Hello everyone, 

Welcome to the 209th edition of The Pattern. 

Last week, Deutsche Bank's investment-bank CIO told a Bengaluru tech event that AI is reshaping how fast the bank works. Projects that used to take two years now take three to six months. Backlogs that ran for months are cleared in weeks. 

On June 24, the regulator moved to put a frame around AI in lending. 

What the RBI proposed 

The RBI released a draft Guidance on Regulatory Principles for Model Risk Management, open for public comment until July 24. It covers nearly everyone it regulates — commercial banks, small finance banks, payments banks, co-operative banks, NBFCs, asset reconstruction companies, and credit information companies. It's a draft, not a final circular, and it builds on the RBI's August 2024 principles on model risk in credit and the August 2025 FREE-AI committee report. 

Most outlets led with the "AI kill switch" — the requirement that lenders be able to override, suspend, or deactivate any AI model the moment it produces harmful or erratic outputs. I believe two other provisions do more to change how lending actually works. 

The parts that change how lending works 

The first is human oversight on decisions that affect customers. Where an AI model influences an outcome like a loan approval or rejection, the borrower must be told AI was involved and be able to reach a human. Lenders also have to guard against what the RBI calls automation bias — staff approving whatever the model says without applying their own judgment. 

The second suggestion should make every fintech-NBFC partnership in the country re-read its contracts. The RBI states that the guidance applies to all models a regulated entity uses, including third-party models and those employing AI or ML — and that the entity is accountable for the outcomes either way. India's digital-lending stack runs on outsourced models: fintech scorecards, bureau-built scores, vendor underwriting engines. The working assumption was that the NBFC distributed the credit but could point at the model provider when a cohort went bad. That door is now shut. The lender owns the outcome. 

Why a good model goes bad 

The draft flags model drift as a core risk — the decay of a model's accuracy as real-world conditions move away from the data it was trained on. The RBI's concern, in plain terms: a sudden macro shift or an unexpected credit cycle can push an automated scoring system into misjudging default probability. 

A few weeks ago, in Edition #207 I made this exact argument about FOIR — that a model set at origination can't see a borrower whose real repayment capacity has shifted under an oil shock. The number on file looks fine while the ground underneath it moves. The RBI has now named that failure mode and is asking every lender to monitor it, validate against it, and be able to halt a model that's drifting. 

Strip away the kill-switch drama and that's what this guidance is: the regulator codifying what good underwriting always demanded. Explainability thresholds for every model. Independent validation. A documented inventory of what's running. Board-level sign-off on the high-risk ones. The RBI wants lenders to treat the model as something that decays when the economy moves, not a fixed truth set once and trusted forever. 

Who carries the weight 

This lands hardest on the lenders running the leanest, fastest, most outsourced AI stacks — the fintechs and NBFCs serving thin-file, small-ticket, new-to-credit borrowers. They built their speed on not having a human in every loop and on leaning heavily on third-party models. These are the lenders least set up to absorb the cost — and the ones whose borrowers can least afford to be cut off if they pull back. 

Explainability, human oversight, and board accountability are good governance and good for borrowers. They also add cost and friction to the instant, algorithmic, thin-file lending that widened credit access in the first place. If every material AI decision now needs a human who can explain it, the economics of a ₹12,000 loan get harder, and consumer-protective model governance could end up tightening access for the very borrowers it's meant to protect. That's not a reason to oppose the rules — it's what the RBI and lenders will have to manage as this moves from draft to circular. 

The takeaway 

The kill switch made the headlines, but accountability is what changes the business. The RBI is telling every lender the same thing: you can use a model to decide who borrows, but you have to be able to explain it, halt it, and own the outcome when it's wrong. So, if you can't explain why the model approved someone, you can't afford to approve them. 

Comments are open until July 24. Anyone lending on a model — which, now, is everyone — has reason to read the draft before it hardens into a rule. 

Reading list 

• RBI’s digital scam compensation pilot | Explained 

• BHIM UPI to offer bouquet of financial products 

• Govt tells public sector banks to expand deeper into rural India 

• Explained: Why Petrol Prices in India Don't Follow Crude's Trajectory 
  
  

Thank you for reading. If you liked this edition, forward it to your friends, peers, and colleagues. You can also connect with me on X here and follow FinBox on LinkedIn to get the latest updates.   
  
Cheers,   
Mayank