India's securities markets move over ₹300 trillion in equity and derivatives turnover annually. A substantial fraction of that volume is driven by AI/ML-powered algorithmic strategies. SEBI has recognized that governing this AI infrastructure requires a framework beyond the 2020 algo-trading circular's speed bumps and randomization requirements.
Two SEBI instruments constitute that framework. The February 4, 2025 circular on algorithmic trading (effective April 1, 2026) imposes pre-exchange approval requirements and non-delegable liability. The June 20, 2025 Consultation Paper on responsible AI/ML usage proposes seven governance pillars that will become mandatory guidelines. The April 2026 deadline has passed. Brokers not yet compliant face examination risk at their next SEBI inspection.
The February 2025 Algo-Trading Framework
SEBI Circular SEBI/HO/MIRSD/MIRSD-PoD/P/CIR/2025/0000013, issued February 4, 2025, mandated compliance by April 1, 2026. Three core AI governance obligations:
Pre-approval of AI algo strategies
Every AI/ML-driven algo strategy must be pre-approved by the exchange before live deployment. The exchange review requires: model logic description, backtesting and forward-testing performance data under stressed conditions, risk management parameters including kill-switch configurations, and named senior officer accountability. Strategies modified after approval require re-approval before the modified version deploys — this creates a model change management obligation: brokers must maintain version control and ensure the approved version is the deployed version at all times.
Non-delegable liability for procured AI
The broker or trading member is solely liable for AI/ML systems used in operations, including systems procured from third-party vendors. A broker using an AI order management system, AI signals platform, or AI portfolio construction tool is fully accountable for that system's compliance. The vendor cannot be contractually assigned the liability. The broker must independently explain, document, and defend the AI system in a SEBI inspection without relying on vendor-provided documentation alone.
Risk checks and automated position limits
All AI-driven strategies must operate within automated risk check infrastructure: real-time position monitoring, automated halt triggers when limits are breached, and mandatory order-level risk filters. The AI strategy cannot override the risk infrastructure — signals beyond approved limits must be blocked at the risk layer, not just flagged for manual review.
The June 2025 Consultation Paper: Seven Governance Pillars
Pillar 1: Senior Management Accountability
Board-level ownership for AI/ML systems used in regulated activities. A designated board committee or named executive must own the AI governance framework. A broker whose AI governance exists only as a technology team responsibility, without board-level ownership, does not meet this pillar.
Pillar 2: Investor Protection
Transparency to clients about AI/ML use in their services. Investment advisers using AI for recommendations must disclose AI involvement. Brokers using AI for order routing must maintain records demonstrating AI-driven routing does not systematically disadvantage client order fills. AI-related complaints must be classifiable and trackable in SEBI's SCORES system.
Pillar 3: Model Testing and Continuous Monitoring
Pre-deployment validation and ongoing production monitoring. Monitoring must detect performance degradation, behavioral shifts under new market regimes, and correlation breakdown between the model's signal and the outcome it was trained to predict. This pillar aligns with RBI MRMF Section 6 — dual-regulated entities (SEBI-registered advisers that are also NBFCs) face simultaneous obligations under both frameworks.
Pillar 4: Fairness and Bias Mitigation
Testing for systematic discrimination in AI outputs. For securities participants, this means assessing whether AI systems produce systematically worse outcomes for specific client categories — retail vs institutional, small-ticket vs large-ticket, geographic segments. AI-driven order routing that systematically provides worse execution to retail accounts is a fairness concern under this pillar.
Pillar 5: Data Privacy Safeguards
DPDP Act 2023 alignment for all AI training and inference data. Securities AI uses extensive personal data: trading history, financial profiles, behavioral patterns, device identifiers. Training data must have a lawful basis under DPDP; inference-time processing (real-time decisions about a specific customer's order) requires consent or a recognized legitimate purpose. Brokers that implemented AI models without a DPDP data mapping exercise have a gap here.
Pillar 6: Third-Party Vendor Risk
Non-delegable liability operationalized as a vendor governance obligation. The regulated entity must have governance visibility into third-party AI tools — independent validation of vendor AI models, contractual audit rights, change notification before model updates, and documented accountability for erroneous outputs.
Pillar 7: Operational Resilience
Circuit breakers, fallback mechanisms, and AI incident response. The kill-switch obligation intersects with SEBI's RAIDAR (Risk-Adjusted Intraday Drawdown Automated Response) framework. Operational resilience requires: defined triggers for AI halt, tested fallback to non-AI execution, incident response protocols, and post-incident review documentation. SEBI inspection teams test kill-switch activation records, not just policies.
SEBI AI Governance Compliance Checklist
- ☐ AI/ML system inventory: all systems in production listed with function, deployment date, accountability owner
- ☐ Exchange pre-approval: all AI algo strategies have exchange approval; approved documentation matches deployed version
- ☐ Model change management: version control for AI algo strategies; re-approval process for material changes
- ☐ Senior management accountability: board-level ownership designated and documented
- ☐ Vendor AI governance: contracts include model documentation access, change notification, audit right, non-delegation acknowledgment; independent validation of vendor AI models
- ☐ Risk check infrastructure: automated position limits active and tested; AI strategy cannot override risk layer
- ☐ Kill-switch: activation criteria defined; tested within last 12 months; test records on file
- ☐ Investor disclosure: clients informed of AI/ML use; disclosure mechanism documented
- ☐ Fairness testing: systematic outcome analysis across client categories; differential execution quality checked
- ☐ DPDP data mapping: training and inference data mapped; lawful basis documented
- ☐ Ongoing monitoring: daily performance monitoring operational; alert thresholds and escalation process defined
- ☐ SEBI inspection pack: AI inventory, pre-approval records, validation records, monitoring records, kill-switch test records — organized and accessible