India's Markets Get an AI Kill Switch: Inside SEBI's New Rulebook

The moment India stopped asking whether markets would use AI#

For years, the debate about artificial intelligence in Indian finance stayed comfortably theoretical. That phase is over. On 19 August 2026, at the Federation of Indian Chambers of Commerce and Industry's 23rd Capital Markets Conference in Mumbai, the chairman of the Securities and Exchange Board of India (SEBI), Tuhin Kanta Pandey, put the question plainly: "The question is not whether markets will use AI, the question is how we use it responsibly while preserving trust." He then confirmed that SEBI will "shortly" issue binding guidelines requiring human oversight, data controls and "kill-switch" mechanisms for AI and machine-learning systems used across India's securities markets.

That sentence changes the tone. India is no longer merely encouraging AI adoption; it is preparing to make firms answerable for it. For anyone who trades, builds models, runs a fund or writes financial software in India, the coming rulebook is the most consequential AI-in-finance development of the moment.

What happened#

Speaking on the sidelines of the FICCI conference, Pandey said SEBI's proposed framework for the responsible use of AI and ML would follow a tiered approach with clearly defined accountability and governance controls. The core principle is blunt: every SEBI-regulated entity remains fully responsible for any AI or ML tool it uses, whether the tool is built in-house or bought from a third-party vendor. That responsibility extends to the privacy, security and integrity of investor data as well as to the outputs the system produces.

The safeguards named so far are concrete rather than aspirational. Regulated firms will be expected to maintain a "kill switch" (a mechanism to halt an AI system on demand), alongside "humans in the loop" and controls over the data that feeds the models. Pandey framed AI as a genuine asset for market surveillance, risk assessment, fraud detection and investor servicing, while flagging its risks around opacity, bias, cybersecurity, data protection and accountability.

Importantly, this is not a standing start. SEBI floated a consultation paper on guidelines for responsible usage of AI/ML in Indian securities markets in June 2025. Pandey acknowledged that the earlier consultation had not been concluded, but said developments since then (the arrival of far more capable generative and "agentic" systems) prompted the regulator to fold newer issues into the framework before finalising it.

Why a "kill switch" and a "human in the loop" matter#

To see why these safeguards are the crux of the matter, it helps to unpack the risks specific to AI in markets.

The first is the black box problem, the lack of explainability. Modern machine-learning models, especially deep neural networks, can produce accurate predictions without offering a legible reason for them. The Reserve Bank of India's deputy governor, T. Rabi Sankar, described this directly in an October 2025 address: algorithmic opacity makes it "hard for regulators and auditors to understand how decisions are made, which, in turn, undermines accountability." When a customer is denied credit or a trade is flagged, the law generally requires a reason. A model that cannot explain itself makes that hard.

The second risk is herding. If many trading desks lease similar AI models trained on similar data, they may react to the same signal at the same instant. Rabi Sankar warned that this can amplify volatility and that "AI misjudgements can trigger market dislocations." A "kill switch" is the market's circuit-breaker for that scenario: a pre-defined way to pull a misbehaving system offline before a feedback loop becomes a flash crash.

The third is accountability drift. As Rabi Sankar put it, it can "become difficult to assign responsibility when an AI makes a harmful or erroneous decision." SEBI's answer is to remove the ambiguity in advance: the licensed entity owns the outcome, full stop. A "human in the loop", meaning a person with the authority and the information to override the machine, is the mechanism that keeps a named human accountable for automated decisions.

These are not abstract worries in a market this size. Pandey noted that equity issuances crossed ₹4.5 lakh crore in FY25-26, including roughly ₹1.9 lakh crore raised through 366 initial public offerings; corporate bond issuance topped ₹9 lakh crore in FY26; alternative investment fund commitments reached about ₹7 lakh crore by end-July 2026; and market capitalisation stood at around 132% of GDP. India now counts close to 14.9 crore unique investors and mutual-fund assets of roughly ₹86 lakh crore. AI errors in a market of that depth propagate quickly.

Market implications: who feels this first#

The nearest-term effects land on the parts of the market where AI is already doing real work.

Quantitative and algorithmic trading desks are the most obvious. A firm running execution or signal-generation models will need a demonstrable ability to halt them and a governance trail showing a human can intervene. That raises the compliance bar but rewards firms that already invested in model risk management. AI in the stock market is no longer a niche concern.

Robo-advisers and wealth-tech platforms that use ML to recommend portfolios must be able to explain and audit their outputs. The tiered design suggests that a platform steering millions of retail investors will face heavier obligations than a back-office tool with no client-facing impact.

Brokerages, exchanges and depositories, the market's plumbing, sit squarely inside the perimeter. SEBI already runs its own AI: Project Sudarsan and R(AI)DAR to spot suspicious financial promotions and misleading advertisements, plus a Cyber Suraksha Portal for market-wide cyber resilience. The regulator is, in effect, holding the industry to standards it is applying to itself.

Fintechs and third-party AI vendors face the sharpest commercial change. Because the buyer of an AI tool remains fully liable for it, procurement contracts, model documentation, data-lineage disclosures and audit rights become negotiating points. Vendors that can supply explainability and clean data provenance gain an edge; opaque "black box" offerings become harder to sell into regulated firms.

Beyond equities, the ripples reach fixed income and foreign exchange through the same surveillance and risk-model logic, and banking, where the RBI is moving in parallel. Governor Sanjay Malhotra in mid-August 2026 urged lenders to accelerate AI investment while warning of the attendant risks. Both of India's principal financial regulators are now pushing adoption and guardrails at once.

Technical deep dive: what a compliant AI stack looks like#

Stripped of jargon, SEBI is asking regulated firms to engineer three properties into any AI system that touches the market.

The first is explainability. Where a decision affects a customer or a market outcome, the firm should be able to reconstruct why the model produced it. In practice this means favouring interpretable models where possible, and, where complex models are used, layering on tools that approximate the reasons for a given output and logging them for audit.

The second is controllability, the kill switch. Technically, this is a governance and systems-design requirement: an AI-driven trading or decision engine needs a pre-authorised, tested pathway to be paused or reverted, ideally automatically when it breaches risk thresholds. Rabi Sankar's framing of "safety by design rather than safety as an afterthought" captures the expectation that these controls are built in from the start, not bolted on after an incident.

The third is data governance. Models inherit the biases of their training data. As Rabi Sankar noted, systems "trained on biased historical data are likely to perpetuate or amplify historical discrimination" in areas such as credit profiling. SEBI's data controls push firms towards documented data lineage, consent-compliant sourcing and testing for skew before deployment. Complementary techniques the RBI has flagged, such as stress-testing models under diverse scenarios and "red-teaming" to surface vulnerabilities, are the market equivalent of crash-testing a car before it is sold.

None of this is uniquely Indian; it mirrors global model-risk practice. What is distinctive is the pairing of hard accountability with an explicit "facilitate, don't obstruct" stance: Pandey stressed that SEBI would "certainly help in AI technology to be facilitated, but then we must have responsibility and accountability more clearly outlined."

Critical analysis: strengths, gaps and the enforcement question#

The framework's strengths are its clarity and its timing. By assigning liability to the regulated entity regardless of whether the AI is built or bought, SEBI removes the vendor-blame loophole before it can be exploited. A tiered, risk-calibrated design is sensible: it avoids smothering low-risk back-office automation with rules meant for client-facing systems.

A "kill switch" is only as good as the humans and thresholds behind it; switching off a trading model mid-session can itself crystallise losses or freeze liquidity, so the cure carries its own risk. "Explainability" remains contested. For the most capable models, current techniques offer approximations rather than true transparency, and a determined firm can produce plausible-looking rationales that satisfy a checklist without illuminating the model. Defining the tiers will be where the hard bargaining happens; draw the lines poorly and you either over-burden start-ups or under-regulate systemically important platforms.

There is also a competing-viewpoints tension worth naming. Rabi Sankar cautioned that if frameworks are "too rigid, they can dissuade experimentation, reducing AI to a tool deployed only by the largest players." Compliance costs fall hardest on smaller firms, which risks entrenching incumbents, the opposite of the inclusion AI is meant to deliver. The counter-risk, "unbridled adoption," is equally live given that 62% of Indian businesses reported a rise in fraud attempts in an August 2026 survey, many of them AI-enabled. SEBI is threading a needle, and the final text, not the speech, will determine whether it succeeds.

Evolution, not rupture#

This is best understood as the tightening of a long screw rather than a sudden turn. SEBI first asked market infrastructure institutions to disclose their AI systems back in 2019, when it directed exchanges, clearing corporations and depositories to report AI and ML tools, and separately asked mutual funds to submit AI-system details each quarter. Those were disclosure rules. The 2026 framework moves from tell us what you use to you are accountable for what it does.

It also dovetails with the RBI's work. In August 2025 the central bank published its Framework for Responsible and Ethical Enablement of AI (FREE-AI), a 26-recommendation report built around principles such as trust, fairness and "understandable by design," and proposed a public fund, reported at around $575 million, to build shared datasets, compute and a regulatory sandbox. The RBI's own AI tools, such as MuleHunter.ai for detecting mule accounts, are already deployed across roughly 20 commercial banks. Meanwhile industry has moved from pilots to production: an August 2026 ASSOCHAM-KPMG paper described AI evolving into a "core intelligence layer" across lending, underwriting and servicing, powered by India's digital public infrastructure.

Read together, this is a structural change in how India governs financial AI, a deliberate attempt to institutionalise accountability while the technology is still young, rather than a cyclical tweak or a one-off reaction to a scandal.

Key takeaways#

Accountability is the headline. Every SEBI-regulated firm will be fully liable for its AI tools, whether built in-house or bought, including data privacy and model outputs.
The safeguards are specific: kill switches, humans in the loop and data controls, applied through a tiered, risk-based framework rather than a blanket rule.
The perimeter is wide, touching algorithmic trading, robo-advisory, brokerages, exchanges, depositories and the vendors that sell to them.
It is a two-regulator pincer. SEBI's markets rules sit alongside the RBI's FREE-AI framework and its push for faster, safer AI adoption in banking.
The draft, not the speech, decides the outcome. How the tiers and explainability standards are defined will determine whether the rules build trust without freezing out smaller innovators.

Frequently asked questions#

What exactly is a "kill switch" in this context? It is a pre-authorised, tested mechanism to stop an AI system on demand. For example, it can halt an automated trading engine that breaches risk limits. Think of it as an emergency brake with a named person responsible for pulling it.

Does this ban AI in Indian markets? No. SEBI has been explicit that it wants to facilitate AI while defining responsibility clearly. The framework governs how AI is used, not whether.

Are the rules already in force? Not yet. As of the 19 August 2026 announcement, SEBI said it would issue the guidelines "shortly." A June 2025 consultation preceded this, and the regulator is updating it before finalising. Treat implementation timing as forward-looking.

Who is covered? SEBI-regulated entities (brokers, exchanges, clearing corporations, depositories, mutual funds, portfolio managers and advisers) and, by extension, the third-party AI vendors that supply them, since the buyer retains liability.

How is this different from the RBI's rules? SEBI governs the securities markets; the RBI governs banks and other lenders. They are moving in parallel: the RBI's FREE-AI framework covers banking, while SEBI's guidelines cover capital markets. A firm active in both will need to satisfy both.

Will smaller fintechs be able to comply? That is the open question. A tiered design should spare low-risk use cases, but compliance costs still fall hardest on smaller players, a tension regulators themselves have acknowledged.

What should firms do now? Inventory where AI touches regulated activity, document data sources, build and test halt mechanisms, and ensure a qualified human can review and override automated decisions. This is operational preparation, not investment advice.

Glossary#

Machine learning (ML): A branch of AI in which systems learn patterns from data rather than following hand-written rules.

Kill switch: A control that lets operators immediately stop an automated system, limiting damage from errors or runaway behaviour.

Human in the loop: A design in which a qualified person can review, approve or override an automated decision, keeping accountability with a human.

Black box problem: The difficulty of explaining why a complex model produced a given output, which complicates audit and accountability.

Herding: When many market participants act on the same signal at once. The risk grows when they run similar AI models, which can magnify volatility.

Tiered (risk-based) framework: Rules scaled to the risk of a use case, so high-impact, client-facing systems face stricter requirements than low-risk internal tools.

Model risk management: The discipline of identifying, testing and controlling the ways a model can fail, including stress-testing and independent validation.

Digital public infrastructure (DPI): Shared national digital rails, such as digital identity, real-time payments and consent-based data sharing, that lower the cost of building financial services.

References#

ANI, "SEBI to soon issue AI/ML guidelines for capital markets, mandate human oversight, kill-switch controls: Chairman Pandey," 19 August 2026.
Business Standard, "Sebi to overhaul SME framework, tighten accountability for AI, ML tools," 19 August 2026.
Securities and Exchange Board of India, "Consultation Paper on Guidelines for Responsible Usage of AI/ML in Indian Securities Markets," June 2025 (regulatory consultation).
T. Rabi Sankar, Deputy Governor, Reserve Bank of India, "Responsible Artificial Intelligence (AI): balancing innovation with financial stability," keynote at the Global Fintech Festival, hosted on BIS, 7 October 2025.
KPMG in India, "RBI's FREE-AI Committee report in the financial sector," September 2025.
Computer Weekly, "Reserve Bank of India proposes framework for AI adoption in India's finance sector."
Bloomberg, "India's Banking Regulator Urges Lenders to Accelerate AI Spend," 11 August 2026.
The Tribune, "AI evolving as core intelligence layer in India's fintech transformation: ASSOCHAM-KPMG report," August 2026.
ANI, "62% of Indian businesses report rise in fraud attacks: Report," 25 August 2026.
Business Standard, "Sebi issues directive to exchanges, clearing corps, depositories using AI tools," 2019.
Business Standard, "Mutual funds to submit details about artificial intelligence-based systems on quarterly basis: Sebi," 2019.

This article is for information only. It is not investment advice, a recommendation to trade, or a forecast of market outcomes. Statements attributed to regulators and officials reflect their public remarks; references to forthcoming guidelines are forward-looking and subject to change until formally issued.