AI Governance for Algorithmic Trading in Banking: SEBI April 2026, ESMA MiFID II, and Pre-Trade AI Controls

SEBI's retail algorithmic trading framework became fully binding on April 1, 2026, after a phased rollout that started with broker registration milestones in October 2025 — and it did something no other major market had done before: it put formal regulatory accountability for retail investors' AI trading algorithms onto the brokers who run them. ESMA's MiFID II supervisory work extends the same instinct to kill switches, pre-trade controls, and governance expectations for AI trading systems generally. The governance problem underneath both is specific to AI: a trading algorithm that learned manipulative-looking patterns from historical data, with nobody programming that behavior in, is still market manipulation in the eyes of the rule — and the firm running it is still liable.

Algorithmic trading governance was built for rule-based systems whose logic you could read in the code and whose behavior you could predict from the rules. AI trading systems — ML models generating signals from pattern recognition, NLP models trading on news sentiment, reinforcement learning agents optimizing strategy over time — need the same governance discipline but can't satisfy it the same way. You audit a rule-based algorithm by reading its code. You audit a trained ML trading model by testing its behavior across market conditions, probing it with adversarial scenarios, and watching its live behavior against what you expected — because the model itself won't tell you what it will do next.

SEBI's framework puts accountability for AI retail trading algorithms on the broker, not just the client who wrote or licensed the algorithm. ESMA's position requires governance arrangements to address AI specifically: non-deterministic behavior, pattern-based learning that can produce strategies nobody designed on purpose, and pre-trade controls that constrain the AI without quietly strangling legitimate trading performance in the process.

The market manipulation risk here is genuinely different in kind from human or rule-based trading risk. A human trader who layers orders intends to manipulate the market. An AI trading model exhibiting layering-like patterns may have learned them from historical data with zero manipulative intent anywhere in its design — but market abuse regulation evaluates the behavior in the market, not the intent behind it. That means the governance job is to prevent the pattern from emerging at all, whether someone designed it in or the model found it on its own.

Four AI Algorithmic Trading Governance Obligations

Development Governance

AI trading algorithms need a development governance framework calibrated to trading-specific risk, covering training data provenance (what periods and market conditions does it span, and were known manipulation episodes excluded or flagged?), architecture documentation, backtesting governance with overfitting controls like walk-forward validation and out-of-sample testing, simulation testing under realistic liquidity and market-impact conditions before anything goes live, and documented capacity limits on position size, order rate, and market impact. SEBI's April 2026 framework requires brokers to maintain documentation of every client algorithm running through their infrastructure — which functionally pushes a documentation obligation down to whoever actually built the algorithm, broker or not.

Pre-Trade Controls

Pre-trade controls have to catch an AI system generating orders at rates, sizes, or prices that exceed what's operationally safe — including in response to novel market conditions the training data never covered. Both MiFID II Article 48 and SEBI's framework require these controls to sit in the broker's infrastructure, outside the algorithm's own reach, so the AI cannot turn off its own risk controls under any circumstance. That means maximum order size limits the signal strength can't override, order-to-trade ratio limits, price-range bands against the current best bid/offer, gross position limits by security and sector, and volatility circuit breakers that force human review before trading resumes. A kill switch nobody has ever actually tested is not a compliant kill switch — it's a hope.

Behavior Surveillance

Post-deployment, AI trading algorithms need surveillance for manipulation patterns: order-pattern analysis distinguishing legitimate market-making from layering, spoofing, or quote-stuffing look-alikes; cross-market behavior for algorithms trading the same name across venues or derivatives; momentum-pattern analysis checking whether the algorithm's order flow looks designed to push price in a direction that benefits its own book; and news-event correlation checking whether trading around announcements is creating artificial price movement. This surveillance has to run at the speed and scale of the trading it's watching — manual human review of high-frequency AI trading data simply isn't operationally possible — and the surveillance system's own false-positive rate needs calibration, because genuine compliance signals buried in noise are as good as no signal at all.

Model Change Management

AI trading models change constantly — scheduled retraining, drift-triggered retraining, hyperparameter tuning, feature updates — and every change is a new model from a governance standpoint, even when backtested performance looks similar to its predecessor. Change management needs a documented rationale for the change, pre-deployment evidence from backtesting, simulation, and paper trading, sign-off from risk personnel who actually understand the strategy and its risk profile, parallel running where feasible, and a rollback path that can reinstate the previous version immediately if the new one misbehaves live. SEBI's framework requires broker audit trails to capture algorithm version changes — a formal versioning obligation for every AI trading model running through broker infrastructure.

A trading algorithm that learned manipulative-looking patterns from historical data didn't set out to manipulate anything. Market abuse regulation doesn't ask about intent. The firm answers for the AI's market behavior whether that behavior was designed in deliberately or discovered by the model on its own — which is precisely why pre-deployment behavior testing, live surveillance, and training data governance exist: they're the controls that keep this liability from ever materializing in the first place.

SEBI's Framework and India's Market Microstructure

SEBI's retail algo framework is the first time a major market has extended formal regulatory oversight to retail investors' AI trading algorithms. It requires brokers to register retail client algorithms through their API infrastructure, maintain complete audit trails with algorithm identifier and version for every algorithmically generated trade, apply the same pre-trade risk controls to retail algorithmic orders as to proprietary ones, and take accountability for client algorithms complying with SEBI's market conduct rules. Brokers are now the front-line AI governance entity for retail algorithmic trading in India — a job most of them did not have the infrastructure for a year ago.

NSE and BSE run dedicated market surveillance functions calibrated specifically to India's market microstructure — liquidity distribution, trading participant mix, and patterns particular to the Indian context — which means an AI algorithm that behaves fine in US or European markets can still trip India's surveillance simply because it was never tested against Indian conditions. Firms deploying AI trading systems here need to test against India's market microstructure directly, not assume a globally calibrated algorithm will clear India's surveillance bar by default.

SEBI's audit trail requirements go further than most jurisdictions: order logs timestamped to microsecond precision, algorithm identifiers linking every order to its source algorithm, risk-check logs showing which pre-trade controls applied and any breaches, and user activity logs for anyone who can override or pause the algorithm. For AI specifically, the audit trail has to capture model version and parameter state at the moment of each trade — so that if SEBI opens an investigation months later, it can pull the exact model that generated a specific trade and examine its actual behavior, not a later version that happens to share a name.

AI trading algorithms trained mainly on US or European data genuinely won't capture India's market dynamics: the F&O market structure, the NSE/BSE dual-listing dynamic, India's pre-open price discovery mechanism, exchange-specific circuit breaker mechanics, and how FPIs, domestic mutual funds, and retail investors behave around India-specific events like Budget announcements, RBI policy decisions, and earnings season. A model without India-specific calibration doesn't just underperform here — it can generate patterns that India's surveillance flags as manipulative even when the algorithm was genuinely designed to be compliant in its home market.

What to Do About It

• Test every kill switch end-to-end at least once before relying on it in production — an untested kill switch is a governance gap disguised as a control.
• Document capacity limits (position size, order rate, market impact) for each AI trading model explicitly, and confirm they're enforced in broker infrastructure the algorithm cannot override.
• Build a dedicated adversarial/drift testing program for AI trading models separate from standard backtesting — walk-forward validation alone won't catch manipulation-pattern drift.
• Calibrate AI trading algorithms against India-specific market microstructure data before deployment here, even for models already live and compliant elsewhere.
• Confirm your broker audit trail captures model version and parameter state at the moment of each trade, not just at deployment — SEBI's framework expects the former.

Zytra governs AI trading algorithms with development documentation, pre-trade control validation, behavior surveillance for manipulation patterns, and model change management — producing the audit trail SEBI's April 2026 framework and ESMA's MiFID II supervisory guidance require.

Talk to sales