• Download the draft guidelines here. 

Banks and other regulated entities (REs) may soon have to ensure every artificial intelligence (AI) system they deploy can be overridden, suspended or deactivated through “kill-switch arrangements”, under the Reserve Bank of India’s (RBI) draft Guidance on Regulatory Principles for Model Risk Management, 2026. 

Released for public consultation on Wednesday, the draft proposes a broad framework governing how banks, NBFCs and other financial institutions develop, deploy and oversee models, including AI systems. Stakeholders can submit comments until July 24, 2026. The guidance would apply to commercial banks, small finance banks, payments banks, local area banks, co-operative banks, regional rural banks, NBFCs, all-India financial institutions, asset reconstruction companies and credit information companies. The guidelines also clarify that “further requirements, if any, applicable to AI models may be issued later”. 

Why is the RBI issuing these guidelines? The RBI said REs are increasingly using models to improve customer service, automate business processes, strengthen risk management and defend against cyber attacks. It attributed this to the growing scale and complexity of financial activities, digitalisation of financial services, advances in analytical and computational capabilities, AI and machine learning (ML), and greater reliance on third-party providers.

The draft defines a “model” broadly. It includes any internally developed or third-party system that uses data and statistical, mathematical, financial or AI/ML techniques to generate outputs used for business operations or decision-making. It also covers algorithms, analytics, applications, decision-based rules and other computational tools that materially affect business decisions, regardless of whether an RE formally classifies them as models.

The AI-specific provisions apply to AI and ML models, including foundational AI models and frontier AI models. The RBI said REs should define their scope and implement additional controls based on their potential impact on customers, business operations and financial outcomes.

Board oversight and governance: The RBI proposes that every RE establish a Board-approved Model Risk Management Framework (MRMF) covering governance, model tiering, inventory, documentation, validation, approvals, monitoring, change management, business continuity and decommissioning.

The Board would oversee the framework, while the Risk Management Committee of the Board would review high-risk models, monitor third-party and AI models, review model tiering at least annually and oversee breaches or other material concerns. The RBI also makes clear that REs remain accountable for the outcomes of every model they use, whether developed internally, sourced from third parties or a combination of both.

AI should only be used where risks can be managed: The RBI said REs should assess whether risks arising from AI models can be adequately identified, measured, monitored and managed before deploying them. AI models should only be used in business processes and use cases where those risks can be effectively managed.

The draft also requires REs to classify every model according to risk. For AI models, this assessment should additionally consider “the extent of reliance and the level of autonomy placed on the model outputs for decision-making.” Models with greater autonomy or greater reliance on their outputs could therefore attract higher risk classifications.

For material third-party AI models, datasets and dependencies, REs should also consider risks arising from dependence on a limited number of providers, supply chain risks, limitations in independently validating models and behavioural changes resulting from provider-driven updates.

AI must remain under human control: REs would have to establish robust human oversight arrangements for AI models, including automated decision-making systems. These arrangements should include:

• Human-in-command mechanisms, such as human-in-the-loop or human-on-the-loop oversight.
• Override, suspension or deactivation mechanisms, including kill-switch arrangements.
• Periodic human review of model outputs and AI-driven decisions to identify anomalies.

The oversight mechanism should also account for automation bias, over-reliance on AI outputs and decision fatigue. Additionally, personnel responsible for overseeing AI systems should possess sufficient expertise to “effectively challenge, override, or escalate issues/concerns in model outputs where required.” REs should periodically review human interventions, overrides, incidents and near misses to strengthen oversight arrangements.

How RBI wants AI models to behave: The draft requires REs to define explainability and transparency thresholds for every AI model. Models relied upon for material decision-making or having significant impact on customers or operations should meet higher explainability standards.

Where full explainability cannot be achieved, the RBI said REs should implement enhanced validation and testing mechanisms to verify and corroborate model outputs before use, more frequent validation and monitoring, usage restrictions and other compensating controls.

The draft also directs REs to establish “appropriate control boundaries” to mitigate hallucinations, particularly in generative AI models and use cases where AI outputs directly or indirectly influence customer interactions or decision-making.

Additionally, REs should identify risks of bias and discriminatory outputs, particularly where AI could unfairly treat certain customer groups. They should conduct fairness assessments and implement mitigants, including recalibration or redesign where necessary.

The RBI further said REs should ensure AI models:

• are not overfitted to training data and can generalise to real-world conditions;
• do not rely on spurious correlations or unintended relationships;
• do not produce excessive or unexplained variation under similar inputs, and where such variability, stochastic behaviour or model uncertainty exists, it should be managed through measures such as confidence scores and probability outputs; and
• continuously monitor and address data quality issues, non-representative or incomplete datasets, intellectual property risks, data drift and concept drift.

How RBI wants AI models tested:

• Independently validate every model, including third-party models, regardless of any validation, certification or assurance provided by the service provider. Validation should assess the model’s inputs, assumptions, conceptual soundness, design, performance and alignment with its intended use.
• Validate models throughout their lifecycle, including before deployment, after deployment, following modifications, on internal or external triggers and periodically under the MRMF.
• Test AI models under stressed conditions to ensure vulnerabilities do not emerge through edge cases, abnormal inputs, manipulations and adversarial conditions.
• Carry out structured challenge processes, including red-teaming or equivalent testing, particularly for models involving customer interaction or generative AI capabilities.
• Apply additional controls to models that receive dynamic or automatic updates, including defining what can be updated automatically, justifying automatic updates, conducting enhanced data quality checks, and monitoring such models more frequently.
• Maintain enhanced documentation for AI models to ensure traceability, reproducibility, and auditability, reflecting their complexity, self-adapting nature, and reliance on training data.

Stronger controls for third-party AI: Before acquiring or using a third-party model, REs should carry out due diligence covering the credibility of the service provider, the model’s methodology and limitations, and the suitability and quality of data used.

Where a third-party provider does not disclose sufficient information about an AI model, REs should identify risks arising from those constraints and implement mitigants, including limiting the model’s use.

The RBI also proposes that contracts with third-party providers should provide REs with access to sufficient technical documentation to understand, validate and audit models, while also covering audit rights, continuity arrangements and exit mechanisms.

Deployment controls: REs should ensure model outputs are replicated and stable in the production environment before deployment. AI models should also not introduce vulnerabilities into either the model itself or the RE’s production environment. REs should implement access controls, cybersecurity safeguards and controls covering APIs, external interfaces and third-party integrations.

For customer-facing AI systems, including generative AI models, REs should implement additional cybersecurity controls against prompt injection, adversarial inputs, persistent sessions and anomalous usage patterns.

They should also clearly disclose that users are interacting with an AI or ML-based system, explain its limitations and provide customers with the option to switch to human assistance whenever requested.

Also read:

• AI in Finance: RBI Committee Calls for Sector-Specific Models, AI Sandbox Etc.
• Lowdown: RBI’s new advisory pushes for tighter data use, AI safeguards, vendor checks
• Explained: How the RBI Draft Amendments Seek to Curb Dark Patterns in Banking