result. It is the one that continues approving borrowers, rejecting applicants, and pricing risk with confidence, long after the environmental conditions and population attributes that once made those decisions reliable have changed. And it happens more often, and sometimes more prominently, than we realise; for instance, cyclones or poor monsoons can fundamentally shift risk understanding.
The Reserve Bank of India's draft Guidance on Regulatory Principles for Model Risk Management, released on June 24, closed for public comment on July 24. The final version is still pending, but its direction seems already settled. The draft widens the definition of a "model" well beyond AI and machine learning systems to cover rule engines, scoring algorithms, and even the spreadsheets that, with some cajoling from underwriters, decide who gets a loan and at what price. It asks for board-approved oversight across the full lifecycle of every such model, a three-lines-of-defence structure separating those who build models from those who validate and audit them, and the ability to switch a model off. On finalisation, the guidance will replace the credit-risk-model chapter of RBI's 2002 Guidance Note on Credit Risk Management, a rulebook written for a lending world without AI in it.
That reframes the questions around AI governance. A lender must now determine what evidence would make it stop trusting a model, who has the authority to act, and how credit decisions continue safely once intervention begins.
{{/usCountry}}That reframes the questions around AI governance. A lender must now determine what evidence would make it stop trusting a model, who has the authority to act, and how credit decisions continue safely once intervention begins.
{{/usCountry}}In our experience, credit models rarely fail through one dramatic event. Borrower populations change, sourcing channels attract different customer profiles, partnerships start driving volumes, macroeconomic conditions alter repayment behaviour, new technology upends older methods of onboarding, and once-predictive variables lose relevance. The system may appear normal even as the relationship between its predictions and real borrower risk weakens.
Waiting too long for defaults to reveal that deterioration is risky because credit has a delayed feedback loop. By the time delinquency confirms that a model has weakened, an entire lending vintage may already have been originated. Oversight, therefore, needs to detect weakening confidence before portfolio losses provide definitive proof.
Standard statistical checks, such as population and characteristic shift measures, rank-ordering tests, and calibration reviews, remain important, but none can answer the intervention question alone. A model may still separate relatively good borrowers from bad ones while underestimating absolute risk, or remain stable overall even as the population entering it changes. Model drift is real too. Even widely used thresholds need judgement; a common shift-index cut-off is an industry convention, not a universal line between a safe and unsafe model. Governance risks ending up as a dashboard exercise where one metric turns red and somebody presses a button.
The draft guidance pushes in the opposite direction by naming seven specific AI risk dimensions lenders must track: Explainability, hallucination, bias, overfitting, spurious correlation, output variability, and data risk. That list is useful precisely because it forces institutions to look for compound signals rather than a single alarm. Unexpected movements in approval rates, unexplained rejection patterns, shifts in score distributions, weakening rank ordering, and altered variable behaviour, appearing together, tell a more reliable story than any one of them alone. Increasing disagreement between experienced underwriters and the model can be equally revealing.
Having a human in the workflow is also not the same as having meaningful human control. A credit officer who can technically override an AI recommendation but lacks the information, confidence, time, or institutional authority to do so is not exercising genuine oversight.
Human overrides should become part of model monitoring. If experienced underwriters repeatedly reverse a model for similar cases, those interventions are evidence about how the model is behaving. Conversely, a high-impact model facing virtually no challenge may reflect strong performance, or automation bias.
That leads to the hardest governance question: If independent model validation believes a live underwriting model should be restricted while the business wants originations to continue, whose judgement prevails?
The answer cannot be negotiated after deterioration begins. Institutions need predefined escalation rights and intervention stages. Early warnings may require closer monitoring, while more serious deterioration may justify tighter cut-offs, lower exposure limits, manual review, or restrictions on certain segments. Suspension should become necessary when data integrity, model performance, or institutional confidence falls beyond an agreed level; the "kill switch" the draft guidance now asks boards to hold, and be prepared to use.
Those decisions must also weigh the cost of being wrong in either direction. Keeping a deteriorating model live can scale poor lending decisions across an entire portfolio; suspending a functioning model can disrupt credit access and business continuity for borrowers who did nothing wrong. Intervention must, therefore, consider model materiality, exposure, borrower impact, and how quickly harm could accumulate.
Fallback decisioning is inseparable from the ability to stop a model. Material credit models need tested alternatives, the last validated model, a conservative rules engine, restricted eligibility, or additional manual underwriting. Those alternatives must also work at real lending volumes; otherwise operational pressure will force institutions back to a model they no longer trust. The tussle will push the modernisation and inclusion drive backwards, because a lender who loses confidence in a model without a workable fallback usually defaults to the manual underwriting AI was meant to replace, and thin-file borrowers are the first casualty of that retreat. The draft's requirement to retain decommissioned models and their records for ten years is a tacit admission of the same point: A model taken offline may need to be understood, or reinstated, long after the decision to stop it was made. Sometimes, history is the way forward: the model retired last year, with its audit trail intact, may be exactly the fallback a stressed portfolio needs next year.
The same discipline applies to third-party models. Accountability cannot disappear into outsourcing. Lenders still need enough visibility, contractual control, and internal capability to understand material changes, challenge performance, and restrict or stop use when required, whatever the vendor agreement says. While FLDG, the first-loss default guarantee that co-lending partners often provide, offers credible risk coverage, the risk nuance and assessment need to be the regulated entity's primary concern; a guarantee covers losses, not governance.
The scale that makes AI attractive for lending also magnifies governance failure. AI can widen access for thin-file and new-to-credit borrowers, but a mistaken assumption embedded in a model can spread across an entire portfolio far faster than a conventional underwriting error.
RBI is now weighing industry feedback on the draft, and a final circular is expected within the coming months. Whenever it lands, the test it sets will not be how many models a lender deploys. It will be whether the institution knows what evidence should trigger intervention, who has the authority to act, and how lending continues safely once that authority is exercised. A lender that cannot answer those questions does not have humans in command of AI. It merely has humans watching it.
(The views expressed are personal)
This article is authored by Amit Das, founder & CEO and Suryadip Ghoshal, co-founder & chief analytics officer, Think360.ai.