AI Governance Framework for Indian Enterprises 2026

AI governance in India has moved from slideware to scheduled board agenda in under two years. The trigger is the Digital Personal Data Protection Act 2023, but the pressure compounds from every side — RBI's Master Direction on IT Governance, SEBI's Cybersecurity and Cyber Resilience Framework, IRDAI's evolving InsurTech rules, and enterprise clients who now ask for an ISO/IEC 42001 attestation before signing. A CIO who shipped five AI pilots in 2025 now owns fifty in production and cannot point to a single authoritative inventory.

This guide lays out a working AI governance framework for Indian enterprises in 2026. It is framework-agnostic at the top — anchored on NIST AI RMF 1.0 because it is free, mature, and board-readable — and maps down to India-specific obligations and a concrete org chart you can copy.

What You'll Learn

• The four NIST AI RMF functions mapped to Indian regulatory context
• A 6-phase model lifecycle from intake to retirement
• Governance org chart template with RACI
• DPDP Act alignment: what Significant Data Fiduciaries must actually do
• A 90-day roll-out plan with gates and artefacts
• Key Takeaways for board-level sponsorship

Why AI Governance Became a 2026 Board Topic

Three forces converged. First, the DPDP Act's Significant Data Fiduciary (SDF) designation brings mandatory DPIAs for new AI/ML models and an independent data auditor. Second, sectoral regulators — RBI's IT Governance Master Direction (effective 1 April 2024), SEBI's CSCRF (circular SEBI/HO/ITD-1/ITD_CSC_EXT/P/CIR/2024/113, 20 August 2024), and IRDAI's 2025 Regulatory Sandbox Regulations requiring Explainable AI and audit trails in underwriting — have made AI risk a specific inspection item. Third, enterprise buyers (especially US and EU clients of Indian IT services) now ask for ISO/IEC 42001:2023 alignment or a SOC 2 add-on covering AI controls.

A governance program is the load-bearing artefact that answers all three without triplicating effort.

Foundation: NIST AI RMF 1.0 Mapped to India

NIST AI RMF 1.0 organises AI risk work into four functions: Govern, Map, Measure, Manage. Use them as the outer loop for every AI initiative.

Govern — Culture, Policy, Accountability

Govern sits above the other three. It answers: who is accountable, which policies apply, and how do we know the program is working.

| Govern sub-function | India-specific artefact | |---|---| | AI policies and procedures | Written AI Acceptable Use Policy, Model Development Standard, Third-party AI Policy | | Roles and accountability | AI Governance Committee charter, RACI per model, DPO sign-off for personal-data models | | Workforce competence | Mandatory "AI & DPDP" training, annual re-certification for prompt engineers and MLEs | | External input and oversight | Independent data auditor (for SDFs), board risk sub-committee quarterly report |

Map — Context and Risk Framing

Map establishes what the AI does, who it affects, and what can go wrong before anything is built.

Every Indian enterprise should run Map as an intake gate. Outputs: a one-page Model Card, a filled Risk Classification Matrix, and a go/no-go decision with a recorded rationale.

| Risk tier | Example use cases | Required artefacts | |---|---|---| | Tier 1 — High | Credit underwriting, fraud flag, clinical support, HR shortlist | DPIA, bias test, red-team report, Govt./Regulator notification where applicable | | Tier 2 — Medium | Customer-service chatbot, personalised recommendations, document summarisation | DPIA (if personal data), eval report, human-in-the-loop plan | | Tier 3 — Low | Internal productivity (email drafting, code suggestions on non-prod) | Acceptable Use attestation, audit log |

Measure — Quantitative and Qualitative Evaluation

Measure is where most Indian enterprises under-invest. Build a shared eval harness so every model — internal or vendor — reports the same metrics before production.

• Performance — task-specific accuracy against a locked eval set; must include Indic-language samples where the model faces Indian users.
• Robustness — adversarial prompts, out-of-distribution inputs, prompt-injection suite.
• Bias & fairness — disparity across gender, religion, caste proxies (state, mother tongue), age; log the test design not just the numbers.
• Privacy leakage — PII regurgitation, membership inference where training data was tuned.
• Explainability — required by IRDAI's 2025 Regulatory Sandbox for insurance AI; Explainable AI (XAI) outputs must be capturable for audit.

Manage — Prioritise, Respond, Communicate

Manage closes the loop: allocate treatment resources, respond to incidents, communicate with affected principals and regulators.

• Incident categories: data leak via LLM, hallucinated advice, unauthorised training, regulator notice, copyright complaint.
• India breach-notification clock under DPDP Rules 2025: without delay to the Data Protection Board; internal clock target of 24 hours.
• Post-incident: add the failure mode to your Map intake questions so the next model inherits the lesson.

The 6-Phase AI Model Lifecycle

Wrap NIST's functions around a concrete lifecycle your engineering teams can actually run.

1. Intake — business sponsor files a one-page problem statement; AI Governance Committee triage within 10 working days; risk tier assigned.
2. Design — Model Card v0, data source inventory, DPIA (if personal data), vendor due diligence (if third-party model); go/no-go to build.
3. Build & Train — in controlled environment, training data provenance recorded, bias and safety tests during development not after.
4. Validate — independent validation by a team other than builders; eval report signed by Risk; DPO attestation for personal-data models; pen-test for customer-facing systems.
5. Deploy — staged rollout (canary to 5%, then 25%, then full) with kill switch; monitoring dashboards live before 100% cutover; human-in-the-loop where risk tier requires.
6. Monitor & Retire — drift detection, quarterly review for Tier 1, annual re-validation, and a dated retirement plan — models do not live forever.

For enterprises building a dedicated capability, the V.A.U.L.T. framework is the transformation overlay; the lifecycle above is the operating model.

Governance Org Chart Template

Copy this structure and adjust titles to your hierarchy.

Board Risk Sub-Committee
        │
        ▼
AI Governance Committee (meets monthly)
├── Chair: CIO, CRO, or Chief AI Officer
├── Members: CISO, DPO, Head of Legal, Head of HR, 2 Business CXOs
└── Standing invitees: Head of Data, Head of MLOps, Lead Model Validator

        │
        ▼
AI Center of Excellence (operational)
├── Head of AI CoE
├── Model Validators (independent from builders)
├── AI Platform / MLOps Engineers
├── Responsible AI Lead (bias, fairness, explainability)
└── AI Incident Response rotation

        │
        ▼
Business Units (build AI use cases)
└── Accountable Product Owner per model — single name, not a team

For step-by-step build-out of the operational layer, see the AI Center of Excellence guide.

RACI at a Glance

| Activity | Product Owner | AI CoE | DPO | CISO | AI Gov Committee | |---|:---:|:---:|:---:|:---:|:---:| | Use-case intake & risk tiering | R | A | C | C | I | | DPIA | R | C | A | I | I | | Model validation | C | A | I | I | R | | Production go-live | C | C | C | C | A | | Incident response | R | A | C | R | I | | Annual review | R | C | C | C | A |

R = Responsible, A = Accountable, C = Consulted, I = Informed.

DPDP Act Alignment for AI Systems

The DPDP Act 2023 (No. 22 of 2023) is the statutory floor. Governance reads the Act through three questions:

1. Lawful basis & consent. Section 6 requires free, specific, informed, unconditional, unambiguous consent with a clear affirmative action. Generic ToS text fails. For AI systems, consent notices must state that AI will process the data and why.
2. Significant Data Fiduciary duties. If the Government designates you an SDF (based on data volume, sensitivity, risk, and use of emerging tech — which explicitly includes AI), Section 10 adds obligations: appoint an India-based DPO, an independent data auditor, and conduct DPIAs for new AI/ML models.
3. Rights of data principals. Correction and erasure (Sections 11–12) create practical pressure to prefer RAG architectures over fine-tuning so deletion is feasible.

For the detailed compliance matrix covering HIPAA, PCI-DSS, and SOC2 alongside DPDP, see the enterprise AI compliance guide.

ISO/IEC 42001 Layer for Certifiable Maturity

ISO/IEC 42001:2023, published December 2023, is the first certifiable AI management system standard. Treat it as the PDCA overlay on your NIST-based program. The clauses that matter most for Indian enterprises:

• Clause 4 — organisational context and stakeholders (maps to Govern).
• Clause 6 — AI objectives, risk assessment, and system impact assessment (maps to Map).
• Clause 8 — operational controls including data, third-party AI, and impact monitoring (maps to Measure + Manage).
• Annex A — 38 controls covering data quality, human oversight, AI system lifecycle, and provider obligations; use as your compliance checklist.

Certification via an accredited body (e.g., BSI, TÜV SÜD, SGS) is a 9–12 month effort once a program exists; start the readiness assessment at the 6-month mark of your governance program.

Prompt Logging, Audit Trail, and Red-Teaming

Governance without evidence is a slide deck. Three non-negotiable telemetry streams:

• Prompt log — every LLM call captured with a request hash, model ID, user ID, tenant ID, timestamp, and content-safety verdict. Encrypt at rest; retain per your data-retention policy (typically 18–36 months).
• Audit trail — every configuration change (system prompts, model swaps, guardrail thresholds) versioned with approver name and ticket link.
• Red team — a standing internal or contracted red team runs prompt-injection, jailbreak, PII-leak, and bias probes quarterly for Tier 1 models. Report to the AI Governance Committee.

For the hardening playbook, see the enterprise AI security guardrails guide.

90-Day Roll-out Plan

| Day | Milestone | Owner | |---|---|---| | 0–15 | Charter AI Governance Committee; publish Acceptable Use Policy | CIO / CRO | | 15–30 | Complete AI system inventory (shadow AI amnesty); risk-tier each | AI CoE | | 30–45 | Stand up prompt logging and audit trail in platform | AI CoE + CISO | | 45–60 | Run DPIAs for all Tier 1 and Tier 2 models with personal data | DPO | | 60–75 | First quarterly board risk report with AI section | AI Gov Committee | | 75–90 | ISO/IEC 42001 gap assessment scheduled with external party | Head of Compliance |

Key Takeaways

• AI governance in India is not one law but the intersection of DPDP Act, sectoral regulators, and voluntary standards — you must synthesise, not wait.
• Anchor on NIST AI RMF 1.0 because it is free and fast; layer ISO/IEC 42001 when clients demand attestation.
• Risk-tier every model at intake; high-risk models need DPIA, independent validation, and quarterly review.
• A named accountable product owner per model is the single most impactful governance control — more than any document.
• Prompt logs and versioned system-prompt history are the evidence base; without them, compliance is unverifiable.
• ISO/IEC 42001 certification is a 9–12 month journey that should start with a gap assessment, not a gap.

Official Resources

• NIST AI 100-1: AI Risk Management Framework 1.0 (PDF) — full text of the RMF
• NIST AI RMF Playbook — actionable guidance for each function
• ISO/IEC 42001:2023 — AI management systems — the certifiable standard
• DPDP Act 2023 (MeitY PDF) — official gazette text
• RBI Master Directions page — IT Governance, Risk, Controls & Assurance Practices
• SEBI CSCRF circular (Aug 2024) — for AMCs, stock brokers, and depositories

Next Steps

• Benchmark your regulator-specific controls for RBI, SEBI, and IRDAI
• Build the operational layer using the AI Center of Excellence playbook
• Extend the compliance matrix with the HIPAA, PCI-DSS, and SOC2 guide
• Implement the telemetry and hardening in the enterprise AI security guardrails guide
• Apply the transformation overlay with the V.A.U.L.T. framework

Last updated: April 19, 2026