SEBI's June 2025 consultation paper on responsible AI sets out fairness, accountability, transparency, robustness, privacy, and explainability expectations for every intermediary using AI or ML in securities markets — stockbrokers, investment advisers, research analysts, portfolio managers, exchanges, depositories, mutual funds, and AIFs. It's the first SEBI document to set governance principles for AI specifically, reaching beyond the existing algo-trading oversight regime into every AI system an intermediary runs.
Two things make this consultation worth treating seriously now, before a final circular forces the issue. First, it's a genuine scope expansion — SEBI is no longer just watching trading algorithms, it's watching AI in onboarding, advisory, surveillance, and compliance too. Second, its six principles overlap heavily with RBI's FREE-AI work and with global frameworks like the EU AI Act and NIST AI RMF, which means an intermediary that builds one control framework correctly can satisfy several regulators from the same evidence, rather than running separate exercises for each.
Scope. The consultation applies to every SEBI-registered intermediary using AI or ML in any capacity — onboarding, advisory, research, trading, execution, surveillance, risk management, operations, compliance. There's no materiality threshold. If an AI system touches a regulated function, it's in scope, full stop.
The six principles, and what they actually ask for
- Fairness — AI systems can't produce discriminatory outcomes across client segments. Demographic parity, equal opportunity, and equitable access to AI-driven services are all explicitly in scope.
- Accountability — every AI-driven decision needs a clear, documented, human owner — board, senior management, or a named function.
- Transparency — investors must be told when AI materially influences a decision about them, and given a real way to understand the basis for it, not a disclaimer buried in a terms-of-service document.
- Robustness — AI systems have to perform reliably across market conditions, including stress scenarios, with adversarial inputs and drift detected and managed rather than discovered after the fact.
- Privacy — personal data processed by AI must comply with DPDP Act 2023 obligations: consent, purpose limitation, data minimisation, localisation for Indian investor data.
- Explainability — material AI-driven decisions — credit, KYC, trade execution, advisory recommendations — must be explainable to investors in plain language. A black box that produces a recommendation nobody can walk back through does not satisfy this.
Mapping the principles to actual controls
| SEBI principle | Controls required | Evidence for examination | Where most firms are actually sitting today |
|---|---|---|---|
| Fairness | Pre-deployment bias assessment, protected-attribute analysis, continuous fairness-metric monitoring | Per-model bias assessment report; fairness metrics tracked over time | Bias testing done once at launch, never repeated; no formal threshold defined |
| Accountability | Named owner per AI system, board-approved policy, committee approval for high-risk AI | Accountability register; board resolution; signed committee approvals | Accountability assigned informally; board policy exists but doesn't cover every AI system in production |
| Transparency | Client disclosure for AI-driven decisions, model card documentation, investor grievance channel | Disclosure templates; model cards; grievance resolution records | No AI disclosure reaches investors at all; grievance process only covers human decisions |
| Robustness | Adversarial testing, drift detection with alert thresholds, stress scenario testing | Adversarial test results; drift logs; stress test reports | Production monitoring exists with no formal drift threshold; no testing against market-shock scenarios |
| Privacy | PII detection at inference, consent verification, purpose-limitation enforcement | PII detection logs; consent records; DPDP compliance documentation | DPDP handled at the data layer only; the inference pipeline itself is never checked |
| Explainability | Attribution (SHAP/LIME or equivalent), human-readable explanation generation, escalation for unexplainable decisions | Explanation logs; method documentation; investor explanation records | Explanations generated ad hoc after a complaint, not produced by default |
Who this actually bites, by entity type
- Stockbrokers — AI in KYC, execution, order routing, margin calculation, client communication. Fairness, robustness under stress, and disclosure of AI use in trading are the principles that get tested first.
- Investment advisers and research analysts — AI in portfolio construction and recommendation generation. Explainability — can the investor actually understand why the AI recommended this — and fairness across client segments carry the most weight.
- Portfolio managers — AI in allocation, rebalancing, risk management. All six principles apply, but accountability for AI-driven allocation decisions is the first question a regulator asks.
- Exchanges and depositories — AI in surveillance, settlement, risk management. Robustness under market stress and transparency of surveillance methodology dominate.
- Mutual funds and AIFs — AI in fund management, risk scoring, investor communication. Fairness across investor classes and explainability of fund-level AI decisions are what gets scrutinized.
The dual-regulated problem: when one AI system answers to two regulators
A bank that's also a SEBI-registered investment adviser — or a universal bank with a broking arm, or an NBFC with an investment management license — can face SEBI's six principles and RBI's supervisory expectations simultaneously, from the same AI system. The principles overlap substantially: fairness and robustness themes recur across both regulators' language, and explainability requirements look nearly identical on paper.
The real cost isn't the overlap — it's that most institutions still run two separate compliance exercises for the same AI system, producing two separate evidence packs with two separate owners, neither of which talks to the other. I've watched a model risk team and a SEBI compliance team independently commission two bias assessments on the identical advisory model in the same quarter, neither aware the other existed. That's not a regulatory problem; it's an org-chart problem wearing a compliance costume.
What to do about it
- Inventory every AI system that touches a SEBI-regulated function, separately from your RBI-facing AI inventory, then find the overlap — it's usually larger than either team expects.
- Run one bias assessment per model, tagged to both SEBI's fairness principle and whatever RBI fairness obligation applies, instead of commissioning two.
- Stand up an investor-facing disclosure template for AI-influenced decisions now, ahead of SEBI's final circular — this is the control most firms have zero of today.
- Define your drift and stress-test thresholds in writing, with sign-off, before an examiner asks what "robustness" means in your institution specifically.
- If you're dual-regulated, name a single owner for AI governance evidence across both regulators — not two owners who occasionally email each other.
Complete India AI governance framework
SEBI AI governance for intermediaries
Algo trading oversight, advisory fairness, market surveillance — sector pages for all SEBI entity types.
Explore →RBI eMRM 2026: independent validation
Mandatory independent validation for all AI models. What banks and dual-regulated entities must implement.
Read →RBI FREE-AI Framework: 26 recommendations
Fairness, Reliability, Explainability, Ethics — how SEBI Principle 1/2/3 maps to RBI FREE-AI controls.
Read →SEBI responsible AI ready — before the circular lands
Talk to Zytra's India team to map your AI systems against SEBI's responsible AI principles and RBI's requirements in a single integrated governance framework.