Short answer: as of 20 November 2026, SEBI's AI/ML guidelines for securities markets are still a consultation paper, but one rule is already in force. The SEBI (Intermediaries) (Amendment) Regulations, 2025 added Regulation 16C on 10 February 2025. Under it, a person regulated by SEBI that uses AI or ML tools, whether it built them or bought them from a technology provider, is "solely responsible" for the privacy, security and integrity of investor data, for the output of those tools, and for compliance with applicable laws. The longer guidelines are a different matter. SEBI published them as a consultation paper on 20 June 2025, took comments until 11 July 2025, and we found no final circular. So Regulation 16C is in force, and the consultation paper is the best guide to what SEBI may ask you to show.
This post is for brokers, asset managers, advisers and other intermediaries, and for the engineering teams and vendors that build AI for them. It sets out what is in force, what is only proposed, and what your systems should be able to prove either way. It is an engineering view, not legal advice. Whether a particular entity or tool is covered goes to your counsel.
What is in force, and what is only proposed?
Keep these two documents apart.
| Regulation 16C | Consultation paper | |
|---|---|---|
| Document | Regulation 16C: SEBI (Intermediaries) (Amendment) Regulations, 2025 | Consultation paper: "Consultation Paper on guidelines for responsible usage of AI/ML in Indian Securities Markets" |
| Date | Regulation 16C: Notified 6 February 2025, published in the Gazette on 10 February 2025 | Consultation paper: Issued 20 June 2025; comments closed 11 July 2025 |
| Status | Regulation 16C: In force from publication in the Gazette | Consultation paper: A proposal. No final circular found as of 20 November 2026 |
| Binding? | Regulation 16C: Yes. SEBI may act under Chapter V of the Intermediaries Regulations for a violation | Consultation paper: No. Its text says what SEBI proposed, not what it requires |
When this post says "SEBI proposed", it means the consultation paper. When it says "Regulation 16C requires", it means the regulation. We will update this post if SEBI issues a final circular.
What does Regulation 16C make you responsible for?

Regulation 16C(1) applies to a person regulated by the Board that uses "artificial intelligence and machine learning tools and techniques, either designed by it or procured from third-party technology service providers, irrespective of the scale and scenario of adoption". That person is solely responsible for three things:
- Data. The privacy, security and integrity of investors' and stakeholders' data, including data it holds in a fiduciary capacity, "throughout the processes involved".
- Output. The output arising from the use of the tools and techniques it relies upon or deals with.
- Compliance. Compliance with the applicable laws in force.
Three details matter for engineering. First, there is no size threshold. The words "irrespective of the scale and scenario of adoption" mean a small pilot is in scope as much as a large deployment. Second, buying a tool does not move the responsibility. A vendor's model, API or embedded feature is your responsibility to the same degree as one you built. Third, the Explanation to the regulation describes the tools widely: an application, software program or system offered to investors, or used internally to support investing and trading, to spread investment strategies and advice, or to carry out activities that include compliance work. Under Regulation 16C(2), SEBI may act under Chapter V of the Intermediaries Regulations for a violation.
The regulation lists no technical controls. It sets an outcome and puts the burden on you. When SEBI, an auditor or a client asks how you know a tool is safe with investor data, and who checked what it produced, you need a document, a log or a test result, not an assurance.
Does this cover AI tools you buy, and tools staff use informally?
On the text, yes to the first. Regulation 16C names tools "procured from third-party technology service providers", and the consultation paper goes further: it proposed that AI and ML services provided by third-party vendors be "deemed to be provided by the market participants", who stay responsible for compliance with all applicable laws.
Informal use is where firms lose track. An analyst who pastes client holdings into a public chatbot to draft a note has made that chatbot a tool "used internally", and the investor data in the prompt is now outside your control. We cannot say how SEBI would treat a specific case. The safe design is to know every tool in use, know what data goes into it, and be able to show you decided it was acceptable. A shadow AI audit is the usual way to find the tools nobody told you about.
What did SEBI's consultation paper propose?
The paper sets out five guiding principles, a list of model governance points, plus Annexure B, which lists possible controls for six risks. These are proposals only. We have grouped them by what engineering has to build.
| Principle in the paper | What SEBI proposed | What it means for your systems |
|---|---|---|
| Model governance | What SEBI proposed: A designated senior manager for the AI/ML models; an internal team that can monitor, test and explain them; documentation of development, validation and versioning; "ability to do replay"; periodic review; back-up or fall-back plans if the AI fails | What it means for your systems: A model register with an owner, a version history, a way to rerun a past decision, and a manual fallback you have tested |
| Third-party vendors | What SEBI proposed: A clear SLA and contract; monitoring of vendor performance; vendor AI "deemed to be provided by the market participants" | What it means for your systems: Vendor due diligence, contract clauses on logging, change notice and incident notice, and your own evaluation of vendor output |
| Testing | What SEBI proposed: Testing in an environment separate from live; shadow testing with live traffic before production; continuous monitoring; documentation of models; input and output data stored "for at least 5 years" | What it means for your systems: A staging environment, a shadow mode, drift monitoring, and an input and output store with a retention rule |
| Investor protection and disclosure | What SEBI proposed: Tell clients when AI/ML affects them, for example in algorithmic trading, portfolio management or advisory services. Disclose features, purpose, risks, limitations, accuracy results, fees and data quality in plain language | What it means for your systems: A client-facing disclosure that matches what the system does, kept current as the model changes |
| Fairness and bias | What SEBI proposed: Models should not favour or discriminate against one group of clients; data should be of adequate quality and broad enough; processes to find and remove bias | What it means for your systems: Bias tests on the data and the output, with results kept on file |
| Data privacy and cyber security | What SEBI proposed: A clear policy for data security, cyber security and privacy for AI/ML; lawful processing of investors' personal data; communicate technical glitches and data breaches to SEBI and others as the existing framework requires | What it means for your systems: Data flow maps, access control, an incident process that covers AI failures |
Some further proposals are very specific. The paper suggested that logs be kept "with full verbosity" so events can be rebuilt in order; that any request to unmask data be recorded; that AI/ML systems and their test cases be independently audited by a team "that has no role in development", with the findings shared with SEBI; that accuracy results be shared with SEBI periodically; and that firms have control to switch between manual and automatic feedback. Treat these as likely direction, not obligations.
What would the proposed tiered approach change?
The paper proposed a "regulatory lite" tier. It would apply to AI/ML used for anything other than operations that directly affect clients, such as internal compliance, surveillance and advanced cyber security tools. For those uses, only a short list of the governance points, the testing section and the data privacy and cyber security section would apply. The operations that SEBI treated as client-facing for disclosure were algorithmic trading, asset or portfolio management, and advisory and support services.
That split helps planning even before a final circular. A surveillance model that flags suspicious orders for your compliance team is not the same risk as a chatbot that tells a client what to buy. If SEBI keeps the tiers, your register should record whether each AI system touches clients. Our post on RBI's draft model risk guidance describes the same tiering idea for banks and NBFCs, and a firm that has both a broker and a lender arm can build one register for both.
What should engineering be able to prove?
Whatever SEBI finalises, a regulator, an auditor or a client will ask the same questions. This table turns Regulation 16C and the paper's proposals into evidence you can build and keep.
| Question you will be asked | Evidence to keep |
|---|---|
| Which AI/ML tools do you use, and who owns each? | Evidence to keep: A register listing every tool, whether built or bought, its purpose, owner, vendor, data it sees and whether it affects clients |
| What investor data goes into each tool? | Evidence to keep: A data flow map per tool, with the fields sent, where they are processed and whether the vendor keeps or trains on them |
| Can you explain or reproduce a past output? | Evidence to keep: Stored prompt or input, model name and version, configuration, retrieved documents and output, so a past decision can be replayed |
| How do you know it works before it goes live? | Evidence to keep: Test sets, results, a staging run, and shadow-mode comparison against the current process, signed off by a named person |
| How do you know it still works? | Evidence to keep: Drift and quality monitoring, with alert thresholds and a log of reviews |
| Who checked the checkers? | Evidence to keep: Review by a person or team that did not build the system, with a written report |
| What happens when it fails or goes wrong? | Evidence to keep: A manual fallback, a tested way to switch the AI off, and an incident runbook |
| What do clients know? | Evidence to keep: A disclosure of where AI is used, its limits and any accuracy figure you quote, with version history |
| Is investor data secure? | Evidence to keep: Access control, encryption, logging, and a record of every unmasking of protected data |
| How do you manage the vendor? | Evidence to keep: A contract with logging, change-notice and incident-notice terms, plus your own periodic review of the vendor's output |
Two of these deserve a closer look because they are the hardest to retrofit.
How do you build replay, logs and a kill switch?

Replay. SEBI's proposed "ability to do replay for diagnosis" is a design choice from day one. For each AI-assisted decision, store the exact input, the model name and version, the settings (such as temperature and system prompt version), any documents retrieved, the output and the person or process that acted on it. If a vendor can change the model under you without notice, replay will not give the same result. Pin versions where the vendor allows it, and write change notice into the contract. Our guide to LLM observability covers what to log and how to protect personal data inside those logs.
Kill switch and human review. The consultation paper's Annexure B lists "circuit breakers and kill switches" and human-in-the-loop or human-around-the-loop mechanisms as possible controls for runaway or failing models, and it says the governance structure should provide for human accountability for AI-driven decisions. Treat a kill switch as a tested feature, not a promise. A usable one is a flag that stops the AI path within minutes, routes the work to a manual process, and is owned by a named role that has authority to use it. Test it on a schedule.
Input and output retention. SEBI proposed keeping inputs and outputs "for at least 5 years". That is a proposal, not a rule. Build retention as a setting you can change per system, so a final circular does not mean a rebuild. Keep the stored inputs as lean as the replay purpose allows, because they may contain personal data.
What does this mean for AI vendors selling to brokers and AMCs?
If you build AI for a regulated intermediary, Regulation 16C lands on your customer, and the customer will push the evidence requirement down to you. The consultation paper's "deemed to be provided by the market participants" line tells you how SEBI may see the relationship. Expect questions in your security review that read like the table above.
Be ready to supply: a list of the models and sub-processors you use; a statement on whether customer data trains any model; per-customer logging that the customer can export; version-change notice; a documented way to switch the AI feature off for one customer; your own testing and monitoring results; and an incident notice process with times. Our guides to answering AI security questionnaires and to selling a SaaS product to banks under DPDP cover the same evidence from the vendor side. The paper also floated reporting the names of AI vendors to SEBI to monitor concentration, so a customer may ask which cloud and model providers you rely on.
How does this overlap with DPDP and RBI rules?
DPDP sets personal data duties across all sectors. Under Section 8(7) of the DPDP Act, a Data Fiduciary must erase personal data once the purpose is served or consent is withdrawn, unless a law requires retention. Rule 8(3) of the DPDP Rules, 2025 requires a minimum one-year retention of personal data, traffic data and processing logs for certain purposes. Those duties start on 13 May 2027. A SEBI retention period, if it becomes law, could be the "law requires retention" that Section 8(7) refers to. Until it does, do not assume five years is mandatory, and do not delete audit records that another SEBI rule already makes you keep.
RBI is a separate regulator with its own AI texts. If your group also has a bank, NBFC or lender, read RBI's FREE-AI framework page. One tool in a group may fall under both regulators.
What should an intermediary do in the next 90 days?
- Week 1 to 2. Build the register of every AI/ML tool in use, built or bought, with owner, vendor, data and whether it affects clients. Add the tools staff use informally.
- Week 2 to 4. Name a senior owner for AI/ML oversight. This is a proposal in the paper, and it is also the person who will answer if SEBI asks.
- Week 3 to 6. Map the investor data that flows into each tool and mark any that leaves your control. Stop any flow you cannot justify.
- Week 4 to 8. Add replay logging to client-facing tools. Pin model versions where you can.
- Week 6 to 10. Run shadow tests and bias tests on client-facing tools and keep the results. Have a person who did not build the tool review them.
- Week 8 to 12. Write and test the kill switch and the manual fallback. Update your client disclosure to match what the system does.
- Ongoing. Review vendor contracts for logging, change notice and incident notice. Watch SEBI's circulars and consultation page for a final text.
Mono's AI governance work builds this register, the evidence and the engineering controls with your team. Mono is not a law firm and does not certify anyone. Counsel decides whether and how each SEBI provision applies to your entity, and SEBI itself decides what compliance means.
Sources: Securities and Exchange Board of India (Intermediaries) (Amendment) Regulations, 2025, notified 6 February 2025 and published in the Gazette of India, Extraordinary, 10 February 2025, inserting Regulation 16C (sebi.gov.in/legal/regulations/feb-2025). SEBI, Consultation Paper on guidelines for responsible usage of AI/ML in Indian Securities Markets, 20 June 2025, paragraphs 5.1 to 5.5, 6 and Annexure B (sebi.gov.in/reports-and-statistics/reports/jun-2025). Digital Personal Data Protection Act, 2023, Section 8(7). Digital Personal Data Protection Rules, 2025, Rule 8. Last reviewed 20 November 2026.
FAQ
Has SEBI issued final AI/ML guidelines?
Not as far as we could find as of 20 November 2026. SEBI published a consultation paper on responsible usage of AI/ML in securities markets on 20 June 2025 and took comments until 11 July 2025. A rule is already in force: Regulation 16C of the SEBI (Intermediaries) Regulations, inserted on 10 February 2025, which makes a regulated person solely responsible for its AI/ML tools.
What does SEBI Regulation 16C require?
A person regulated by SEBI that uses AI or ML tools, whether built in-house or bought from a third-party technology provider and at any scale, is solely responsible for the privacy, security and integrity of investor data, for the output the tools produce, and for compliance with applicable laws. SEBI may take action under Chapter V of the Intermediaries Regulations for a violation.
Does Regulation 16C apply if we use a vendor's AI tool?
Yes, on the text. Regulation 16C refers to tools procured from third-party technology service providers. Responsibility stays with the regulated person. The consultation paper also proposed that vendor AI services be deemed to be provided by the market participant. Keep vendor contracts, logs and your own testing so you can show you managed the risk.
Does SEBI require AI audit logs to be kept for 5 years?
Not yet. The consultation paper proposed that market participants store input and output data for at least 5 years, and keep logs with full verbosity. It is a proposal. Build retention as a configurable setting, and take your counsel's advice on how it sits with DPDP erasure duties under Section 8(7), which start on 13 May 2027.
Does SEBI require a kill switch or human oversight for AI?
Regulation 16C does not mention either. The consultation paper's Annexure B lists circuit breakers, kill switches and human-in-the-loop controls as possible measures for model failure, and it asks for human accountability for AI-driven decisions. These are suggestions in a consultation paper. They are still a sensible design for any client-facing AI system.
