AI in trading is no longer just a niche experiment – it’s becoming part of the market’s operating fabric. This week’s developments show that shift in motion: Europe’s EU AI Act has moved beyond transparency to a Code of Practice on safety, security, and governance; the SEC has launched a dedicated AI Task Force; and market participants are pushing the boundaries of deployment – from human-in-the-loop agent systems to detecting reinforcement learning’s cartel-like behaviours, tackling orchestration risks, and advancing enterprise platforms to combine research with trading. Here’s what I learnt last week and why it matters for AI in Trading.

1. EU AI Act – From Transparency to Full Accountability

Last week, we looked at the EU AI Act’s August 2 training-data disclosure requirement under Article 53(1)(d), which made dataset provenance a core compliance obligation. This week marks the second milestone: the GPAI Code of Practice, expanding from transparency alone to a full framework for accountability and enforcement. For AI in trading, this means embedding transparency, safety, and security-by-design across the lifecycle—from model development to deployment—addressing risks like data/model poisoning and unlawful data use. While voluntary, the Code is expected to set the EU’s de facto AI security baseline, and breaches of the AI Act can carry fines of up to 7% of global turnover.

In the US, the SEC has launched an AI Task Force led by Chief AI Officer Valerie Szczepanik to accelerate responsible AI adoption across enforcement, market surveillance, and operational oversight. The team will coordinate initiatives across divisions, encourage cross-disciplinary collaboration, and deploy AI tools to enhance investor protection and market integrity.

Why it matters for trading: Compliance now goes beyond proving lawful data sourcing. Firms must show model integrity, build anti-manipulation safeguards, and strengthen governance and third-party risk management. Even fine-tuning a model can classify a firm as a provider, triggering new obligations. This raises AI oversight to the same standard as market, credit, and operational risk, requiring readiness for tighter regulator scrutiny in a fast-changing environment.

Note: The UK FCA’s AI Live Testing programme is accepting applications until 20 August 2025. It allows firms to trial fully developed AI systems – covering models, data, governance, and human oversight – in live market conditions with regulatory guidance.

2. AI Agents – Keeping Humans in the Loop

The paper Magentic-UI: Towards Human-in-the-loop Agentic Systems introduces an open-source interface and architecture for building AI agents with human oversight by design. Thanks to Stuart Winter-Tear for sharing, the system supports a flexible multi-agent setup (extendable via the Model Context Protocol) with access to tools like a browser, code execution, and file operations, while the UI provides explicit control points for people. It defines six core mechanisms—co-planning, co-tasking, multi-tasking, action guards, and long-term memory—making oversight routine and low-cost.

Why it matters for AI in trading: Trading workflows need automation with guardrails. Action guards can require approval for order routing or risk-sensitive changes; co-planning/co-tasking allow intervention during regime shifts or data anomalies; multi-tasking and memory support supervised, auditable execution across instruments; and MCP integration makes it easy to connect market-data APIs, backtesting tools, OMS/EMS platforms, and risk checks. This enables faster iteration and broader strategy coverage without losing control, auditability, or safety – key requirements for moving AI from research to live trading.

3. From Herding to Algorithmic Collusion

A new Wharton–HKUST study by Dou, Goldstein, and Ji finds that reinforcement-learning (RL) trading agents in simulated limit-order markets can spontaneously collude for supra-competitive profits without explicit agreements. First posted on SSRN in May 2023 and revised into NBER Working Paper No. 34054 in July 2025, the research identifies two collusion paths:

1. Artificial intelligence – agents learn contingent price-trigger strategies that reduce competition.

2. Artificial stupidity – under-exploration or over-pruning locks agents into low-aggression, mutually beneficial policies.

These behaviours persisted across market conditions and RL parameters, raising questions for market regulation and surveillance.

Why it matters for AI in trading: As RL-driven execution and market-making become more common, risks shift from simple herding to unintended collusion. This could widen spreads and impair price discovery, especially in thin or stressed markets. Firms may need AI to detect coordinated pricing patterns, embed anti-collusion logic into algos, test with adversarial “red-team” bots, and use market-level stress tests or circuit breakers. Expect regulators to increase focus on “algorithmic antitrust.”

4. From Code to Crisis – Orchestration Risks

AI-generated code is now being used to build execution algorithms, risk models, and data pipelines—sometimes deployed directly into live systems. Without robust validation, errors or malicious code can cause market-wide disruption: mispriced trades, leaked data, or outages. TurinTech notes a recent Amazon case where the issue was not a traditional bug but an AI orchestration failure, where harmful prompts passed review and reached production.

Why it matters for AI in trading: Speed-to-market pressures and agentic systems automating strategy updates, compliance tasks, or connectivity scripts increase the risk. Without guardrails, sandbox stress testing, and adversarial validation, one flawed AI output could lead to unauthorised trades, regulatory breaches, or manipulation vulnerabilities. In trading, AI is only as safe as its orchestration safeguards—without them, fragility gets embedded into core systems.

5. One Step Closer – RBC’s AI-Driven Workflow

RBC has integrated Aiden QuickTakes into its AI platform, cutting post-earnings note drafting from 45 minutes to just 15. This lets analysts and traders act on market-moving news far faster. Led by RBC’s new AI & Digital Innovation team and built on Databricks for secure, governed AI, Aiden now links research, pre-trade planning, execution, and automation in one workflow.

Why it matters for AI in trading: This is a move toward end-to-end AI in trading. By uniting research and execution under one AI framework, firms can respond to new data in minutes. QuickTakes keeps a human in the loop for accuracy, while Databricks ensures data security and compliance. This sets a model for regulated AI use and points to a future where AI continuously supports decisions across the investment lifecycle.

I’ll be taking a short break from the weekly All Things AI in Trading over August, returning in September – unless something truly market-shifting emerges, in which case I will share. As always, thank you for reading! Let me know what you found helpful, what you disagreed with and what you’d like to see more of.
Many thanks,
Rebecca