India's AI governance discourse in banking focuses on MRMF and FREE-AI. But the third layer — RBI's IT Governance Master Direction (November 7, 2023, effective April 1, 2024) — preceded both and contains AI-specific requirements that most compliance programs have not mapped. When the three frameworks are read together, the compliance obligations are more extensive than any single framework suggests.
IT Direction November 2023
- IT Steering Committee AI oversight
- Vendor AI disclosure requirements
- AI-specific IT risk assessments
- Access control for autonomous AI
- IS audit scope includes AI systems
- BCP/DR for AI infrastructure
Responsible: CIO / IT Governance function
RBI FREE-AI August 2025
- AI risk appetite framework
- Fairness Sutra (caste/religion/language)
- 26 recommendations across six pillars
- AI ethics governance structure
- Customer-facing AI obligations
- Cross-cutting risk monitoring
Responsible: CRO / AI Risk team
RBI MRMF June 2026
- Model inventory and risk tiering
- Independent model validation
- Continuous performance monitoring
- Explainability and human oversight
- Kill switch (tested capability)
- Third-party model governance
Responsible: MRM function
Four AI-Specific Provisions in the IT Governance Master Direction
IT Steering Committee AI Oversight Mandate
RBI-regulated entities must maintain IT Steering Committees (ITSCs) that include AI/ML systems within their governance scope. ITSCs must: review and approve AI strategy (what AI will be deployed, for what purposes, at what risk appetite); approve significant new AI deployments before they go to production; receive periodic AI risk reporting (model performance, incidents, governance gaps); and provide board-level reporting on AI governance status. This is the IT governance committee mandate for AI — separate from, but coordinated with, the model risk committee mandate under MRMF and the AI risk oversight mandate under FREE-AI. Banks must design committee structures that prevent the same AI decisions being reviewed three times by three committees with no shared visibility, and prevent governance decisions falling in the gaps between committees.
Mandatory Vendor AI Disclosure
IT outsourcing contracts with vendors providing AI-enabled services must include specific AI disclosure requirements: what AI/ML is used in service delivery, which AI models are in use, how AI model failures affect service reliability, and how AI model changes will be communicated. This is the IT direction foundation for the more detailed vendor AI governance requirements in IT Outsourcing Directions 2023 (audit rights, RBI access) and MRMF (third-party model documentation, validation). Banks must retrofit vendor AI disclosure requirements into existing vendor contracts — many were signed before this direction was issued and do not include AI disclosure clauses.
Access Control for Autonomous AI Systems
The direction includes access control provisions that apply specifically to autonomous AI systems — systems that can execute actions without human authorization at each step. Privileged access management for AI systems: AI systems with database access, API access, or the ability to initiate transactions must have their access governed under privileged access management frameworks, with audit trails for AI-initiated actions. This provision provides the IT governance foundation for MRMF's autonomous agent authorization boundary requirement and human oversight mandate. Banks building agentic AI applications must implement this IT direction requirement alongside MRMF agentic AI governance.
AI-Specific IT Risk Assessment
IT risk assessments — required under the direction for all significant IT systems — must specifically include AI-related risks: model accuracy failure risk (AI model producing incorrect outputs at scale), AI system outage risk (AI serving infrastructure failure), adversarial attack risk (prompt injection, adversarial inputs, model poisoning), and AI-driven third-party concentration risk (multiple REs depending on the same AI vendor). This extends standard IT risk frameworks (COBIT, ISO 27001, NIST CSF) with AI-specific risk categories. AI risks assessed under the IT direction must be reconciled with AI model risks assessed under MRMF and AI risks assessed under FREE-AI — three separate risk assessment processes with overlapping scope require integration.
Overlap and Integration Points Across the Three Frameworks
| Governance Activity | IT Direction (Layer 1) | FREE-AI (Layer 2) | MRMF (Layer 3) | Integration Approach |
|---|---|---|---|---|
| AI risk assessment | IT-level AI risk: outage, adversarial, concentration | Enterprise AI risk appetite: aggregate risk across AI portfolio | Model-level risk tiering: Critical/High/Medium/Low per model | One integrated risk taxonomy with three scopes. Model-level risks roll up to enterprise risk appetite. IT-level risks feed model risk assessment. |
| Committee governance | IT Steering Committee reviews AI strategy and significant deployments | AI Ethics committee or CRO function reviews fairness and risk appetite | Model risk committee reviews model validation and remediation | Design a unified AI governance committee structure: ITSC covers IT governance; MRM committee covers model risk; shared reporting to board. Avoid three parallel committees with no shared data. |
| Vendor AI management | Vendor AI disclosure requirements; BCP/DR for vendor AI | Third-party AI fairness accountability (FREE-AI Pillar 5) | Third-party model documentation, validation, IT Outsourcing Direction requirements | One vendor AI management program satisfying all three layers. Contract template includes IT disclosure (Layer 1), fairness accountability (Layer 2), and model documentation/audit rights (Layer 3). |
| IS audit vs model validation | IS audit includes AI system controls: access control, change management, data governance | FREE-AI recommends periodic AI ethics audit | Independent model validation: model accuracy, performance, fairness, explainability | Define the boundary: IS audit assesses IT controls around AI (system, access, change); model validation assesses the model itself (accuracy, fairness, documentation). Coordinate to share evidence and avoid duplication. |
| Kill switch / BCP for AI | BCP/DR includes AI system recovery: RTOs, RPOs for AI infrastructure, fallback procedures | FREE-AI human-in-the-loop for customer-facing AI decisions | Kill switch: tested capability to suspend or deactivate AI models | Build one capability: AI system suspension (kill switch per MRMF) + manual fallback (BCP per IT Direction) + human oversight trigger (FREE-AI/MRMF human oversight). One tested mechanism, three regulatory requirements satisfied. |
How to Structure Compliance Across IT Direction, FREE-AI, and MRMF Without Tripling the Work
The three-layer compliance matrix is more complex than any single framework but does not require three parallel compliance programs. A well-designed program maps each governance activity to its responsible function, builds shared evidence infrastructure, and coordinates committee structures.
- Build one AI model inventory that serves all three layers. MRMF requires a model inventory for validation. FREE-AI requires an AI risk catalog for enterprise risk appetite. IT Direction requires vendor AI disclosure management. One model inventory — with fields for model risk tier (MRMF), enterprise risk appetite impact (FREE-AI), vendor AI disclosure status (IT Direction), and IS audit coverage status (IT Direction) — serves all three frameworks from one source of truth. This avoids three separate model lists maintained by three different functions with inevitable inconsistencies.
- Design committee structure before deploying AI governance tools. The three-layer compliance matrix creates the risk of three committees with overlapping but not identical AI governance remits — ITSC (IT Direction), AI Ethics/Risk Committee (FREE-AI), Model Risk Committee (MRMF). Each committee will want its own reports, its own approval workflows, and its own governance tools. Rationalize first: define what each committee decides (not all three should approve every AI deployment) and what shared information each receives. A governance tool that serves one committee creates shadow compliance for the others.
- Retrofit vendor AI disclosure into existing vendor contracts before the next vendor review cycle. IT Direction's vendor AI disclosure requirement applies to existing vendor contracts, not just new ones. Most banks have IT outsourcing contracts predating November 2023 that do not include AI disclosure clauses. Identify the next contract review or renewal date for each vendor using AI-enabled services. Use that renewal as the trigger to insert AI disclosure clauses (what AI is used, model change notification, AI failure reporting) — plus MRMF audit rights and RBI access clauses. A single contract template amendment covering all three framework requirements avoids renegotiating the same contracts three times.
- Integrate IS audit scope with MRM validation planning. IS audit (IT Direction) and independent model validation (MRMF) both cover AI systems — from different angles. Coordinate the annual IS audit plan with the MRM validation calendar: where IS audit plans to review an AI system, share the IS audit findings with the MRM team to avoid duplicate fieldwork. Where MRM plans to validate a model, share documentation with the IS auditor to avoid parallel documentation requests. Annual IS audit is typically more IT infrastructure-focused; model validation is more statistical and model-performance-focused. The division of labor is clear — the coordination eliminates duplicate effort.
- Build BCP/DR and MRMF kill switch as one tested capability. IT Direction requires BCP/DR for AI infrastructure (system recovery). MRMF requires a tested kill switch (model suspension). FREE-AI requires human-in-the-loop fallback for customer-facing AI decisions. These three requirements point to one capability: a tested procedure to suspend an AI model, route its workload to a fallback process (manual or rule-based), and recover the AI model after the incident. Design and test this once, document it as satisfying all three framework requirements, and run annual tests that produce evidence for all three audit/examination trails simultaneously.