AI governance for banks and NBFCs in India: FREE-AI and the RBI model risk draft

Where the RBI stands on AI

The Reserve Bank of India has approached AI in two steps. First came the FREE-AI committee report, published on 13 August 2025. It set out seven sutras and 26 recommendations, many of them addressed to regulated entities. They include a board-approved AI policy, an AI inventory updated at least half-yearly, an AI audit framework, consumer notice whenever AI is in use, and AI disclosures in annual reports. The report is a set of recommendations, not a direction.

The second step is the draft Guidance on Regulatory Principles for Model Risk Management, released on 24 June 2026. The RBI says it builds on its August 2024 draft on credit risk models and on the FREE-AI report. Once final, it would supersede the credit risk models chapter of the 2002 Guidance Note on Credit Risk Management. It covers every model a regulated entity uses, whether built in-house, bought or both, and it treats AI and machine learning models as models with extra controls rather than as a separate regime.

Read together, the two documents describe what supervisors will look for. FREE-AI says what good AI governance in finance contains. The model risk draft says how it should be run.

What the model risk draft asks for

The draft is detailed. These are the parts that shape an AI governance programme.

• A board-approved framework. Paragraphs 9 to 11 require a Model Risk Management Framework approved by the board, covering all models including AI and machine learning. The board approves risk appetite and policy. The Risk Management Committee of the Board oversees implementation, and model owners, validators and internal audit form three lines of defence.
• Risk tiering. Every model in the inventory is classified into a risk tier and the tier is reviewed at least annually. The tier sets how hard the model is validated, who approves it and how closely it is monitored. High-risk models need approval from the Risk Management Committee. Low complexity cannot offset high materiality.
• An inventory with no exceptions. Paragraph 21 says no model may be used unless it is in the inventory. The inventory covers active, inactive and decommissioned models, and decommissioned models are kept for at least ten years.
• Independent validation. Models are validated before deployment, after changes and periodically. Validation reports reach the Risk Management Committee, or a delegated authority, within three months.
• Change control. Model changes need a documented impact assessment and a version log, with thresholds that trigger re-validation.

For AI and machine learning, paragraphs 49 to 57 add explainability thresholds that are stricter for material decisions, with compensating controls where full explainability is not possible. They also cover bias and discriminatory output, drift, output variability and overfitting, and they require red teaming or an equivalent structured challenge for customer-facing and generative models. Automatic model updates need a defined scope and tighter monitoring.

Paragraphs 60 to 63 require human oversight, either in the loop or on the loop, with override, suspension and kill-switch mechanisms. Periodic human review of outputs is required, along with attention to automation bias and staff who know enough to challenge a model. The guide to human oversight of AI covers how to choose between those arrangements.

Third-party and generative AI

Most AI in Indian banks and NBFCs arrives through vendors, from credit scoring engines to chatbot platforms built on foundation models. The draft does not let that shift accountability. The regulated entity stays accountable for outcomes, validates the vendor’s model itself whatever certifications the vendor holds, and gives third-party models enhanced oversight from the Risk Management Committee whatever their tier. For material third-party AI, it must also weigh supply chain risk, concentration on a few providers and the effect of changes the provider makes. Contracts need documentation, audit rights for the entity and the RBI, and exit arrangements. The vendor AI risk guide sets out the due diligence and contract terms in more detail.

Customer-facing generative AI gets its own controls in paragraph 59. Those are protection against prompt injection and adversarial input, limits on session persistence, anomaly detection, a clear notice that the customer is talking to an AI system, and a way to switch to a person. Paragraph 54 asks for control boundaries that limit the risk of hallucination where outputs affect customers or decisions.

Data protection and incident reporting

Banking AI almost always processes personal data, so the DPDP Act and Rules apply alongside RBI expectations. Most of the Rules’ obligations, including notice, security safeguards, breach intimation and data principal rights, take effect on 13 May 2027. A large bank or NBFC may also be notified as a Significant Data Fiduciary, which brings an annual Data Protection Impact Assessment, an independent audit and the Rule 13(3) duty to check that algorithmic software does not put people’s rights at risk.

Cyber incidents involving AI systems also fall under the CERT-In directions of 28 April 2022, which require listed incidents to be reported to CERT-In within six hours of being noticed. An AI incident process should route to the RBI, CERT-In and, once the Rules apply, the Data Protection Board as each one requires. The guide on how to respond to an AI incident sets out the order of work.

Where to start

The inventory comes first. Every other requirement in the draft, from tiering and validation to oversight and board reporting, runs off a complete list of models with an owner, a purpose, a source and a status. Most institutions find more AI than they expected once they look at vendor products and staff use, so discovery belongs in this step. The guide to building an AI system register covers the fields to record.

Then tier each model, record the oversight arrangement and validation schedule, and bring vendor models into the same view. Board reporting follows from that record rather than being assembled by hand each quarter.

Aicura supports this work. Its AI register, discovery, risk and impact assessments, policies, oversight records, vendor due diligence and production monitoring give a regulated entity one place to run the programme the RBI describes, with ISO/IEC 42001 and the NIST AI RMF available as assessment frameworks from the Pro plan. The decisions, and the judgement on how the RBI’s final guidance applies, stay with you.

This page is Aicura’s reading of RBI publications, not legal or regulatory advice. Read the primary documents on the RBI website.