SEBI's kill switch requirement for algo trading is well understood: every algorithmic trading system must have a mechanism that can halt all trading immediately. What is less understood is the governance obligation that arises when the decision to pull that switch is not made by a human but by an AI model.
The mechanism and the trigger are different governance problems. The mechanism — is it fast enough, reliable enough, tested enough — is a systems engineering question. The trigger — when should it fire, is the AI that makes that call validated, can it be explained, what happens when it misfires — is a model risk question. Most SEBI compliance programs have addressed the first. Almost none have addressed the second.
The Four Kill Switch Trigger Modes
Manual Trigger
Human operator pulls kill switch based on judgment. No AI model in the trigger path. Mechanism must be tested and available. Trigger decision is human — accountability is straightforward under Regulation 16C.
Rule-Based Trigger
Pre-programmed threshold fires kill switch: position limit breach, drawdown limit, order rate exceeded. Deterministic rule execution — not a model. Rules must be documented and approved; mechanism tested. Most SEBI algo traders operate here today.
AI-Assisted Trigger
AI model detects anomalies, predicts volatility, or classifies risk state — recommends pull/no-pull. Human confirms before the kill switch fires. Trigger model is in MRMF scope. Recall (misses genuine emergencies) and false-positive rate (disrupts legitimate trading) must both be measured and reported.
Autonomous AI Trigger
AI model fires kill switch without human confirmation. Highest-governance mode. Model makes consequential, time-sensitive, market-impacting decision autonomously. Requires full documentation, independent validation, stress-testing against historical crisis scenarios, human override capability, and a kill switch for the trigger AI itself.
Five Governance Requirements for AI-Driven Trigger Models
Model Documentation
What inputs does the trigger model consume — price, volume, order book depth, volatility surface, external market signals? What model architecture — threshold-based ML, time-series anomaly detection, reinforcement learning, rules augmented by ML? What is the decision boundary — at what output score/probability does the model recommend or execute a kill? Full model documentation must exist before production deployment. Under Regulation 16C, the broker owns this documentation obligation regardless of whether a third-party vendor built the model.
Independent Validation + Historical Backtesting
The trigger model must be validated against historical market stress scenarios. SEBI's August 2026 consultation paper named specific scenarios: March 2020 COVID crash, June 2022 rate shock, India-specific flash crash events. Validation must test both dimensions: recall (would the model have triggered in genuine emergencies?) and precision (would it have avoided false triggers in volatile-but-not-emergency conditions?). Independent validation means the team validating the model is separate from the team that built it — the MRMF standard applies.
Real-Time Performance Monitoring
Trigger models can degrade as market regime changes — a model trained in low-volatility environments may underperform during volatility regime shifts. Continuous monitoring must track: trigger frequency vs. baseline, false trigger rate (triggers followed by normal market recovery within N minutes), missed trigger rate (measurable market stress events where trigger did not fire), and model confidence score distribution drift. Monitoring outputs should feed the MRMF model risk reporting structure.
Explainability for Trigger Decisions
For Mode 4 (autonomous trigger), SEBI's August 2026 consultation paper called for post-hoc explainability: the broker must be able to explain, after the fact, why the AI triggered or did not trigger the kill switch in any given instance. For regulatory examination or market surveillance query purposes, "the model said so" is not a sufficient answer. Either native model explainability (SHAP values, attention weights, feature importance) or post-hoc explanation capability must be built and tested.
Kill Switch for the Trigger AI (Meta-Governance)
The most structurally novel requirement: a kill switch for the kill switch trigger AI itself. If the trigger model malfunctions — begins firing spuriously, becomes unresponsive, is compromised by adversarial market inputs designed to prevent emergency suspension — there must be a mechanism to disable the AI trigger and revert to manual or rule-based trigger mode. This "meta-kill-switch" must itself be tested, have a named responsible owner, and have a documented fallback procedure. MRMF's kill switch requirement for AI models applies to the trigger AI as a governed model, not just the trading algo it governs.
Regulation 16C (SEBI, February 2025) established that stock brokers are solely responsible for all algo trading activity on their systems, including activity by third-party algo providers. This sole accountability principle extends to the kill switch trigger AI when provided by a third-party risk management vendor.
If a risk management SaaS vendor provides an AI-based kill switch trigger module — marketed as "intelligent circuit breaker" or "AI-driven risk controls" — and that module generates a false trigger that halts a client's legitimate trading strategy, or fails to trigger during a genuine emergency, the broker faces SEBI examination on both the incident and the governance of the third-party AI component. The vendor's governance documentation is the broker's starting point — but the broker cannot offload the accountability.
Practical implication: brokers using third-party AI-based risk management must contractually require the same model documentation, validation evidence, and monitoring reports from their vendors that MRMF requires for internally-built models. IT Outsourcing Directions 2023 provide the framework for bank-affiliated brokers; SEBI's IT governance expectations provide the parallel framework for standalone brokers.
SEBI and MRMF Compliance Matrix for Kill Switch Trigger AI
| Governance Activity | SEBI Obligation | MRMF Obligation | Gap / Intersection |
|---|---|---|---|
| Trigger model documentation | SEBI Reg 16C: broker accountable for all algo components including trigger logic | MRMF Model documentation requirement for all AI models in production | BOTH One documentation package satisfies both — write it to MRMF standard, it covers SEBI accountability |
| Independent validation | SEBI Aug 2026 paper: validation against historical stress scenarios required | MRMF Independent validation for High/Critical models before production deployment | BOTH SEBI names the scenarios; MRMF provides the validation methodology. One combined validation exercise. |
| Explainability for trigger decisions | SEBI Aug 2026 paper: post-hoc explanation for trigger/no-trigger decisions | MRMF Explainability for high-stakes AI decisions | BOTH Same technical requirement from two frameworks. SHAP/attention/feature importance output satisfies both. |
| Human override capability | SEBI Kill switch must be available to humans — SEBI circulars require human control capability | MRMF Human oversight requirement: human must be able to override AI decision | BOTH Mode 4 (autonomous trigger) must have human countermand capability tested and documented. One technical implementation. |
| Kill switch for the trigger AI | SEBI Not explicitly required (yet) — Aug 2026 paper suggests forthcoming | MRMF Kill switch required for all governed AI models — applies to trigger AI as a model | GAP MRMF clearly requires it. SEBI has not yet formalized. Build to MRMF standard now — satisfies MRMF and positions ahead of SEBI formalization. |
| Performance monitoring | SEBI Not explicitly specified — SEBI surveillance team will examine if false triggers or missed triggers occur | MRMF Continuous monitoring for all production AI models: performance drift, data drift, output distribution | GAP MRMF monitoring built for trigger AI also prepares for SEBI surveillance queries. No formal SEBI monitoring standard yet — MRMF standard is the build target. |
| Sole accountability for third-party trigger AI | SEBI Reg 16C: broker accountable for third-party components | MRMF Third-party model governance: documentation, audit rights, RBI access (for bank-affiliated entities) | BOTH Vendor contract must require model documentation + validation evidence (MRMF) and model change notification (SEBI Reg 16C change management requirement). |
What to Do Now, Before SEBI Makes the Kill Switch Trigger AI Governance Requirements Explicit
SEBI's August 2026 consultation paper signals formal requirements within 12–18 months. Building the governance program now avoids retrofit under examination pressure and demonstrates proactive compliance — the same strategy that works with MRMF.
- Audit your trigger mode. Determine whether your kill switch trigger is manual (Mode 1), rule-based (Mode 2), AI-assisted (Mode 3), or autonomous AI (Mode 4). If you are in Mode 3 or 4, you have a trigger AI in production that currently has no formal model documentation, validation history, or monitoring program. Document the current state before SEBI asks.
- Build a model card for the trigger AI. Even if the trigger model was built internally with no formal documentation, reconstruct what can be documented now: inputs, architecture, decision boundary, training data vintage, last validation date, performance on known historical events (2020, 2022). This is the foundation for independent validation. For third-party trigger AI, request the vendor's model card and make it the starting point for your own documentation of the broker-facing components.
- Run a historical backtest against SEBI-named scenarios. Test the current trigger model (or reconstruct its behavior) against the March 2020 COVID crash, June 2022 rate shock, and India-specific flash crash events. Measure: would it have fired in genuine emergencies? Would it have fired (incorrectly) in volatile-but-not-emergency conditions? Document the results. This is the validation evidence SEBI's August 2026 paper anticipates requiring. If results are poor, that is information — it drives a model improvement program before an incident forces the question.
- Implement and test a meta-kill-switch (kill switch for the trigger AI). Define the procedure for disabling the AI trigger and reverting to manual or rule-based trigger mode. Test it. Assign a named owner. This is required under MRMF (kill switch for governed AI models), anticipated under SEBI's consultation paper direction, and operationally prudent regardless of regulatory requirement — a malfunctioning autonomous trigger AI is a category of market disruption risk that has no resolution path without a tested meta-kill-switch.