RBI Mandates AI Kill Switch and Human Oversight for Banks

India’s central bank has ordered every bank, NBFC and cooperative lender to install an AI kill switch that can instantly shut down any AI model in their systems. The Reserve Bank of India released its draft “Guidance on Regulatory Principles for Model Risk Management, 2026” on June 24, 2026, and is taking public comments until July 24, 2026. The framework applies to eleven categories of regulated entities, from commercial banks to credit information companies.

The proposal also pushes model risk governance onto the board of every regulated entity for the first time and warns banks against automating decisions they cannot shut down.

What the RBI Just Put on the Table

The draft guidance, issued under press release number 2026-2027/528 and signed by Brij Raj, Chief General Manager of the RBI’s Department of Regulation, covers every regulated lender, NBFC, payments bank, cooperative and asset reconstruction company in India. At its core is a requirement that no AI model operates without a way to be shut down if it produces harmful or erroneous outputs. Banks and other regulated entities must establish “override, suspension and deactivation mechanisms, including kill-switch arrangements, to ensure that no AI model can operate without the ability to be immediately shut down if it produces harmful or erroneous outputs,” the RBI said in the draft.

The framework also makes human oversight of every AI-driven decision non-negotiable, with the central bank calling out automation bias, the tendency for employees to over-rely on AI outputs without applying their own judgement. For customer-facing AI, banks must tell customers they are talking to a machine and offer a human switch on request. The draft further introduces a risk-based tiering structure, requires board-level approval for high-risk models, and explicitly flags supply chain risk from concentration in a small set of global AI providers. Public feedback runs through the RBI’s Connect 2 Regulate portal as part of the public consultation process on the draft, open until July 24, 2026.

The framework’s headline numbers, at a glance.

• Document: Draft “Guidance on Regulatory Principles for Model Risk Management, 2026” (PR 2026-2027/528)
• Issued by: RBI Department of Regulation, signed by Brij Raj, Chief General Manager
• Release date: June 24, 2026
• Comment deadline: July 24, 2026
• Scope: 11 categories of regulated entities, from commercial banks to credit information companies

Who the Rules Cover, Every Model, Every Lender

The applicability list in the draft is unusually wide, with eleven categories of regulated entities falling under the new framework. The draft defines a “model” just as broadly: any system, internally developed, third-party sourced or a mix, that takes inputs, applies statistical or AI-driven processing logic, and produces outputs used in business or decision-making. The framework applies irrespective of whether the regulated entity itself recognises a particular tool as a “model.” Even cooperative banks and credit information companies will sit alongside commercial lenders in the new scope.

Even spreadsheet-based calculators qualify if they materially shape decisions like lending rates. The RBI’s own illustration calls out “a spreadsheet-based loan pricing calculator that takes inputs (borrower type, tenor, credit score, collateral value), applies processing logic (interest rate grids, risk-weighted spreads, margin formulas), and produces a final lending rate affecting business decisions” as something that should be treated as a model. Many NBFCs and cooperatives will discover they are running more models than they had labelled, the consultancy Vinod Kothari noted in its reading of the draft Guidance on Model Risk Management.

From Spreadsheet Calculators to Frontier AI

The draft treats AI and machine-learning systems as a special class of model with stricter requirements. For AI and ML, regulated entities must assess risks arising from hallucinations, bias and discriminatory outcomes, as well as data drift and adversarial attacks. They must also test such models under stressed scenarios before deployment. The framework requires banks to define explainability thresholds, the ability to explain in understandable terms why a model produced a particular output, for every AI system.

For generative AI that touches customers, the rules go further. Banks and NBFCs must guard against hallucinations with system-level controls, build in protections against prompt injection and adversarial inputs, and run structured red-teaming, or challenge testing, on such models, according to the consultancy Vinod Kothari’s breakdown of the draft.

The framework’s model inventory rules leave little room for shadow AI. Every regulated entity must maintain a comprehensive inventory of active, inactive and decommissioned models, and no model can be used unless it is on that list. The inventory must capture model owners, developers, validators, approvers, risk tiers and findings from monitoring exercises. The board-approved MRMF itself must cover the full model lifecycle, from selection and development through validation, approval, deployment, monitoring, change management, business continuity and decommissioning. Risk tier for each model must be reviewed at least once a year, or earlier if specific triggers are met.

The board-approved Model Risk Management Framework must cover the full model lifecycle:

1. Model selection and development
2. Validation
3. Approval structure (including exceptions and risk mitigants)
4. Deployment and monitoring
5. Change management
6. Business continuity management
7. Decommissioning

For customer-facing AI systems, the rules explicitly mandate a disclosure that customers are interacting with AI, a statement of the system’s limitations, and an offer to switch to a human at any point. Generative AI in particular must be tested against prompt injection, adversarial inputs and anomalous usage, with structured red-teaming required for such models, according to the consultancy Vinod Kothari’s reading of the draft. Decommissioned models must be retained for ten years or more, Vinod Kothari noted.

The Kill Switch and Human Oversight

The headline provision is the kill switch itself. Regulated entities must have the ability to instantly override, suspend or deactivate any AI model deployed in their operations, “including a kill switch arrangement,” as part of a sweeping draft framework on Model Risk Management, the RBI said. If the switch is flipped, the model has to stop producing outputs that affect customers, pricing, credit or operations. RBI’s kill switch mandate for banks is the framework’s most visible provision.

Beyond the switch, every AI-driven decision must remain subject to human oversight. The draft frames human-in-the-loop and human-on-the-loop arrangements, periodic human review and explicit override mechanisms as the standard, and warns against automation bias and decision fatigue among staff reviewing AI outputs.

If not effectively governed and managed, model risk can lead to inaccurate outcomes, flawed decisions, financial losses, operational disruptions, compliance failures, and other serious adverse consequences, for the RE itself, its consumers, and the broader financial system.

The Reserve Bank of India laid out the warning in the introduction to the draft Guidance on Regulatory Principles for Model Risk Management, 2026. The document was issued under press release 2026-2027/528, signed by Brij Raj, Chief General Manager of the RBI’s Department of Regulation. It is open for public comments through July 24, 2026.

The Board Is Now on the Hook

For the first time, the RBI is placing AI and model governance squarely at the board level. Every regulated entity must put in place a board-approved Model Risk Management Framework, or MRMF, applicable to all models including AI and ML systems, regardless of origin. The board’s own duties include approving the entity’s risk appetite for model risk and setting policies for model risk tiering. Those policies must be forward-looking and informed by stress testing and scenario analysis, the draft specifies.

The draft leans on a classic three-lines-of-defence structure. Model owners form the first line, an independent model risk management and validation function forms the second, and internal audit forms the third. The Risk Management Committee of the Board, or RMCB, sits over all three lines. It reviews validation reports for high-risk models before they are deployed, reviews model risk tiering at least annually, and oversees monitoring of models approved with exceptions and any model involving AI. Breaches and material concerns are reported up to the RMCB as well.

High-risk models are reserved for the RMCB itself. They cannot be cleared by the technology or risk team alone.

When the assessed risk of a model exceeds the entity’s risk appetite, the framework calls for timely corrective action: enhanced controls, restrictions on use, remediation, or decommissioning. A report on each such case must be placed before the RMCB. Risk tiering itself must reflect materiality, complexity and other supervisory considerations, with one explicit guardrail: a low complexity score cannot drag down the overall risk tier of a highly material model. Tiering must be reviewed at least annually.

For AI and ML specifically, the framework adds explainability, fairness and stress-testing requirements on top of the standard model validation regime. Banks must define explainability thresholds for every AI model and proactively test for discriminatory outcomes. They must recalibrate or redesign where those outcomes show up. The draft’s enhanced requirements for AI systems also cover risks arising from hallucinations, data drift and adversarial attacks. AI models deployed in customer-facing settings must additionally be tested for prompt injection, adversarial inputs and anomalous usage, with structured red-teaming required for generative systems.

Vendors Don’t Get a Pass

The draft takes a hard line on third-party AI. “An RE is accountable for the outcomes of all models used by it, irrespective of whether those models are developed internally, sourced from third parties, or a combination thereof. Delegation to a vendor does not transfer or dilute the RE’s regulatory accountability,” the guidance states. Regulated entities must independently validate third-party models regardless of any certification the vendor provides. They must undertake due diligence before acquisition and deployment.

The RBI has also flagged supply chain risk, the risk arising from over-dependence on a limited number of AI model providers, as a concern banks must actively manage. The Economic Times called this a pointed reference to the growing concentration of AI capabilities in a handful of global technology companies.

Contractual levers come with the rulebook. Banks need contractual rights to technical documentation and audit access over any AI vendor they use. Decommissioned third-party models must be retained for at least ten years. Validation reports for high-risk models must reach the RMCB before deployment, even when the model is sourced from outside. Where risk appetite is breached, the regulated entity itself remains on the hook, regardless of where the model sits on its vendor list.

Why Now, and What Comes Next

The June 24 draft is the third leg of a regulatory arc that began with the RBI’s August 5, 2024 draft on credit model risks and continued with the August 13, 2025 report of the Committee on Framework for Responsible and Ethical Enablement of Artificial Intelligence, known as FREE-AI. The 2026 guidance significantly expands scope from credit models to all models used across the business, including third-party and AI/ML models. Once finalised, it will supersede Chapter 3 of the RBI’s October 12, 2002 Guidance Note on Credit Risk Management.

The timing also tracks global pressure. The May 2026 warning on AI cyber risk in finance from the IMF cited Anthropic’s Claude Mythos Preview model, which the Fund said could find and exploit vulnerabilities in every major operating system and web browser, even when used by non-experts. The IMF’s analysis noted that AI models can dramatically reduce the time and cost needed to find and exploit vulnerabilities, raising the likelihood of simultaneous attacks across institutions. Indian banks, like peers abroad, are watching the same threat curve.

The RBI has signalled that this is not the end of the road. The draft notes that further requirements specifically applicable to AI models may be issued later, as indicated in paragraph I.10 of the RBI’s Utkarsh 2029 strategic document.

The 2026 framework is proportionality-based, meaning regulated entities are expected to apply the principles in step with the nature, scale and complexity of their operations. The consultation closes on July 24, 2026, with feedback routed through the RBI’s Connect 2 Regulate portal or directly to the Chief General Manager, Operational Risk Group, in Mumbai. Banks, NBFCs, cooperative lenders and AI vendors all have a stake in what the final shape looks like. Until then, the framework remains a draft, not a final circular. The RBI will issue the final guidance after considering comments received during the consultation window.

Frequently Asked Questions

What is the RBI’s AI kill switch rule?

The Reserve Bank of India has proposed that every regulated lender, NBFC and cooperative in India be able to instantly override, suspend or deactivate any AI model it runs. The draft Guidance on Regulatory Principles for Model Risk Management, 2026, published on June 24, 2026, calls the mechanism a “kill switch arrangement” and ties it to a wider framework covering board-level governance, model risk tiering and human oversight.

Who has to comply with the RBI’s draft AI framework?

The draft applies to eleven categories of RBI-regulated entities, spanning commercial and foreign banks, all four NBFC layers, payments banks, local area banks, regional rural banks, urban and rural cooperative banks, all-India financial institutions (EXIM Bank, NABARD, NaBFID, NHB, SIDBI), asset reconstruction companies and credit information companies. Foreign banks operating in India are covered as commercial banks under the Banking Regulation Act, 1949.

What is the deadline for feedback on the RBI’s draft AI rules?

The RBI is accepting public comments until July 24, 2026, through its Connect 2 Regulate portal on the RBI website, or by post and email to the Chief General Manager, Operational Risk Group, Department of Regulation, RBI, Mumbai.

What counts as a “model” under the RBI’s new framework?

The draft defines a model as any system, internally developed, third-party sourced or a mix, that takes inputs, applies statistical or AI-driven processing, and produces outputs used in business or decision-making. The RBI’s own illustration calls out a spreadsheet-based loan pricing calculator that drives lending rates, meaning regulated entities may be running more systems that qualify as models than they have formally labelled. The definition is intentionally broad: any tool that materially shapes lending rates, pricing or other business decisions counts, even if the entity does not internally call it a model.

Will the RBI rules affect third-party AI vendors?

Yes. Regulated entities remain fully accountable for the outcomes of any model they use, including third-party AI supplied by a vendor. The draft requires independent validation regardless of any certification the vendor provides, contractual rights to technical documentation and audit access, and active management of supply chain risk from concentration in a small number of AI providers. Vendors that lose the bank’s contract still cannot escape the bank’s obligation to maintain documentation and validation evidence for at least ten years.

Disclaimer: This article is for informational purposes only and does not constitute financial, legal or investment advice. The framework discussed is a draft under public consultation as of publication on June 25, 2026. Specific rules and figures may change before finalisation. Consult a qualified professional for decisions tied to regulatory compliance or AI adoption in financial services.