Short answer: as of 20 November 2026, RBI's model risk management guidance is still a draft, not final. RBI released the draft "Guidance on Regulatory Principles for Model Risk Management" on 24 June 2026 and took comments until 24 July 2026. We found no final version: RBI's draft notifications page still lists it as a draft, and its notifications page shows no final guidance. The draft would apply to banks, NBFCs and other RBI-regulated entities, and its definition of a "model" is wide enough to cover many customer-facing GenAI chatbots. If it is finalised as drafted, a regulated entity would need to tell users they are talking to AI, offer a human on request, defend the bot against prompt injection, red-team it, validate it independently and be able to switch it off. None of that binds anyone yet. But banks and NBFCs may start to reflect it in their vendor questionnaires before it is final.
This guide is for engineering and product leads at NBFCs, fintechs and SaaS companies that sell AI features to them. It explains what the draft says about GenAI, what you can build now, and what to keep out of your claims. It is not legal advice. Where a regulatory reading matters, take it to counsel.
Is RBI's model risk management guidance final or still a draft?
It is still a draft. As of November 2026, RBI has not issued final model risk management guidance. The June 2026 draft is the latest text. We checked RBI's notifications and draft-notifications pages on 20 November 2026. The draft is listed under drafts, and no final version appears among the notifications through 11 November 2026.
| Item | Status | Source |
|---|---|---|
| Document | Status: Draft "Guidance on Regulatory Principles for Model Risk Management" | Source: RBI Press Release 2026-2027/528 |
| Released | Status: 24 June 2026 | Source: RBI Press Release 2026-2027/528 |
| Comments closed | Status: 24 July 2026 | Source: RBI Press Release 2026-2027/528 |
| Final version | Status: None found as of 20 November 2026 | Source: RBI notifications and draft-notifications pages, checked 20 November 2026 |
| Circular number | Status: None. The copy of the draft we read carries placeholders ("DoR.ORG.REC.XXXX", "June xx, 2026"), so cite the press release | Source: Draft text |
| Effective date | Status: None stated in the draft | Source: Draft text |
| What it would replace | Status: Chapter 3 (Credit Risk Models) of the Guidance Note on Credit Risk Management, 12 October 2002 | Source: Draft, para 64 |
| Earlier steps | Status: Draft credit model risk principles, 5 August 2024; FREE-AI committee report, 13 August 2025 | Source: RBI Press Release 2026-2027/528 |
The final text may differ from the draft. The draft itself says (para 2) that further requirements for AI models may be issued later. Treat every provision below as proposed. We write "the draft proposes", never "RBI requires".
Who does the draft apply to, and does it reach a SaaS vendor?
The draft applies to RBI-regulated entities, which it calls REs. Para 4 lists eleven groups: commercial banks (including foreign banks), small finance banks, payments banks, local area banks, regional rural banks, urban and rural co-operative banks, NBFCs in all four layers, All-India Financial Institutions, asset reconstruction companies and credit information companies. Para 6 says the principles apply to every model the RE uses, whether built in-house, bought from a third party or both.
It does not regulate vendors directly. But para 8 and para 45 say the RE is accountable for the outcomes of any model it uses, including third-party ones. So the obligations travel to you through contracts, due diligence and questionnaires. If you sell a chatbot, copilot or scoring feature to a bank or NBFC, your customer will need evidence from you to meet its own duties.
Is a GenAI chatbot a "model" under the draft?
Often, yes. Para 7(3) defines a model as a system, built in-house or bought, that takes data and assumptions as input, uses statistical, mathematical or other cognitive techniques, including AI and machine learning, to process them, and produces results used for business operations and decision-making. It includes "algorithms, analytics, interfaces, applications, decision-based rules, and other computational tools" which, by virtue of their use, have a material impact on decision-making, "irrespective of whether such tools are recognised as models by the RE".
The draft's own example is a spreadsheet loan-pricing calculator. It counts as a model once the RE uses it to set lending rates or credit terms. In our reading, a chatbot that explains loan terms, handles complaints or steers repayments is likely to count too, and so might an internal assistant whose output feeds credit or collections decisions. Confirm the reading for your own systems with counsel.
That matters for one reason in particular. Para 21 says no model may be used, relied upon or deployed unless it is in the model inventory. Shadow chatbots and unlogged pilots would sit outside that rule. If you do not know which AI tools your own teams use, start with a shadow AI audit.
What does the draft propose for customer-facing GenAI?

Most of the GenAI-specific text sits in paras 54 to 63. The table maps the main proposals to what an engineering team would build or keep as evidence. The right-hand column is our interpretation, not RBI text.
| Draft para | What the draft proposes | What you would build or keep |
|---|---|---|
| 59(ii) | What the draft proposes: Tell users they are interacting with an AI/ML system, and state its limitations | What you would build or keep: A visible disclosure at the start of every chat or call, a short limitations note, and a log of which version of the text was shown |
| 59(iii) | What the draft proposes: Let the customer switch to human assistance when asked | What you would build or keep: A tested handoff route, a fallback when no agent is free, and a record of each handoff |
| 59(i) | What the draft proposes: Controls against prompt injection and adversarial inputs, limits on session and context persistence, detection of anomalous usage | What you would build or keep: Input and output filters, least-privilege tool access, session expiry and context reset, rate limits and usage alerts |
| 54(2) | What the draft proposes: Control boundaries, through system controls or model design, to limit hallucinations | What you would build or keep: Answers grounded in approved sources, refusal outside scope, and figures such as rates and fees read from the system of record, never generated |
| 55 | What the draft proposes: Structured challenge processes, including red-teaming, especially for customer-facing and generative models | What you would build or keep: A written test set (injection, jailbreak, off-topic, mis-selling prompts), run again on every model, prompt or retrieval change |
| 56 | What the draft proposes: Enhanced controls for models that update automatically, including a clear scope of what may update | What you would build or keep: Pinned model versions, change control for prompts, indexes and model swaps, and tighter monitoring after any change |
| 57 | What the draft proposes: Enhanced documentation for traceability, reproducibility and auditability | What you would build or keep: Versioned prompts, configuration and evaluation results kept together for each release |
| 25 | What the draft proposes: A grievance process that covers consumer-facing models | What you would build or keep: A way to tag and route complaints about the bot into the normal grievance channel |
Para 54 adds more that applies to GenAI: define explainability thresholds, test for bias, monitor data drift, and manage variability in outputs through measures such as confidence scores. For a chatbot, most of this turns into an evaluation suite and a monitoring dashboard. Buyers may ask for the same evidence in the AI section of a security questionnaire.
How would a chatbot be tiered, and why does the tier matter?
The draft asks each RE to tier every model (paras 17 to 20). Tiering should weigh materiality (importance to the business, effect on outcomes, consumer implications) and complexity (difficulty of oversight, unstructured data, explainability limits). Para 20 says one factor must not water down another: low complexity should not pull a highly material model down a tier. For AI models, para 52 adds two more factors: how much the RE relies on the model's outputs, and how much autonomy it gives them.
The draft sets no tier thresholds. Each RE would write its own method. In our reading, a bot that only quotes published branch hours would sit lower than one that explains loan terms, handles complaints or nudges customers about repayments, because the second affects consumers and financial outcomes.
The tier drives the workload (para 18). Validation intensity and frequency follow the tier. Models with a high or equivalent tier need RMCB approval, which is the Risk Management Committee of the Board (paras 12 and 18). Validation reports go to the RMCB or its delegate within three months of completion (para 33). So a vendor whose product lands in a high tier should expect longer reviews and more document requests.
What do the kill switch and human oversight proposals mean in engineering terms?

Paras 60 to 63 hold the kill-switch proposal that press coverage picked up. The draft proposes three safeguards for AI models: human-in-command arrangements (human-in-the-loop or human-on-the-loop), "override, suspension, or deactivation mechanisms, including kill-switch arrangements", and periodic human review of model outputs and decisions to spot anomalies. It also asks the RE to consider automation bias, over-reliance and decision fatigue, to make sure reviewers understand the model well enough to challenge it, and to review overrides, incidents and near misses over time.
A practical build for a chatbot looks like this:
- One switch, no deploy. A feature flag or gateway setting that stops the bot, or drops it to a fixed FAQ or a human queue, within minutes and without a code release. For a multi-tenant product, make it per customer as well as global.
- Named owners. A short list of who may pull the switch, who is told, and how the fallback is staffed. Para 43 asks for fallback mechanisms such as manual intervention in continuity plans.
- Tested, not assumed. Run the switch in a drill and keep the result. A switch nobody has tried is a claim, not evidence.
- Sampled review. A regular human review of a sample of conversations and escalations, with a form that records what was found and what changed.
- An incident and near-miss log. Record overrides, bad answers caught by a reviewer and fixes, so the review process can improve, as para 63 proposes.
- Reviewer training. Reviewers need to know how the bot fails.
What will banks and NBFCs ask AI vendors to prove?
Paras 45 to 48 and 51 to 53 cover third-party models, and they are the ones a vendor will feel first. The draft proposes that the RE:
- validates third-party models independently, "notwithstanding any validation, certification, or assurance provided by the third-party provider" (para 46)
- puts third-party models under enhanced RMCB oversight whatever their risk tier (para 46)
- carries out due diligence on the provider's credibility, the model's methodology and limitations, and the suitability and quality of its data (para 47)
- secures contract terms for minimum technical documentation, audit rights for the RE and its supervisory authority, and continuity and exit arrangements (para 48)
- identifies risks and limits use where a provider does not disclose enough about the model (para 51)
- considers concentration risk and changes in model behaviour caused by provider updates (para 53)
That points to an evidence pack you can prepare before anyone asks:
| Draft theme | Evidence to have ready |
|---|---|
| Technical documentation (48) | Evidence to have ready: Architecture description, model and provider list, data flows, known limitations, evaluation method and results |
| Independent validation (46) | Evidence to have ready: Test sets and results the customer can re-run, access to a test environment, and a plain note on how you evaluate changes |
| Audit rights (48) | Evidence to have ready: A contract clause that gives the customer, and its regulator on request, a route to inspect the system or your evidence |
| Continuity and exit (48, 43) | Evidence to have ready: A fallback plan, data return and deletion on exit, and a notice period before you retire a feature |
| Provider changes (53) | Evidence to have ready: A change log for model, prompt and provider swaps, and a promise to tell customers before behaviour shifts |
| Oversight (60 to 63) | Evidence to have ready: Kill-switch design and drill record, review process, incident log |
Note para 46. A vendor's certificate or report does not replace the bank's own validation. ISO/IEC 42001 or a SOC 2 report can support due diligence, but under this draft they would not stand in for the RE's testing. Mono is not a certification body and issues no certificate. We help build the controls and the evidence. For the wider questionnaire, see our security questionnaire service and the post on ISO 42001 versus SOC 2.
Where does RBI's FREE-AI report fit?
RBI released the report of its FREE-AI committee on 13 August 2025 (Press Release 2025-2026/902). It sets out seven Sutras and 26 recommendations across six pillars. It is a committee report, not a direction. RBI's June 2026 press release names it as background to the draft. FREE-AI recommendations are not binding RBI directions. As of November 2026, RBI's AI-specific regulatory text is the draft model risk guidance.
In practice the two fit together. FREE-AI is the broader report on responsible AI in finance. The draft turns part of that thinking into model governance proposals. Our page on RBI FREE-AI implementation covers the report and how NBFCs and fintechs can put it into practice.
How does this overlap with DPDP?
A chatbot prompt often holds a name, phone number, account detail or loan status. That is personal data under the DPDP Act. The core DPDP duties, including notice, consent, rights, Rule 6 security safeguards and retention, start on 13 May 2027. So the chatbot needs both lenses: model controls from the draft, and data controls from DPDP. Our guide to DPDP-compliant AI architecture maps where prompts, logs and embeddings live, and our DPDP page for BFSI covers the sector.
One point to watch. The draft asks for model inventory and documentation to be kept for at least ten years after decommissioning (paras 23 and 24). In our reading, that is a record about the model: its owner, tier, tests and approvals. It is not a reason to keep customer conversations for ten years. Keep chat logs under your own DPDP retention schedule, and keep model records free of personal data. Confirm the exact reading with counsel.
What should you do now, while the draft is still a draft?
- Build the inventory. List every model and AI feature, who owns it, what it is used for, which provider it calls and what data it touches. The draft would not allow an unlisted model to run.
- Write down a tier for each AI feature. Use materiality, complexity, reliance and autonomy. Expect your customer to tier you, so know your own view first.
- Add the customer-facing basics. AI disclosure, a human handoff, injection defences and session limits are modest builds, and they match questions banks may ask.
- Install and test a kill switch. Drill it and keep the record.
- Write a red-team test set and run it on every release. Keep the results with the release notes.
- Prepare the vendor evidence pack above. Include a change log and an exit plan.
- Watch RBI's notifications. If final guidance appears, compare it with this draft line by line before you change your claims.
Do not tell buyers you are "RBI compliant" or "aligned with RBI model risk guidance" as if it were final law. Say what you have built, and show the evidence.
Sources: Reserve Bank of India, Press Release 2026-2027/528, "RBI issues draft Guidance on Regulatory Principles for Model Risk Management", 24 June 2026 (rbi.org.in, prid 63006). Draft "Guidance on Regulatory Principles for Model Risk Management, 2026", paras 2, 4 to 8, 12, 17 to 25, 29 to 33, 43, 45 to 48, 49 to 63 and 64 (draft text as hosted by MediaNama; the copy carries a placeholder reference and date). RBI notifications page and Draft Notifications/Guidelines page, checked 20 November 2026. RBI Press Release 2025-2026/902 on the FREE-AI committee report, 13 August 2025. Reading cross-checked against MediaNama, 25 June 2026, and S&R Associates, 24 August 2026. Digital Personal Data Protection Rules, 2025. Last reviewed 20 November 2026.
FAQ
Is RBI's model risk management guidance final or still a draft?
It is still a draft. RBI issued the draft Guidance on Regulatory Principles for Model Risk Management on 24 June 2026 (Press Release 2026-2027/528) and took comments until 24 July 2026. As of 20 November 2026 we found no final version on RBI's notifications or draft-notifications pages. The June 2026 draft is the latest text, and it has no circular number or effective date.
Does RBI's draft model risk guidance apply to a SaaS vendor directly?
No. The draft applies to RBI-regulated entities such as banks, NBFCs, asset reconstruction companies and credit information companies. But it makes the entity accountable for third-party models, and proposes independent validation, technical documentation, audit rights and exit terms in contracts. A vendor would feel the draft through customer due diligence and questionnaires.
Does the draft require a kill switch for AI chatbots?
The draft proposes it, but it is not final. Para 60 says a regulated entity should have human oversight for AI models, including override, suspension or deactivation mechanisms such as kill-switch arrangements, and periodic human review of outputs. Because the text is a draft, write that RBI proposes this, not that RBI requires it.
Does the draft require telling customers they are talking to an AI?
The draft proposes it. For models with a customer or external user interface, including generative AI, para 59 says users should get appropriate disclosures and warnings that they are interacting with an AI/ML system, together with its limitations, and an option to switch to human assistance when they ask. It is not final guidance.
Would ISO 42001 or SOC 2 satisfy the draft's third-party model rules?
Not on their own. Para 46 of the draft proposes that a regulated entity validates third-party models independently, notwithstanding any validation, certification or assurance the provider offers. An ISO/IEC 42001 certificate or a SOC 2 report can support due diligence, but would not replace the bank's or NBFC's own testing.
Is RBI's FREE-AI report binding?
No. The FREE-AI report, released on 13 August 2025, is a committee report with 26 recommendations, not an RBI direction. As of November 2026, RBI's AI-specific regulatory text is the draft model risk management guidance, which is itself still a draft.
