Indian lenders are deploying AI faster than they are governing it. The Reserve Bank's FREE-AI framework — Fairness, Reliability, Explainability, Ethics and Accountability in AI — is the clearest signal yet that credit models will be supervised like any other material risk. The question for most institutions is no longer whether to govern AI, but how to make governance operational without grinding lending to a halt.
This piece lays out a working blueprint: the artifacts, controls and evidence that turn FREE-AI's five principles into something a model risk team can run — and a board can sign off on.
Start with a single source of truth
Every governed estate begins with an inventory. If you cannot list every model touching a credit decision — including the vendor scorecards and the spreadsheet heuristics nobody calls "AI" — you cannot govern them. A usable AI register captures ownership, purpose, training data lineage, and the decision it influences.
- Owner & approver — a named accountable executive, not a team.
- Decision context — where in the lifecycle the model acts: origination, limit-setting, collections.
- Risk tier — driven by materiality and customer impact, refreshed quarterly.
- Evidence links — validation reports, bias tests and monitoring dashboards.
Make explainability contestable, not decorative
Explainability fails the moment a customer disputes a decline and the institution cannot articulate why. FREE-AI raises the bar from technical interpretability to contestability: a borrower must be able to challenge an automated decision and receive a reasoned, human-reviewable response.
If you can't explain a decline to the customer in plain language, you don't have an explainable model — you have a defensible-looking black box.
The practical test is simple. Take ten recent adverse decisions and ask a first-line reviewer to reconstruct the reasoning from the model's outputs alone. If they can't, your reason codes are theatre.
Wire monitoring to the risk appetite
Governance that lives in an annual validation cycle is already stale. Production models drift; portfolios shift; a segment that was fair at launch can skew within a quarter. Continuous monitoring — for performance, stability and fairness — is what converts a policy document into a control.
The three signals that matter
Most drift dashboards drown teams in metrics. In credit, three signals carry the freight: population stability across key segments, approval-rate divergence between protected groups, and override frequency by underwriters — the human tell that a model is losing trust.
Key takeaways
- A live AI inventory is the precondition for every other control.
- Explainability only counts if a customer can contest the decision.
- Continuous monitoring tied to risk appetite beats annual validation.
- The board needs one dashboard, not twelve model reports.
Give the board one picture
Directors don't need twelve model validation reports; they need a single operating picture — a governance score, an exception log, and a clear view of which decisions the institution is now delegating to machines. That is precisely the layer FREE-AI expects boards to own, and the one most institutions are missing today.
Get these four moves right — inventory, contestable explainability, appetite-linked monitoring, and board-level visibility — and FREE-AI stops being a compliance burden. It becomes the operating system for lending you can actually defend.
