Short answer: FREE-AI recommends that every regulated entity, NBFCs included, have a board-approved AI policy, but it does not order one. Recommendation 14 of RBI's FREE-AI committee report says regulated entities (REs) "should establish a board-approved AI policy". The report is a committee report, not an RBI direction, so write "FREE-AI recommends", never "RBI mandates". The policy should cover governance, accountability, risk appetite, safeguards, auditability, consumer protection, AI disclosures, the model life cycle and liability, and it should sort every AI use into low, medium or high risk. Annexure V of the report gives an eight-section outline, and this post turns it into a template. As of 20 November 2026, RBI has not issued final model risk management guidance. The June 2026 draft is its latest AI-specific text.
This guide is for the people who will draft the policy: risk, technology, compliance and the company secretary. It is engineering and governance guidance, not legal advice. Have counsel review the final wording before it goes to the board.
Is a board-approved AI policy mandatory for NBFCs?
No, not on its own. RBI released the FREE-AI committee report on 13 August 2025 (Press Release 2025-2026/902). It sets out seven Sutras (principles), six pillars and 26 recommendations. Recommendation 14 sits under the Governance pillar and is tagged for "REs and Industry" on a medium-term horizon. It is a recommendation. The report is not a direction, circular or master direction.
In your policy and board papers, write "FREE-AI recommends". Do not write "RBI requires". It can still matter: RBI may turn parts of the report into rules, and banks, auditors and investors may ask whether you have followed it. Other RBI directions that already apply to you do not stop applying because a system uses AI. Which ones cover your entity is a question for counsel.
The report speaks to "regulated entities", and its surveys covered banks, NBFCs, asset reconstruction companies and all-India financial institutions. For a wider view, see our page on RBI FREE-AI implementation.
What does RBI's FREE-AI report say a board AI policy should cover?
Recommendation 14 reads: REs "should establish a board-approved AI policy which covers key areas such as governance structure, accountability, risk appetite, operational safeguards, auditability, consumer protection measures, AI disclosures, model life cycle framework, and liability framework. Industry bodies should support smaller entities with an indicative policy template."
| Area in Recommendation 14 | The question your policy must answer |
|---|---|
| Governance structure | The question your policy must answer: Which bodies and people decide on AI, and who can stop a system? |
| Accountability | The question your policy must answer: Who is named as the owner of each AI use, and who answers to the board? |
| Risk appetite | The question your policy must answer: What outcomes will the board accept, and what is off limits (for example, fully autonomous credit decisions)? |
| Operational safeguards | The question your policy must answer: What testing, monitoring, fallback and human review must a system have before and after launch? |
| Auditability | The question your policy must answer: What logs, documents and model records exist, and who can audit them? |
| Consumer protection measures | The question your policy must answer: How do customers learn they are dealing with AI, get an explanation, and reach a person? |
| AI disclosures | The question your policy must answer: What do you tell customers, supervisors and the public about your AI use? |
| Model life cycle framework | The question your policy must answer: How does a model move from idea to approval, live use, change and retirement? |
| Liability framework | The question your policy must answer: Who bears the loss when an AI system errs, and how does that run through vendor contracts? |
The last row needs care. Recommendation 8, which is addressed to regulators, says REs "must continue to remain liable for any loss suffered by customers". So the policy cannot hand liability to a vendor on paper. Contracts can allocate cost, but the customer relationship stays with you.
What is in the suggested policy outline (Annexure V)?

Annexure V of the report gives a suggested outline of a board policy on AI. It says the outline may be customised to the organisation's needs and complexity of use. It has eight sections.
| # | Section | What Annexure V asks the section to cover |
|---|---|---|
| 1 | Section: Purpose and Scope | What Annexure V asks the section to cover: The role of AI, the stakeholders (internal and vendors) and the outcomes you want |
| 2 | Section: Principles | What Annexure V asks the section to cover: Principles aligned with the FREE-AI Sutras, adjusted to your needs |
| 3 | Section: Governance Structure and Roles | What Annexure V asks the section to cover: Decision-making and accountability, oversight and escalation, and a grievance redressal mechanism |
| 4 | Section: AI Life cycle management | What Annexure V asks the section to cover: Testing and product approval before deployment, continuous red-teaming, and controls and responsibilities at each stage |
| 5 | Section: Data Governance and Management | What Annexure V asks the section to cover: Sourcing, cleaning, anonymisation, encryption, sharing and purging of data; bias detection; synthetic data validation |
| 6 | Section: Risk Management and Controls | What Annexure V asks the section to cover: Pre-deployment questions on fairness, accuracy, cyber risk, customer recourse, AI disclosure and data leakage |
| 7 | Section: Third-Party and Vendor Management | What Annexure V asks the section to cover: Scope, responsibility and liability of third-party vendors |
| 8 | Section: Policy Review | What Annexure V asks the section to cover: A review at least once a year to reflect new regulation and technology |
The seven Sutras are: Trust is the Foundation; People First; Innovation over Restraint; Fairness and Equity; Accountability; Understandable by Design; and Safety, Resilience, and Sustainability.
How should the policy classify AI risk?
Paragraph 4.4.44 says the policy should include "a clear risk classification framework" that sorts AI use cases into low, medium and high risk. The factors it names are impact on customers, criticality and potential for harm. It gives indicative examples, and says classifications should be reviewed and updated periodically.
| Tier | Examples in the report | What the tier should trigger in your policy (our suggestion) |
|---|---|---|
| Low | Examples in the report: Internal uses such as document summarisation and email classification, where outcomes have limited impact | What the tier should trigger in your policy (our suggestion): Entry in the inventory, a named owner, approved tools only, light review |
| Medium | Examples in the report: Customer-facing tools such as chatbots, and fraud detection used for preliminary assistance | What the tier should trigger in your policy (our suggestion): Pre-launch testing, AI disclosure to customers, a human route, red teaming at least every six months (para 4.4.60), logging |
| High | Examples in the report: Credit underwriting, and autonomous AI that handles customer interactions, makes financial decisions or moves customer funds | What the tier should trigger in your policy (our suggestion): Risk committee approval, human oversight, independent review, fallback tested in drills, board visibility, the most frequent red teaming |
The last column is our reading, not RBI text. It draws on Recommendation 16 (human oversight, "especially for medium and high-risk use cases"), Recommendation 20 (red-team frequency tied to risk level) and Recommendation 24 (an audit framework aligned with a board-approved risk categorisation, with third-party audits for high-risk or complex uses).
Write criteria, not just examples, so an engineer can tier a new feature quickly. Does it talk to customers? Does it influence a decision about money, credit or access? Does it act without a human step? Does it use personal data? A "yes" to the second or third question should push it up a tier.
Who should own AI risk: the board, the Risk Management Committee or a new AI committee?
Paragraph 4.4.45 says the Risk Management Committee or a similar body should identify, assess and mitigate AI-related risks and fold them into the institution's overall risk framework. It adds that the entity could set up an AI Adoption Committee, or use an existing technology body, to bring business, risk, compliance and technology teams together.
| Role | Suggested duty in the policy |
|---|---|
| Board | Suggested duty in the policy: Approves the policy and risk appetite. Receives a report on AI use, incidents and high-risk systems at a set interval |
| Risk Management Committee (or equivalent) | Suggested duty in the policy: Owns AI risk identification and mitigation. Approves high-risk uses. Reviews the inventory and incident log |
| AI adoption or technology committee (optional) | Suggested duty in the policy: Cross-functional approval of new uses, vendor choices and tier assignments. Reports to the risk committee |
| Named owner per AI system | Suggested duty in the policy: Answers for purpose, data, testing, monitoring and fixes. Appears in the inventory |
Whatever structure you pick, name who can switch each high-risk system off.
What does an NBFC board AI policy template look like?

The template follows Annexure V. Replace the square brackets with your own details. It is a skeleton for your risk and legal teams, not finished policy text.
- 1. Purpose and scope. "This policy sets how [Company] adopts, uses and oversees artificial intelligence. It covers every AI system used by [Company], whether built in-house or supplied by a vendor, including generative AI tools used by staff. It applies to [business lines, subsidiaries, vendors acting for us]."
- 2. Principles. "[Company] adopts these FREE-AI Sutras: [list]. In practice: [one plain sentence per principle]."
- 3. Governance and roles. "The Board approves this policy and our AI risk appetite. The [Risk Management Committee] owns AI risk. Each AI system has a named owner. [Role] may suspend any AI system immediately. Customers can raise AI-related grievances through [channel], and they are answered within [period]."
- 4. AI life cycle. "No AI system goes live until it is in the AI inventory, tiered, tested and approved through our product approval process. Medium and high-risk systems are red-teamed at least every [six] months and after major changes. Every system has a fallback and a defined retirement process."
- 5. Data governance. "AI systems use only data we are permitted to use for the stated purpose. Personal data in prompts, training sets, logs and outputs follows our DPDP retention and erasure schedule. We test for bias before launch and at set intervals."
- 6. Risk and controls. "Before launch, the owner answers these questions in writing: Are outputs fair, accurate and transparent? Does the system add cyber risk? Can customers get recourse? Do customers know they are dealing with AI? Can the system reveal confidential or personal information?"
- 7. Third parties and vendors. "Vendor AI is held to the same standard as our own. Contracts must cover [data use and retention, model change notice, audit rights, incident notice, exit]. We stay accountable to customers for vendor AI outcomes."
- 8. Review. "The Board reviews this policy at least once a year, and sooner when RBI issues final guidance, a major incident occurs or our AI use changes materially."
Legal review note: this skeleton is a starting point drawn from public RBI material. The wording of any clause that creates a duty, a liability or a customer promise should be settled by your counsel and approved by your board before use. Recommendation 14 also expects industry bodies to publish a template for smaller entities. If one appears, compare it with this skeleton. For the staff-facing rules that sit under this policy, see our AI acceptable use policy template.
What evidence should sit behind each policy clause?
Supervisors, auditors and bank clients will ask for records, not statements. Map every clause to something you can show.
| Policy clause | Evidence to keep | FREE-AI reference |
|---|---|---|
| AI inventory | Evidence to keep: A register of every model and use case: owner, purpose, tier, vendor, data, dependencies and grievances. Update at least half-yearly | FREE-AI reference: Recommendation 23; para 4.4.68 |
| Product approval | Evidence to keep: Approval records showing an AI-specific risk review for each AI-enabled product | FREE-AI reference: Recommendation 17 |
| Testing and monitoring | Evidence to keep: Validation results, drift and degradation monitoring, and change logs | FREE-AI reference: Recommendation 16 |
| Red teaming | Evidence to keep: Test plans, findings, fixes and dates. At least every six months for medium and high-risk uses | FREE-AI reference: Recommendation 20; para 4.4.60 |
| Fallback and continuity | Evidence to keep: A tested fallback for each AI system and BCP drill records that cover model degradation | FREE-AI reference: Recommendation 21 |
| Customer protection | Evidence to keep: Disclosure screens, human-handoff logs and complaint records | FREE-AI reference: Recommendation 18 |
| Audit | Evidence to keep: Internal audit reports, third-party audits for high-risk uses, framework review at least every two years | FREE-AI reference: Recommendation 24 |
| Disclosure | Evidence to keep: AI-related disclosures for the annual report and website | FREE-AI reference: Recommendation 25 |
| Incidents | Evidence to keep: An incident log and a form with date, detection, impact and fix. Annexure VI of the report has an indicative sample | FREE-AI reference: Annexure VI; paras 4.4.48 and 4.4.64 |
The tool that holds most of this together is the inventory. If you do not know which AI tools your teams use, start with a shadow AI audit.
How does the draft model risk guidance affect the policy?
RBI issued draft "Guidance on Regulatory Principles for Model Risk Management" on 24 June 2026 (Press Release 2026-2027/528). Comments closed on 24 July 2026. As of November 2026, RBI has not issued final guidance, so the draft is the latest text. It covers models that use AI and machine learning, including third-party models, and applies to banks, NBFCs and other regulated entities. Published summaries of the draft say it proposes, for example, model tiering, an inventory, independent validation of third-party models, human oversight and kill-switch arrangements. We write "the draft proposes", never "RBI requires". Our post on the draft and your GenAI chatbot goes through it.
Write your policy so that it can absorb the final text without a rewrite. Keep principles and roles in the board policy, and put validation methods and tier thresholds in a procedure that management can update. Define tiers once, and point to the inventory instead of copying it. Add a clause that final RBI guidance starts a policy review before the annual date.
How should the policy treat vendor and off-the-shelf AI tools?
Paragraph 4.4.46 says the use of third-party or off-the-shelf AI tools, such as generative AI apps for drafting documents, summarising reports or analysing data, should be governed by the organisation's policies. It also says internal AI policies must be compliant with the broader national AI governance and regulatory frameworks. Annexure V adds that the policy should define the scope, responsibility and liability of vendors.
The policy should require answers to three questions before a vendor tool is approved. Where does our data go, including region, retention, training use and sub-processors? What changes without telling us, such as model updates and new terms? What happens if it fails or leaves, meaning the fallback, the exit plan and getting your records back? For staff use of public chat tools, add a plain approved-tools rule with a route to request new ones. Our guide to ChatGPT and AI tools at work under DPDP covers the data side.
How does a board AI policy fit with DPDP?
Recommendation 15 of the report asks REs to set up data governance for AI covering collection, access, usage, retention and deletion, and says it should comply with applicable laws such as the DPDP Act. Prompts, retrieved documents, logs and training data often hold personal data, so your AI policy and your DPDP programme should share one data map and one retention schedule.
The core DPDP duties start on 13 May 2027. Our post on DPDP-compliant AI architecture shows where personal data hides in an AI feature, and our DPDP page for BFSI covers the sector view.
How should the board approve and review the policy?
- Inventory first. List every AI use and vendor. You cannot set an appetite for what you cannot see.
- Draft with the owners. Risk, technology, compliance, data protection and the business lines write the sections together.
- Test it on three real systems. Run a low, medium and high-risk use through the draft.
- Legal review. Counsel checks the wording, the liability clause and how the policy sits with your other RBI obligations.
- Risk committee, then board. The risk committee reviews and recommends. The board approves and the minutes record it.
- Report on a schedule. The board sees the inventory, high-risk systems, incidents and red-team results at a fixed interval, and reviews the policy at least once a year.
Mono does not certify anyone, and it is not a law firm. We help NBFCs and fintechs build the inventory, controls and evidence behind the policy. See our RBI FREE-AI implementation service.
Sources: Reserve Bank of India, Report of the Committee on Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI), released 13 August 2025 (Press Release 2025-2026/902), including paras 4.4.43 to 4.4.46, 4.4.48, 4.4.59 and 4.4.60, 4.4.64, 4.4.68, Recommendations 8, 14 to 18, 20, 21, 23 to 25, and Annexures V and VI. RBI, Press Release 2026-2027/528, draft Guidance on Regulatory Principles for Model Risk Management, 24 June 2026. Digital Personal Data Protection Rules, 2025. Last reviewed 20 November 2026.
FAQ
Does RBI require NBFCs to have a board-approved AI policy?
Not through the FREE-AI report on its own. Recommendation 14 says regulated entities should establish a board-approved AI policy, but the report is a committee report with recommendations, not an RBI direction. RBI could turn parts of it into rules later. Ask counsel which RBI directions already bind your entity.
What should a board-approved AI policy for an NBFC include?
Recommendation 14 lists governance structure, accountability, risk appetite, operational safeguards, auditability, consumer protection, AI disclosures, the model life cycle and liability. Annexure V of the report suggests eight sections: purpose and scope, principles, governance and roles, AI life cycle, data governance, risk and controls, third-party management, and an annual review.
How should an NBFC classify AI use cases by risk?
FREE-AI paragraph 4.4.44 asks for a low, medium and high framework based on customer impact, criticality and potential for harm. Its examples are document summarisation as low, chatbots and fraud detection as medium, and credit underwriting or autonomous AI that moves customer funds as high. Review the tiers periodically.
Who should own AI risk at an NBFC?
Paragraph 4.4.45 of the report says the Risk Management Committee or a similar body should identify, assess and mitigate AI risks. It also says the entity could set up an AI Adoption Committee, or use an existing technology body, to link business, risk, compliance and technology teams. Each AI system should also have a named owner.
How often should the board review its AI policy?
Annexure V of the FREE-AI report says the AI policy may be reviewed at least once annually to reflect new regulation and technology. Add triggers too, such as final RBI model risk guidance, a major AI incident or a material change in the AI systems you use.
Does the board AI policy need to cover third-party AI tools?
Yes. FREE-AI paragraph 4.4.46 says third-party and off-the-shelf AI tools, such as generative AI apps used for drafting or analysis, should be governed by the organisation's policies. Annexure V asks the policy to define the scope, responsibility and liability of vendors. Recommendation 8 says REs remain liable for customer losses.
