AI Compliance for RBI, SEBI & IRDAI 2026

BFSI is where Indian AI compliance gets specific. A bank's credit-underwriting copilot, a broker's research assistant, an AMC's portfolio-commentary generator, and an insurer's claims-summary model all carry live regulator exposure in 2026. Unlike earlier technology waves where regulators eventually caught up, RBI, SEBI, and IRDAI have all issued frameworks that explicitly contemplate AI within 24 months of LLMs reaching the Indian enterprise.

This guide is the regulator-by-regulator playbook. It assumes you already have a baseline governance program in place — for that layer, see the AI governance framework guide. Here we go deep on what each sectoral regulator expects and how to operationalise prompt logging, audit trails, and Explainable AI for each.

What You'll Learn

• How RBI's IT Governance Master Direction applies to AI systems in banks and NBFCs
• SEBI CSCRF mapping to AI components for brokers, AMCs, and depositories
• IRDAI Explainable AI and audit-trail requirements for insurers
• Prompt logging schema for BFSI
• Vendor-AI due diligence checklist aligned to sectoral rules
• Inspection-readiness artefacts

The Three-Regulator Reality

Indian BFSI firms rarely fall under just one regulator. A universal bank also runs an AMC subsidiary and a life insurance JV. That is three sets of IT/AI rules running in parallel. The table below shows the anchor instruments.

| Regulator | Instrument | Issued | Effective | Scope | |---|---|---|---|---| | RBI | Master Direction — Information Technology Governance, Risk, Controls and Assurance Practices | 7 Nov 2023 | 1 April 2024 | SCBs (excl. RRBs), SFBs, Payments Banks, NBFC Top/Upper/Middle Layers, CICs, large AIFIs | | SEBI | CSCRF circular SEBI/HO/ITD-1/ITD_CSC_EXT/P/CIR/2024/113 | 20 Aug 2024 | 1 Jan 2025 (existing REs) | Stock brokers, depositories, AMCs, MIIs, portfolio managers, AIFs and others | | IRDAI | Regulatory Sandbox Regulations | 3 Jan 2025 | 2025 cohort onwards | Life, general, and health insurers; reinsurers; insurance intermediaries |

All three sit above the horizontal DPDP Act 2023 floor.

RBI — IT Governance Master Direction Applied to AI

What the Master Direction Covers

The RBI MD consolidates earlier instructions on IT governance, risk management, business continuity, and information-system audit into one document for regulated entities. It does not name "AI" in every clause, but the definition of IT systems and the vendor-risk clauses are broad enough that AI/ML falls squarely inside.

Key focus areas the Board must demonstrably own:

• Strategic IT alignment — AI investments must trace back to the IT strategy the Board approved.
• IT risk management — AI model risk, data risk, and concentration risk (single LLM vendor) are IT risks.
• Resource management — AI talent, compute, and data are in-scope resources.
• Performance management — model performance metrics, SLA tracking with AI vendors.
• Business continuity and disaster recovery — fallback plans when an LLM API fails.

Governance Structure RBI Expects

• Board IT Strategy Committee with defined charter.
• IT Steering Committee (executive) reporting to the Board committee.
• Segregation of duties between development, testing, and production.
• Independent audit function with access to AI systems, logs, and vendor contracts.

Vendor & Third-Party AI Risk

The MD devotes significant attention to vendor risk. For AI, that translates to:

• Vendor due-diligence on LLM providers (financial, technical, security, regulatory posture).
• Concentration-risk mitigation — do not have 90% of AI workflows depending on a single model vendor.
• Exit plan — documented process to switch providers within a bounded time window.
• Sub-processor visibility — cloud regions, model hosts, and data storage locations must be listed.

For a hands-on comparison of LLM-provider enterprise platforms used by Indian banks, see VertexAI vs Bedrock vs Azure.

AI Use Cases Under Extra Scrutiny

• Credit underwriting and re-underwriting models.
• Fraud detection and anti-money-laundering flagging.
• Customer-facing chatbots where advice could be construed as financial advice.
• Early-warning-signal models feeding into NPA classification.
• Collections prioritisation models.

Each of these must be Tier 1 in your governance classification with DPIA, independent validation, and quarterly review.

SEBI — Cybersecurity and Cyber Resilience Framework (CSCRF)

What the CSCRF Requires

SEBI's CSCRF is built on five goals: Anticipate, Withstand, Contain, Recover, Evolve. For AI systems, each goal maps to a concrete control set.

| Goal | AI control | |---|---| | Anticipate | Threat modelling for prompt injection, model poisoning, data exfiltration via LLM | | Withstand | Rate limits, content safety filters, guardrail API, sandboxed tool-use | | Contain | Kill switch per model, network segmentation, PII redaction layer | | Recover | Incident runbooks, model rollback procedure, client communication templates | | Evolve | Quarterly red team, incident post-mortems feeding back into intake |

Compliance was expected by 1 January 2025 for existing SEBI REs and 1 April 2025 for newly regulated entities. SEBI issued technical clarifications (circular of August 2025) addressing AIF, InsurTech-adjacent, and smaller RE implementation questions.

What Brokers, AMCs, and Depositories Must Document

• Asset inventory that includes AI/ML systems as distinct assets with owner, data classification, and criticality.
• Role-based access controls on AI configuration and model deployment.
• Logging and monitoring with SOC integration; AI events must flow to the same SIEM as other IT events.
• Incident reporting to SEBI within the timeline specified in the CSCRF for material incidents.
• Annual CERT-In empanelled auditor review covering AI systems.

Research Desk and Algo-Trading AI

For brokers operating research desks, AI-generated research must be reviewed by a registered research analyst before publication. Algorithmic and AI-assisted trading systems are subject to SEBI's separate algorithmic trading framework and must be approved by stock exchanges; layering LLM prompts on top does not bypass approval. See also our AI for chartered accountants and finance teams for finance-adjacent AI workflows.

IRDAI — Explainable AI and Audit Trails for Insurance

The 2025 Regulatory Sandbox Signal

IRDAI notified the Regulatory Sandbox Regulations on 3 January 2025. The scope widened materially — sandbox applications are now permitted across the insurance value chain except for prudential and financial-stability matters. The InsurTech Working Group flagged four technology clusters reshaping insurance: AI/ML in underwriting, wearables in life and health, telematics/IoT in motor, and chatbot/voicebot-enabled claims handling.

What IRDAI Expects From AI Deployments

• Explainable AI (XAI). Recommendation logic must be expressible in human-understandable terms. Black-box models without an XAI wrapper fail IRDAI scrutiny.
• Audit trails. The complete decision-making process must be captured — input features, model version, intermediate scores, final recommendation, and human-override if any.
• Hallucination controls. For GenAI used in policy wordings, claim explanations, and support chatbots, insurers must implement review workflows to prevent plausible-but-incorrect outputs reaching policyholders.
• Fraud monitoring technology. IRDAI's 2025 Insurance Fraud Monitoring Framework expects real-time oversight, auditability, and traceability — squarely aligned with AI-driven fraud detection provided XAI is present.

Data Privacy Overlay

The DPDP Act 2023 sits on top. Insurers process sensitive personal data (health, financial) so SDF designation is likely for mid- and large-sized carriers, triggering mandatory DPIAs for each new AI model, an India-based DPO, and an independent data auditor.

Prompt Logging Schema — BFSI Reference

Implement this as a structured log table keyed on request_id. Every LLM call writes one row.

| Field | Example | Retention driver | |---|---|---| | request_id | UUID v4 | — | | timestamp_utc | 2026-04-19T07:32:15Z | Sectoral reg + DPDP | | employee_id or service_account_id | emp_41827 | SEBI CSCRF access control | | customer_reference | pseudonymised token | DPDP Act | | tenant_id | org_12 | Multi-tenant audit | | use_case_id | underwriting_copilot_v3 | RBI MD model inventory | | model_provider | anthropic or google or azure | Vendor risk | | model_id | claude-sonnet-4.6 | Drift analysis | | system_prompt_version | sp-7c2a | Change control | | input_hash_sha256 | hex | Privacy — never raw PII | | output_hash_sha256 | hex | Privacy — never raw PII | | input_tokens | 1,284 | Cost & sizing | | output_tokens | 612 | Cost & sizing | | latency_ms | 840 | SLA | | guardrail_verdict | PASS or BLOCKED or REDACTED | Safety | | human_reviewed | true or false | XAI / IRDAI | | override_reason_code | 14 — pricing exception | XAI / IRDAI |

Store raw inputs and outputs, if retained at all, in a separate encrypted store with tighter access than the log table. For the hardening stack around this, see enterprise AI security guardrails and secure AI prompting in regulated industries.

Vendor AI Due-Diligence Checklist

Run this for every AI vendor before procurement, and re-run annually.

• [ ] Incorporation, financial health, concentration of revenue on a single foundation-model provider.
• [ ] DPDP-compliant data processing agreement with India-specific clauses.
• [ ] Sub-processor list including foundation-model provider and cloud region.
• [ ] Training data exclusion commitment for customer prompts and outputs.
• [ ] Data residency attestations (Indian region availability for your use case).
• [ ] SOC 2 Type II, ISO 27001, and ISO/IEC 42001 posture.
• [ ] Security incident notification clock.
• [ ] Model versioning and deprecation policy.
• [ ] Red-team and safety testing public documentation.
• [ ] Exit plan — data return, contractual notice period, export of fine-tuned weights if applicable.

Inspection-Readiness Artefacts

When RBI, SEBI, or IRDAI knocks, produce within 48 hours:

1. AI system inventory — one row per model with owner, risk tier, regulator exposure.
2. Board-approved AI policy and Acceptable Use Policy.
3. DPIA file for every Tier 1 model with personal data.
4. Vendor contracts with AI-specific addenda.
5. Prompt logs sampled across a 12-month window.
6. Incident register with root-cause and closure.
7. Quarterly AI Governance Committee minutes.
8. Independent validation reports for Tier 1 models.

Key Takeaways

• RBI, SEBI, and IRDAI each now have anchor instruments that explicitly or implicitly cover AI — none of them permit "AI is a grey area" as an inspection answer.
• Prompt logging, audit trails, and Explainable AI are not advanced practices; they are baseline compliance for Indian BFSI.
• Model inventory and named accountable owners are the first ask in every regulator inspection.
• Vendor concentration risk on a single LLM provider is a material finding — plan multi-provider architecture early.
• Sectoral rules stack on top of DPDP; do not treat DPDP compliance as sufficient for BFSI.
• Annual CERT-In empanelled audit plus ISO/IEC 42001 readiness assessment together form a credible assurance narrative.

Official Resources

• RBI Master Directions page (IT Governance, Risk, Controls & Assurance)
• SEBI CSCRF circular (Aug 2024)
• SEBI Technical Clarifications to CSCRF (Aug 2025)
• IRDAI Guidelines page
• DPDP Act 2023 (MeitY PDF)
• NIST AI RMF 1.0 (PDF)

Next Steps

• Stand up the governance layer with the AI governance framework guide
• Extend the control stack with enterprise AI security guardrails
• Harden inputs with secure AI prompting for regulated industries
• Pick a platform using the AI vendor selection playbook for CIOs
• Review DPDP data residency and security practice

Last updated: April 19, 2026