India AI & sector rulesUpdated 22 November 20269 min read

Board AI Policy Template for Indian Banks & NBFCs

A copy-ready board AI policy skeleton for Indian banks and NBFCs, mapped clause by clause to RBI FREE-AI recommendations and to DPDP Rules obligations.

By Suhas Bhairav · Reviewed against the DPDP Rules, 2025 (G.S.R. 846(E), 13 Nov 2025)

Recommended by
RBI FREE-AI committee report, August 2025
Approval level
Board of directors (or board committee)
Binding status
Becomes mandatory via RBI directions or circulars
DPDP core obligations
From 13 May 2027
DPDP log retention
At least one year (Rule 6(1)(e), Rule 8(3))
Breach report to DPB
Detailed report within 72 hours (Rule 7(2)(b))

Quick answer

A board AI policy for an Indian bank or NBFC should cover scope, risk appetite, an AI inventory, risk-based classification, approval gates, customer disclosure, human escalation, AI incident reporting, validation and audit, vendor controls and DPDP compliance. RBI's August 2025 FREE-AI report recommends a board-approved AI policy; it becomes mandatory only once RBI issues directions, so track circulars.

  • RBI's FREE-AI report recommends that every regulated entity adopt a board-approved AI policy; binding force depends on RBI issuing directions.
  • A workable policy has about a dozen clauses: scope, principles, roles, inventory, risk tiers, approvals, disclosure, escalation, incidents, validation, vendors and review.
  • Each clause should name an owner and an evidence artefact so the board can see the policy operating, not just approved.
  • DPDP obligations from 13 May 2027 — security, breach, retention and notice — belong inside the AI policy, not in a separate silo.
  • Review the policy at least annually and whenever RBI issues an AI-related direction.

Who this is for: Boards and risk committees of banks, NBFC and HFC leadership, Chief Risk Officers and Chief Compliance Officers, Heads of data science and model risk, Company secretaries drafting board papers

Why banks and NBFCs need a board AI policy

RBI's FREE-AI committee report (August 2025) recommends that regulated entities put in place a board-approved AI policy. It sits alongside other asks: an AI inventory, risk-based classification, customer disclosure, human escalation, AI incident reporting, audits and cybersecurity for AI.

Two reasons to draft the policy now rather than after a circular:

  1. Supervisory evidence. The report proposes a tolerant approach to first-time AI errors where safeguards existed. A board-approved policy, with minutes showing it was discussed, is the first safeguard a supervisor will look for.
  2. Fixed DPDP dates. From 13 May 2027, DPDP Rules 3, 6, 7, 8 and 14 apply to every AI system that touches personal data. A single policy can govern both.

The policy below is a skeleton. It names clauses, owners and evidence. Your legal and compliance teams should adapt wording to your entity type, size and existing frameworks (outsourcing, IT, cyber, model risk). Where RBI later issues directions with specific requirements, those directions override anything here. Check rbi.org.in before final approval.

For background on the framework itself, read the RBI FREE-AI explainer.

Copy-ready board AI policy skeleton

[ENTITY NAME] — POLICY ON THE USE OF ARTIFICIAL INTELLIGENCE
Version: 1.0   Approved by: Board / [Risk Committee]   Date: [ ]
Next review: [12 months from approval or on new RBI direction]

1. PURPOSE AND SCOPE
   1.1 Applies to all AI and ML systems built, bought or used by the
       entity, its employees and its outsourced service providers.
   1.2 Includes generative AI, foundation-model APIs, AI agents,
       vendor products with embedded AI, and employee use of public tools.

2. PRINCIPLES
   2.1 Adopt the FREE-AI sutras as guiding principles.
   2.2 Customer interest, fairness and accountability override speed.

3. GOVERNANCE AND ROLES
   3.1 Board: approves policy and risk appetite; receives reports.
   3.2 AI Oversight Committee: [CRO chair], CTO, CISO, CCO, DPO/privacy
       lead, business heads. Meets [quarterly].
   3.3 Use-case owner (business) and model owner (technical) per system.

4. AI INVENTORY
   4.1 Every AI use case registered before production use.
   4.2 Fields: owner, purpose, model/vendor, data categories, DPDP
       ground, risk tier, customer-facing (Y/N), validation dates.

5. RISK CLASSIFICATION
   5.1 Tiers: High / Medium / Low by customer impact, autonomy,
       data sensitivity and scale.
   5.2 Credit, pricing, collections and KYC decisions default to High.

6. APPROVAL GATES
   6.1 High: Oversight Committee approval + independent validation.
   6.2 Medium: model risk sign-off.  6.3 Low: owner attestation.

7. CUSTOMER DISCLOSURE AND HUMAN ESCALATION
   7.1 Customers told when they interact with AI.
   7.2 A human escalation route on every customer-facing AI channel.

8. AI INCIDENT MANAGEMENT
   8.1 Definition of AI incident; reporting to Oversight Committee.
   8.2 Linkage to DPDP breach workflow and CERT-In reporting.

9. VALIDATION, MONITORING AND AUDIT
   9.1 Pre-deployment testing: performance, fairness, robustness.
   9.2 Ongoing drift monitoring; periodic internal audit.

10. CYBERSECURITY FOR AI
11. THIRD-PARTY AND VENDOR AI
12. DATA PROTECTION (DPDP ACT 2023 AND RULES 2025)
13. TRAINING AND AI LITERACY
14. REPORTING TO THE BOARD
15. EXCEPTIONS, BREACHES OF POLICY AND REVIEW

How each clause maps to FREE-AI and DPDP

Use this table in the board paper so directors can see why each clause exists.

Policy clauseFREE-AI recommendation areaDPDP linkEvidence artefact
1 ScopePolicy pillars.3 applicabilityScope statement
3 GovernanceGovernance oversights.8(1) accountability, incl. processorsCommittee charter, minutes
4 InventoryAI inventory/repositoryData map for Rule 3 notices and Rule 14 rightsInventory export
5 Risk tiersRisk-based classificationRule 13(3) algorithmic due diligence (SDFs)Tiering rationale per row
6 ApprovalsGovernance oversight—Approval records
7 DisclosureCustomer disclosureRule 3 noticeScreenshots, scripts
7 EscalationGrievance and human escalations.8(10), Rule 14(3) ≤ 90 daysTicket logs
8 IncidentsAI incident reportingRule 7 breach intimationIncident register
9 ValidationAudits and model validationRule 13(1) DPIA and audit (SDFs)Validation reports
10 CyberCybersecurity for AIRule 6 safeguardsTest results
11 VendorsGovernance, assuranceRule 6(1)(f) contract clausesContracts, assessments
13 TrainingCapacity pillar—Attendance records

The SDF-specific rows apply only if the entity is notified as a Significant Data Fiduciary. No SDF has been notified at the time of writing; treat Rule 13 as a good-practice reference until then.

Drafting the cyber, vendor and data protection clauses

Clauses 10–12 are where generic AI policies tend to be thin. Suggested content:

Clause 10 — Cybersecurity for AI

  • Threats in scope: prompt injection, data poisoning, model extraction, insecure tool calls by agents, leakage through outputs.
  • Controls: input/output filtering, least-privilege credentials for agent tools, rate limits, red-team testing before launch for High-tier systems.
  • Logs from LLM gateways and model APIs feed the security monitoring stack. See CERT-In rules for AI systems for the 6-hour reporting and 180-day log retention directions.

Clause 11 — Third-party and vendor AI

  • No vendor AI in production without inventory registration and tiering.
  • Contracts must cover: data use limits (no training on entity data without approval), security safeguards, incident notification, audit and validation access, data location, sub-processor disclosure, and erasure on exit.
  • Payment data must stay within RBI's payment data localisation rules regardless of where the model runs.

Clause 12 — Data protection

  • Each AI use case records its DPDP ground: consent under section 6 or a section 7 legitimate use.
  • Personal data in prompts, embeddings and logs is protected per Rule 6: encryption or masking, access control, logging, backups.
  • AI-related personal data breaches follow the Rule 7 workflow: intimation without delay, detailed Board report within 72 hours.
  • Logs and associated data kept at least one year (Rule 8(3)), then erased unless other law requires longer.

What the board should receive and how often

A policy only works if the board sees it operating. A quarterly AI report to the board or risk committee can be short:

  1. Inventory movement — new, retired and re-tiered use cases.
  2. High-tier systems — status, validation results and open findings.
  3. AI incidents — count, severity, root cause, whether any were also personal data breaches or reportable cyber incidents.
  4. Customer outcomes — escalations, grievances and resolution times against the published grievance period.
  5. Vendor exposure — AI vendors in use, assessment status, any contract gaps.
  6. Regulatory watch — new RBI directions on AI, IT Rules amendments, DPDP milestones.
  7. Exceptions — any approvals granted outside policy, with rationale.

Metrics worth tracking

  • percentage of AI use cases with a completed inventory row and risk tier;
  • percentage of High-tier systems validated in the last twelve months;
  • median time from customer AI escalation to human response;
  • number of AI vendors without signed data protection clauses.

Avoid vanity metrics such as the number of pilots launched. Directors need to know whether risk is under control, not how much experimentation is happening.

Adopting the policy: a 90-day sequence

  1. Weeks 1–2: Gap review. Compare existing IT, outsourcing, cyber and model-risk policies against the skeleton. Many clauses can cross-refer rather than duplicate.
  2. Weeks 2–6: Inventory sprint. Populate the AI inventory. Include vendor products and employee use of public generative AI tools.
  3. Weeks 4–8: Drafting. Adapt clauses, define risk-tier criteria and set approval thresholds. Involve the privacy lead so clause 12 matches your DPDP compliance roadmap.
  4. Weeks 8–10: Committee review. Present to the AI Oversight Committee and risk committee with the mapping table.
  5. Weeks 10–12: Board approval. Minute the discussion, not just the resolution.
  6. After approval: run training, start quarterly reporting and schedule the first annual review.

Common pitfalls

  • A policy that names no owners.
  • Excluding employee use of public chatbots.
  • Copying another entity's risk tiers without calibrating to your products.
  • Forgetting that DPDP applies to AI logs and embeddings, not only to databases.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Confirm current RBI position

    Check rbi.org.in for any direction implementing FREE-AI before finalising the policy.

    RBI FREE-AI report

  2. Populate the AI inventory

    Register every AI use case with owners, data categories, DPDP ground and risk tier.

    Policy clause 4

  3. Set risk-tier criteria

    Define High, Medium and Low by customer impact, autonomy, sensitivity and scale.

    Policy clause 5

  4. Constitute the AI Oversight Committee

    Approve a charter with membership, quorum, frequency and reporting line to the board.

    Policy clause 3

  5. Implement AI disclosure and escalation

    Ensure every customer-facing AI channel discloses AI use and offers a human route.

    Policy clause 7; Rule 14(3)

  6. Link AI incidents to breach workflow

    Map AI incident categories to DPDP Rule 7 and CERT-In reporting paths.

    DPDP Rule 7

  7. Update vendor contracts

    Add data use, security, incident, audit and erasure clauses for AI vendors.

    DPDP Rule 6(1)(f)

  8. Set log retention

    Retain AI access and processing logs for at least one year, then erase unless law requires longer.

    DPDP Rule 8(3)

  9. Schedule annual review

    Diarise policy review at twelve months and on any new RBI AI direction.

    Policy clause 15

Key dates

  1. DPDP Rules notified; Board rules in force

  2. DPDP core obligations apply to AI systems processing personal data

Answers

Frequently asked questions

Is a board-approved AI policy mandatory for Indian banks?

RBI's FREE-AI committee report recommends one, but the report is not itself binding. It becomes mandatory when RBI issues a direction or circular requiring it. Check rbi.org.in for the current position. Many banks are adopting a policy anyway because supervisors look for documented safeguards.

What should a bank AI policy contain?

At minimum: scope, principles, governance roles, an AI inventory, risk classification, approval gates, customer disclosure and human escalation, AI incident management, validation and audit, cybersecurity for AI, vendor controls, data protection under DPDP, training, board reporting and a review cycle. Each clause should name an owner and an evidence artefact.

Does the AI policy need to cover ChatGPT use by employees?

Yes. FREE-AI's inventory idea covers all AI use cases, and employee use of public generative AI tools can expose customer personal data. The policy should either prohibit, restrict or approve specific tools, and require that no customer personal data is entered into unapproved services.

How does the AI policy relate to the DPDP Act?

AI systems that process personal data must meet DPDP Rules from 13 May 2027: Rule 3 notices, Rule 6 security safeguards, Rule 7 breach intimation, Rule 8(3) one-year retention, and Rule 14 rights. Putting these inside the AI policy avoids two frameworks governing the same model differently.

Can an NBFC use the same template as a bank?

Yes, with calibration. Smaller NBFCs can merge the AI Oversight Committee into an existing risk committee and use simpler tiering. The clause structure stays the same. Any RBI direction specific to NBFCs or scale-based regulation layers would override the template.

How often should the board review the AI policy?

At least annually, and whenever RBI issues a new AI-related direction, the entity launches a High-tier AI use case, or a significant AI incident occurs. Quarterly AI reports to the board or risk committee keep directors informed between reviews.

Legal references and sources

  • RBI FREE-AI Committee Report — Board-approved AI policy recommendation (official text)
  • Digital Personal Data Protection Rules, 2025 — Rules 6, 7 and 8(3) (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 13 (if notified as SDF) (official text)

This guide explains the law for engineering and business teams. It is not legal advice; confirm how it applies to your organisation with qualified counsel, and check the Gazette for later notifications or amendments.

About the author

Suhas Bhairav

Suhas Bhairav

Systems architect and applied AI practitioner who designs production AI systems for Indian and global companies, with a focus on privacy-by-design under the DPDP Act and the EU AI Act.