RBI and SEBI AI Rules: What Indian Financial Firms Should Do Now
By Vaaryan Team • • 6 min read
India's financial regulators are converging on the same message: you may use AI, but you own every model you use, you must be able to explain and override it, and customers must know when they are dealing with one. In mid-November 2026, SEBI's chairman said the regulator will soon issue detailed guidelines on responsible AI in capital markets. The RBI's draft model risk management guidance, released in June 2026, already spells out what that looks like for banks and NBFCs. Neither is final, but together they are specific enough to act on now.
This post explains what was said, why it matters and what Indian banks, NBFCs, brokers, asset managers and the fintechs that serve them should do next. For the broader use-case picture, see our AI in Indian banking and financial services playbook.
What did SEBI say about AI this week?
Speaking at the inauguration of SEBI's pavilion at the India International Trade Fair in New Delhi, chairman Tuhin Kanta Pandey said SEBI will soon issue detailed guidelines on the responsible use of AI in the securities market (report). The emphasis, as reported:
- Disclosures — market participants would need to disclose how they use AI.
- Accountability — regulated entities remain responsible for what their AI systems do.
- Human in the loop — critical decisions stay subject to human supervision.
- Investor protection and market integrity — innovation must not come at their expense.
No draft text or date was given. The guidelines are expected to build on SEBI's June 2025 consultation paper on guiding principles for AI and machine learning in the securities market. Until the final text is out, treat the specifics as direction rather than obligation.
What does the RBI's model risk draft require?
The RBI's draft Guidance on Regulatory Principles for Model Risk Management, 2026 was put out for consultation in June 2026 and, as far as we can tell, has not yet been finalised. It is broader than an "AI rule": it defines a model as any system whose outputs materially affect decisions, which captures credit scorecards, pricing spreadsheets, machine learning models and generative AI alike (summary by Vinod Kothari Consultants).
The main expectations in the draft, paraphrased:
| Area | What the draft expects |
|---|---|
| Governance | A board-approved model risk management framework; high-risk models reviewed at board risk committee level |
| Inventory | A living inventory of all models, including decommissioned ones |
| Risk tiering | Models classified by risk, with controls proportionate to tier |
| Validation | Independent validation by the institution, even when a vendor certifies the model |
| Third-party models | Contractual rights to documentation and audit; outsourcing the model does not outsource the risk |
| Explainability | Higher bar for material decisions such as credit; extra controls where a model cannot explain itself |
| Fairness | Assess and correct discriminatory outputs |
| Human oversight | Human-in or on-the-loop review, override and kill-switch mechanisms |
| Customer-facing generative AI | Tell customers they are talking to AI, offer a human on request, control hallucinations, defend against prompt injection, red-team before launch |
Once final, it is expected to replace the credit-risk-model chapter of the RBI's 2002 guidance note on credit risk management. It sits alongside the RBI's 2025 FREE-AI committee report, which set out principles for responsible AI adoption.
Why does this matter beyond compliance teams?
Three reasons.
1. Vendor AI is now your AI. Many institutions use AI through vendors — a document-extraction service, a voicebot, a fraud engine, a general-purpose LLM API. The RBI draft is explicit that buying a model does not transfer the risk. Procurement, legal and technology teams will need validation evidence and audit rights in contracts, not just a security questionnaire.
2. Generative AI gets specific controls. Chatbots and voicebots are singled out: disclosure, a path to a human, hallucination controls and prompt-injection defences. These are product and engineering decisions as much as policy ones, and they shape the user experience.
3. "Kill switch" is an architecture requirement. Being able to disable or override a model quickly means the surrounding workflow must keep working without it. That is far easier to design in than to bolt on.
What should Indian financial firms do now?
You do not need to wait for final text. The following steps are useful under any reasonable final version of either regulator's rules.
A 30-day checklist
- Build an AI and model inventory. Every model, rule engine, scoring sheet, LLM prompt flow and vendor AI service, with an owner, purpose, data used and the decisions it affects.
- Tier each entry by risk. Customer-impacting decisions (credit, pricing, claims, investment advice) at the top; internal drafting and search assistants lower.
- Review vendor contracts. Do you have rights to model documentation, validation evidence, change notifications and audit? Where is data processed and retained?
- Map human oversight. For each high-tier system, who reviews outputs, when, and with what authority to override?
- Check customer-facing assistants. Do they disclose they are AI? Can customers reach a human easily? Is there logging good enough to reconstruct a conversation?
- Define the off switch. For each high-tier system, what happens operationally if it is turned off tomorrow?
Over the next quarter
- Stand up an evaluation practice. Validation for generative AI means test sets, regression runs on every change and tracked quality metrics — the same discipline we describe in our LLMOps guide.
- Red-team customer-facing AI. Test for prompt injection, data leakage, mis-selling language and confident wrong answers, and record what you fixed.
- Write the board paper. Boards will be asked to approve the framework. A short paper with the inventory, tiering and the top risks is a good start.
- Align with DPDP work. Personal data flowing into models and vendors needs the same mapping for data protection; do both exercises together (responsible AI in India).
What should fintechs and AI vendors expect?
If you sell into banks, NBFCs, brokers or AMCs, expect sharper due diligence: requests for model documentation, evaluation results, explainability approaches, incident processes and audit access. Vendors who can hand over a clean model card, an evaluation report and a clear data-flow diagram will close deals faster than those who cannot. Treat this as a product requirement, not paperwork.
What is still uncertain?
- Final text and timelines. Neither the RBI's model risk guidance nor SEBI's AI guidelines are final at the time of writing, and details — including effective dates and transition periods — may change.
- Proportionality for smaller entities. How tiering applies to smaller NBFCs and intermediaries will matter a great deal in practice.
- Interaction with other rules. How these guidelines line up with DPDP obligations, outsourcing rules and IRDAI's approach for insurers is still to be seen.
We will update this post when final versions are published.
How Vaaryan approaches this
Our view is that good regulation and good AI engineering ask for largely the same things. An inventory, evaluation sets, human review of consequential outputs, clear disclosure and a graceful fallback are what make AI systems reliable — regulators are simply writing them down.
When we work with financial services teams, we start by making those controls part of the design rather than a review at the end: deciding where a human approves, what the assistant says when it is unsure, what gets logged and how the workflow runs if the model is switched off. We prefer to stay model-agnostic so that a model can be replaced if it fails validation, and we document evaluation results in a form a risk team can read. That is slower than a demo for the first two weeks and much faster from then on.
Frequently asked questions
Has SEBI issued AI guidelines for capital markets?+
Not yet in final form. In mid-November 2026 SEBI's chairman said detailed guidelines on responsible AI use in capital markets would be issued soon, focusing on disclosures, accountability of regulated entities and human oversight of critical decisions. They build on a consultation paper SEBI published in June 2025.
What is the RBI's draft model risk management guidance?+
It is a draft released for public consultation in June 2026, titled Guidance on Regulatory Principles for Model Risk Management, 2026. It asks regulated entities to run a board-approved framework covering all models, including AI, machine learning, generative AI and vendor-supplied models, with an inventory, risk tiering, independent validation, monitoring and human oversight.
Do the RBI rules apply to AI models bought from a vendor?+
Yes, as drafted. Using a third-party or vendor model does not move the risk to the vendor. The regulated entity is still expected to validate it, monitor it and secure contractual rights to documentation and audit access.
Should we pause AI projects until the final rules arrive?+
No. The direction of both regulators is clear enough to build for now: an inventory of AI systems, documented validation, human oversight of material decisions, disclosure to customers and the ability to switch a model off. Designing these in today costs far less than retrofitting them later.
Do these rules apply to fintechs that are not regulated entities?+
Directly, they apply to regulated entities. In practice, fintechs and AI vendors that serve banks, NBFCs or market intermediaries will be asked to meet the same expectations through contracts, due diligence and audit rights.
Work with Vaaryan
Turning AI ideas into products that ship
Vaaryan is an AI consulting and product company based in India. We help teams find the right use cases, design the experience, and engineer AI systems that hold up in production.