Does SEBI Clause 26 system audit cover AI models in stockbroker OMS and RMS?

SEBI Master Circular for Stock Brokers (SEBI/HO/MIRSD/MIRSD-SEC-3/P/CIR/2023/025, July 31, 2023) Clause 26 mandates annual system audit by a CERT-In empanelled auditor covering the broker's Order Management System (OMS) and Risk Management System (RMS). Since virtually all modern OMS and RMS platforms incorporate AI/ML — AI-powered smart order routing, ML-based risk scoring, algorithmic order slicing — these AI components fall within Clause 26 scope as subsystems of the covered platforms. The governance gap: Clause 26 was designed as a cybersecurity and operational resilience audit standard. It certifies that the system works, not that its AI decisions are sound. A CERT-In empanelled auditor conducting a Clause 26 audit of a broker's AI-powered SOR will verify that the system is accessible only to authorized users, has disaster recovery procedures, and meets uptime requirements — not that the SOR AI's routing logic is free from adverse selection patterns, not that its venue preference does not systematically disadvantage certain order types, and not that its training data represents the current market microstructure. The result: stockbrokers comply with Clause 26 and still have no governance framework for the AI inside their compliant systems. This is a structural gap that grows as OMS/RMS AI becomes more sophisticated — an annual IT audit covering a system that makes 10,000 AI-mediated routing decisions per second is not proportionate to the AI risk.

Who owns the AI governance obligation for smart order routing in multi-venue Indian markets?

SEBI's SOR guidelines require each smart order router to be approved by the exchange where the broker operates. India's stockbrokers operate simultaneously on NSE, BSE, and NSE-SME, with some also active on MCX-SX. A multi-venue SOR that routes order flow across these exchanges uses AI to optimize execution quality — routing to the venue offering the best price, deepest liquidity, or lowest market impact at a given millisecond. The approval process requires each exchange to separately approve the SOR operating on its platform. But no entity reviews the SOR AI model holistically — neither NSE nor BSE has visibility into how the AI routes orders between them, only into how it behaves on their own platform. SEBI, as the cross-exchange regulator, has not published guidance on multi-venue SOR AI governance. The resulting governance gap: the AI logic that determines how order flow is allocated between exchanges — which directly affects price formation on each exchange — has no governance owner at the level of the AI model as a whole. Each exchange approves a partial view; the broker owns the whole model but faces no MRMF-equivalent validation obligation for it.

What is the broker's AI governance liability when DMA clients deploy their own AI models?

SEBI's Direct Market Access circular (SEBI/MRD/SE/Cir-4/2008, April 3, 2008) makes brokers providing DMA to institutional clients liable for all orders routed via DMA, including orders generated by client-side AI models. The broker is required to implement pre-trade risk filters — automated checks on order price, quantity, and value — but no SEBI circular defines the broker's obligation to validate the client's AI logic that generates those orders before granting DMA access. Three scenarios illustrate the governance gap: (1) A hedge fund's reinforcement learning model develops a market-making strategy that generates unusual order patterns consistent with layering — the broker's risk filter passes the individual orders (each within limits) but the AI pattern is manipulative. SEBI can pursue the broker under FUTP Regulations for inadequate surveillance of DMA orders; (2) A client's AI order-routing model malfunctions during a market stress event, flooding the exchange with orders that trigger circuit breakers. The broker bears SEBI liability for the market disruption even though the AI was entirely client-owned; (3) A client updates their AI model without notifying the broker — the new model behavior is outside the risk parameters the broker evaluated when granting DMA. No SEBI circular requires brokers to implement model change notification in DMA agreements.

What AI governance obligations arise when a broker re-enables trading after a kill-switch event?

SEBI's algorithmic trading circulars (CIR/MRD/DP/01/2013 and subsequent updates) require stockbrokers to maintain a kill switch to immediately halt all algorithmic trading in an emergency. For AI-driven trading systems — reinforcement learning market-making models, ML-based execution algorithms — the kill switch stops the AI. SEBI's circulars are silent on what model governance obligations apply to re-enabling the AI after a kill-switch event. Three unaddressed questions: (1) If the kill switch was triggered because the AI was generating anomalous order patterns, is the broker required to investigate and remediate the AI before restart — or can it simply restart and hope the anomaly was transient? (2) If the AI was halted during a market stress event and market conditions have changed significantly, is the AI still operating within its validated range after restart? The MRMF concept of 'material change in operating environment' would suggest a validation requirement, but MRMF does not explicitly cover trading AI restart scenarios; (3) If multiple brokers' AI systems are simultaneously restarted after a market-wide circuit breaker event, coordinated restart behavior could itself be market-destabilizing. SEBI has no guidance on AI restart sequencing or coordination. The kill-switch governance framework is essentially a one-way instrument — SEBI has detailed stop-the-AI obligations and zero re-enable-the-AI governance framework.

AI Governance · SEBI · Stockbroker · Smart Order Routing · Clause 26 · India BFSI

SEBI Stockbroker AI Governance: Smart Order Routing AI Ownership, Clause 26 System Audit Gap, DMA Client AI Liability and Kill-Switch Restart Obligations India 2026

SEBI's broker regulatory stack — Master Circular Clause 26 system audit, algorithmic trading circulars, DMA rules — governs AI in stockbroker OMS and RMS without any AI-specific methodology. Smart order routing AI ownership is fragmented across exchanges with no holistic model governance. DMA broker bears SEBI liability for client-side AI. Kill-switch restart governance is a complete void. Quarterly AI monitoring mandatory for AMCs is absent for brokers. Zero global competitor coverage.

Zytra Research Team · October 2026 · 12 min read

📖 Part of the India BFSI AI Governance Complete Guide 2026 — RBI, SEBI, IRDAI, NPCI & DPDP in one place.

India's stockbroker community — NSE and BSE registered brokers ranging from full-service brokerages to discount brokers — deploys AI throughout the trading infrastructure: smart order routers deciding execution venue millisecond by millisecond, risk management AI scoring client positions in real time, pre-trade compliance AI checking order parameters against client mandates, and DMA infrastructure enabling institutional client AI models to interact directly with exchange matching engines.

SEBI's broker regulatory stack was designed for an era of rule-based algorithmic trading. The Master Circular's Clause 26 system audit, the algo trading circulars from 2013–2021, and the DMA framework from 2008 collectively create governance obligations that apply to AI-powered trading systems — but through regulatory language that predates the AI era. The result is a set of structural governance gaps that are invisible to global AI governance vendors focused on US/EU frameworks and visible only through careful reading of India's SEBI broker-tier regulation.

Clause 26 System Audit: The AI Methodology Vacuum

SEBI Master Circular Clause 26 mandates annual system audit by a CERT-In empanelled auditor. The audit covers the broker's OMS, RMS, and trading infrastructure. Modern OMS platforms (Omnesys, FinAstra, proprietary broker platforms) incorporate AI for order routing, execution quality optimization, and position risk scoring. Modern RMS platforms use ML for margin breach prediction, concentration limit monitoring, and client behavior anomaly detection.

A Clause 26 audit of these AI-embedded platforms certifies: access controls are adequate, disaster recovery is tested, data integrity is maintained, and change management procedures exist. It does not assess: whether the AI's routing decisions are systematically biased toward or against certain instrument types, whether the risk scoring AI has been validated on stressed market conditions, whether the AI's performance has degraded since last validation, or whether the AI training data reflects current market structure.

The practical consequence: brokers conduct Clause 26 audits annually, receive clean audit reports, and have no documented evidence that their AI systems are performing correctly. SEBI, reviewing broker audit reports, has no visibility into AI model risk within compliant systems. This is the stockbroker equivalent of the credit model risk management gap that led to RBI's MRMF — but for SEBI-regulated trading infrastructure, there is no MRMF equivalent.

Smart Order Routing AI: The Cross-Exchange Governance Void

SEBI's SOR framework requires exchange-level approval for smart order routing systems. Each exchange reviews the SOR's interaction with its own trading platform — does it respect circuit breakers, does it comply with tick size rules, does it handle exchange-specific order types correctly? Exchange approval is venue-specific.

Multi-venue SOR AI — the AI logic that decides, given an order to execute, whether to route to NSE, BSE, or split across venues — is reviewed by no single entity holistically. NSE and BSE each see a partial view of the SOR's behavior on their own exchange. Neither exchange can assess how the AI allocates flow between exchanges, because neither exchange has visibility into the other's execution data for the same orders.

This creates three ungovernmed AI risk scenarios:

Systematic venue preference: SOR AI that learns to route predominantly to one exchange (because historical data shows marginally better fills) concentrates market impact and potentially disadvantages clients who might benefit from cross-venue liquidity — without any governance mechanism to detect this pattern
Adverse selection: SOR AI that routes toxic flow (orders likely to move the market) to one exchange while routing benign flow to another — a behavior that could constitute market manipulation under FUTP Regulations if deliberate, or model drift if not
Cross-venue coordination: Multiple brokers' SOR AI systems making similar routing decisions simultaneously could create coordinated market impact without any single entity's intent — an emergent AI market risk that SEBI's current framework cannot detect or address

DMA Client AI: Broker Liability Without Broker Visibility

DMA clients — institutional investors, proprietary trading firms, and hedge funds — deploy their own AI models on the broker's DMA infrastructure. The broker's pre-trade risk filter checks individual order parameters. The broker cannot assess the AI model generating those orders.

Minimum AI governance provisions for DMA agreements (no SEBI circular currently requires these):

Model disclosure: DMA agreement should require the client to disclose AI model type, trading strategy category, and expected order frequency range — giving the broker a baseline for anomaly detection
Change notification: client must notify broker before material model changes — new training, new strategy parameters, new instrument universe — so the broker can reassess risk filter adequacy
Kill-switch coordination: when the broker exercises the kill switch, what is the protocol for client AI model restart — does the client need broker approval? SEBI's circulars are silent
Incident liability: when DMA client AI causes a market incident, the broker-client DMA agreement determines liability allocation — a governance decision that should be made in advance, not in post-incident litigation

Kill-Switch Restart: The One-Way Governance Instrument

SEBI's kill-switch obligation for algorithmic trading — the requirement to immediately halt all algo trading — is well-documented. The restart governance obligation is a complete regulatory void. Post kill-switch restart governance questions that SEBI has not addressed:

Who authorizes AI restart after a kill-switch event — is the broker's compliance officer sufficient, or does an exchange approval step apply?
If the kill-switch was triggered by AI-generated anomalous behavior, what remediation evidence is required before restart?
If market conditions have changed materially between the halt and restart (a circuit breaker period spanning a news event), is the AI still operating within its validated parameters?
For market-wide halt events (SEBI-directed circuit breakers), is there any coordination framework for AI restart sequencing across brokers to prevent correlated restart-time order surges?

SEBI Stockbroker AI Governance Checklist

☐ Clause 26 AI extension: annual system audit scope extended to include AI-specific review — model performance against market microstructure, routing bias assessment, validation currency check
☐ SOR AI model card: document SOR routing criteria, venue preference logic, training data date range, validation methodology — available for SEBI or exchange inspection
☐ Multi-venue SOR governance: cross-venue routing analysis conducted periodically (not only per-venue); systematic venue preference or adverse selection patterns documented and remediated
☐ DMA client AI provisions: DMA agreement includes model disclosure, change notification, kill-switch coordination, incident liability allocation
☐ DMA risk filter adequacy: risk filters reviewed when client AI model changes materially; filter parameters documented against strategy type
☐ Kill-switch restart governance: documented restart protocol — authorization required, remediation evidence for AI-triggered halts, market condition assessment, restart log
☐ AI performance monitoring: periodic AI performance review (not only annual Clause 26 audit) — SOR execution quality trending, RMS false-positive rate on margin calls, compliance AI accuracy on client mandate checks
☐ SEBI examination readiness: SOR AI documentation, DMA agreement provisions, Clause 26 AI extension scope — packaged for SEBI examination on request