And just like that it is already March – AI continues to push a redesign of how markets are supervised, controlled and stabilised under stress. Each day there seems to be yet another LinkedIn post on innovation in agents for trading workflows – the question is no longer whether AI can optimise execution but instead whether market infrastructure can currently govern AI systems operating at millisecond speed – particularly during increasing volatility. Here’s this week’s thoughts on what I learnt last week on AI in Trading:

1. AI Orchestration Is Now Inside the Regulatory Perimeter

ESMA has sharpened the perimeter of algorithmic trading under MiFID II, making clear that any automated system that determines an individual order parameter – including price, size, timing, initiation, execution strategy selection, slicing, or post-submission order management – constitutes algorithmic trading. A human “click to approve” does not remove the activity from scope where parameters are algorithmically set, and outsourcing trading technology does not shift regulatory responsibility. Only routing orders to venues without influencing trading parameters remains outside the definition. Where automated logic influences order behaviour beyond simple routing, firms fall within Article 17 and RTS 6 organisational requirements. Read more here – https://www.esma.europa.eu/document/supervisory-briefing-algorithmic-trading-eu

What this means for Trading: The industry’s move toward AI-driven orchestration layers – smart execution selectors, adaptive routers, liquidity-seeking engines, automated allocators – is no longer a question of adjacent analytics or decision support. Where these systems determine how orders behave in the market, they now fall under regulation. Firms may outsource infrastructure, models, or execution tools, but they cannot outsource accountability. ESMA has reinforced several supervisory expectations: pre-trade controls must be technically non-bypassable; parent and child orders must be cumulatively constrained to prevent runaway execution loops; repeated recalibration or adaptive model updates can amount to a material change requiring testing; and outsourcing arrangements must preserve full oversight, auditability, and kill-switch capability.

2. Volatility Is Increasingly Structural so Governance Must Operate at Market Speed

The UK are also taking note. In a speech this week, “Renaissance at Market Speed,” FCA CEO Nikhil Rathi noted how intraday volatility is becoming sharper, with execution speeds compressing – making it impossible for humans to sit inside every trading decision loop. While further automation is inevitable, Rathi emphasised that if markets operate at machine speed, controls must also operate at machine speed. The FCA recognises that automation is necessary but governance frameworks designed for slower, human-mediated trading environments are no longer sufficient for AI-driven markets. Read more here – https://www.fca.org.uk/news/speeches/renaissance-market-speed-uk-wholesale-finance-2026.

Why this matters for Trading: Similar to the ESMA guidance, governance can no longer be periodic and oversight can no longer be retrospective. Firms need to demonstrate:

• Real-time monitoring of live AI systems,
• Deterministic kill-switch capability,
• Escalation pathways that function during stress,
• Drift management in adaptive models.

But this is not just compliance evolution: it is infrastructure evolution. Supervision now needs to be engineered – not just documented.

3. Trading Surveillance Is Moving AI-Native

Banks including Deutsche Bank and Goldman Sachs are now deploying agentic AI models directly into trading surveillance workflows. Read more here – https://www.bloomberg.com/news/articles/2026-02-25/deutsche-bank-goldman-look-to-ai-to-flag-trader-misconduct. According to Bloomberg, both firms are exploring advanced AI systems that go beyond traditional rule-based monitoring to analyse orders, trades and trader behaviour in real time, flagging potential misconduct for human compliance review.

At Deutsche Bank, executives are working with Google Cloud to build large language models that can spot anomalies across order and execution data and monitor communications across internal channels, enabling the system to reduce false positives and surface meaningful signals more efficiently. Nomura is also reportedly exploring cross-bank collaboration on shared AI surveillance model training. Goldman Sachs is studying similar tools to analyse trading patterns for suspicious signals.

Why this matters for trading: Markets are moving into an era where AI not only executes but also monitors trades in real time across asset classes and venues. Surveillance is shifting from static, rule-based alerts to dynamic, model-based detection as liquidity formation and oversight begin to operate on the same computational layer. This evolution only increases the need for robust governance, real-time controls, explainability and model risk management at the participant level to uphold market integrity and meet regulatory expectations (back to point 1 of the newsletter this week).

4. Machines Can Mimic Most Fund Behaviour – But Not the Alpha

A new paper, Mimicking Finance claims that a neural network trained on historical data can predict 71% of mutual fund buy, sell, and hold decisions using observable inputs such as fund size, flows, and macro conditions. The key insight: the 29% of trades the model could not predict are the ones most linked to outperformance. The authors argue that managers tend to be less predictable when they have higher personal ownership and trade actively in top positions or in growth, high-R&D, and earnings-surprise stocks. Conversely they are more predictable when long-tenured, managing multiple funds or styles, operating in less competitive categories, and focused on stable, value-oriented companies – read more here https://www.nber.org/system/files/working_papers/w34849/w34849.pdf

Why this matters for trading: If a significant share of active management is systematic and machine-replicable, this makes flows and positioning more predictable and susceptible to geopolitical volatility shifts and crowding risk. Excess returns will increasingly concentrate in decisions that are harder to model shifting the use of AI more novel insights. For example, Norway’s $2.2 trillion sovereign wealth fund (NBIM) uses Anthropic’s Claude to screen more than 7,200 companies across 60 countries for ethical and reputational risks. Generating daily ESG risk reports flags issues such as forced labour, corruption, and fraud before they are fully reflected in market prices. This is particularly valuable in small and emerging-market firms with limited analyst coverage, enabling earlier exits and potential loss avoidance. As CEO Nicolai Tangen commented – sustainability and governance are inseparable from financial performance – read more here https://www.nbim.no/contentassets/99fa5525f1a947d6b7231101832bfb40/responsible-investment-2025.pdf.

5. Compute Concentration Is Becoming a Structural Market Variable

As AI becomes embedded across investment research, trade execution and market surveillance, financial markets are growing increasingly dependent on large-scale computing infrastructure. At the same time, regulators are drawing attention to the risks associated with heavy reliance on a small number of dominant cloud and AI providers. If liquidity provision, pricing engines and risk systems all depend on concentrated compute infrastructure, then infrastructure fragility can quickly translate into market fragility. Compute risk is evolving into systemic risk.

Why this matters for Trading: infrastructure can no longer be seen as a technical consideration. Vendor concentration exposure, latency variability across cloud regions, resilience and failover capacity under stress, and audit access to third-party AI systems have become core strategic questions. Decisions about compute architecture are no longer IT decisions – they are increasingly strategic market access decisions.

Markets are not becoming autonomous; they are becoming engineered – faster, more controlled and deeply compute-dependent. Infrastructure design increasingly determines resilience, execution quality and competitive positioning.

Technological shifts inevitably change roles, but they also create opportunity. The firms that will outperform over the long term are those that embed AI within a coherent firmwide strategy, driven from the top and supported by strong governance and data discipline. The ability to extract actionable insight from both structured and unstructured data at speed is becoming a defining advantage. In this environment, data quality, infrastructure resilience and leadership clarity are emerging as the quiet differentiators in modern trading.

As always thank you for reading – any comments/feedback welcome

Rebecca