Your model can score a borrower. Can it defend the decision?
Why explainability just became an origination problem, not a data science one
A customer applies for a loan. Your system declines it in eleven seconds. Six months later, that customer files a grievance, and your ombudsman asks a simple question: why was this application rejected?
For most Indian lenders, answering that question today takes days. It involves pulling records from a credit system, reconstructing which version of the policy was live that week, finding out which officer or which model made the call, and hoping the deviation note was written down. Sometimes the honest answer is that nobody knows.
That gap is no longer just operationally awkward. It is now a regulatory exposure.
What Changed
The RBI released its Framework for Responsible and Ethical Enablement of Artificial Intelligence, FREE-AI, on 13 August 2025. It sets out seven guiding principles and 26 recommendations across six pillars, and a draft Model Risk Management guidance followed in 2026.
Two of those seven principles land directly on the credit function. Understandable by Design requires that AI decisions be transparent and documented, so that customers and regulators can see how a credit or fraud decision was reached. Accountability makes clear that decision-making responsibility cannot be handed to a machine. Boards and senior management remain answerable for AI outputs.
The framework is specific about digital lending. AI-based credit assessments must be auditable rather than opaque, with access controls and tamper-evident logs applied to models and data. For high-stakes decisions such as credit, adverse onboarding outcomes and fraud flags, there must be a human override path and an explanation that can be shown both to the customer and to an auditor, with a record of when and why each override occurred.
Customers must also be told when they are dealing with an AI system and given a route to contest the outcome. Independent audits, bias testing and impact assessments are expected to become routine parts of compliance, and external certification of high-stakes models such as credit scoring is a realistic prospect.
The direction of travel was reinforced at the highest level. A review chaired by the Finance Minister with the RBI, MeitY and bank chiefs on 23 April 2026 examined systemic AI risk in the financial sector.
Credit growth must not come at the expense of underwriting standards. RBI Deputy Governor Shirish Chandra Murmu, 7th NBFC and HFC Summit, Mumbai, 3 September 2026.
Why This Is an Origination Problem
Most institutions have read FREE-AI as a model governance exercise and handed it to the data science team. That is the wrong owner.
Explainability is not produced at the model. It is produced at the point where the decision is assembled, which is your loan origination system. The model contributes a score. The decision is the score, plus the policy version it was matched against, plus the eligibility rules applied, plus the deviation raised, plus the officer who approved it, plus the timestamp on each of those events.
If those elements live in separate systems, or if some of them live in an email thread, no amount of model interpretability will let you reconstruct the decision. SHAP values explain a score. They do not tell a regulator who waived the FOIR breach on 14 March.
There is a second reason this sits with origination. Under the Digital Lending Directions, accountability cannot be outsourced. When a lending service provider or a fintech partner contributes to underwriting, the regulated entity still owns the decision and must be able to explain it. If the evidence trail sits in the partner's stack, you own a liability you cannot document.
What You Should Be Able to Produce on Demand
Four things, for any loan file, without a manual reconstruction exercise.
The policy version. Which product policy, which eligibility criteria, which thresholds were live when this application was assessed. Policies that live in circulated PDFs cannot answer this.
The decision path. Every check the file passed or failed, in sequence. Bureau, KYC, FOIR, exposure, collateral. Not a final verdict. The path to it.
The human layer. Which deviations were raised, who approved or rejected each one, on what basis, and when. This is precisely what FREE-AI means by recording overrides.
The rejection reason, in language a customer can be shown. Not an internal code. A statement that survives a grievance process.
Governance and Speed Are Not Opposites
The instinctive reading is that all of this slows lending down. In practice the reverse holds.
When product policy sits inside the origination system rather than in a document, eligibility is confirmed and deviations flagged automatically, so files reach an underwriter already sorted. When the rule engine, decision engine and workflow engine run on one platform, straight through processing becomes possible precisely because the rules are explicit enough to be trusted without a human read. Files that clear bureau, KYC and asset norms can move directly to sanction, while genuinely marginal cases get the human attention they deserve.
The audit trail is a by-product of that structure, not an additional burden layered on top of it.
That is the reframe worth taking to your board. Institutions still treating explainability as a compliance cost will end up building documentation around a process they cannot see into. Those that rebuild the origination layer get the governance and the throughput from the same investment.
The regulator has asked a reasonable question. Most lenders simply are not yet built to answer it.
See how a structured origination platform produces this by design: www.celusion.com/origin
______________________________________________________________________________________________________________
Sources: RBI FREE-AI Framework (13 August 2025); RBI draft Model Risk Management guidance (2026); RBI Digital Lending Directions, 2025; Deputy Governor's address, 7th NBFC and HFC Summit, Mumbai (3 September 2026); Finance Ministry review on systemic AI risk (23 April 2026).
