AI-enabled tools continue to roll out across institutional and retail markets. Global banks are deploying AI-enhanced execution and market-intelligence platforms, such as HSBC’s AI Markets tools in FX and macro analysis (https://www.hsbc.com/investors/results-and-announcements/investor-updates). OMS and EMS providers are embedding AI to support close-price prediction, strategy analysis, and workflow efficiency. AI-driven sentiment, macro, and cross-asset analytics are expanding, particularly in crypto and multi-asset markets. At the same time, data quality is emerging as a binding constraint on AI effectiveness, a theme highlighted by CME Group (https://www.cmegroup.com/education/articles-and-reports.html). Here’s what I learnt this week on AI in Trading:

1. AI is becoming market infrastructure, not a trading enhancement

AI in trading is moving from signal generation and task automation into the core operating fabric of markets. Persistent, agent-driven systems can now operate continuously across research, execution, monitoring, and system development, maintaining long-horizon context and closing feedback loops at low marginal cost – see the latest from Openclaw (https://openclaw.ai/blog/introducing-openclaw). This enables a shift away from vendor-locked SaaS tools toward open, self-hosted, model-agnostic orchestration layers where firms retain control over data, memory, tooling, and execution.

Unlike session-based assistants, these systems run continuously: monitoring news, filings, and prices in real time, writing and deploying code, coordinating agents, and acting without explicit prompts. While this lowers barriers to advanced automation and accelerates information processing and execution, it introduces new governance, auditability and security risks – with one example of the challenges here https://www.404media.co/exposed-moltbook-database-let-anyone-take-control-of-any-ai-agent-on-the-site/.

Competitive advantage is moving away from model performance toward system architecture – specifically, how tools, memory, evaluation, and execution are orchestrated. This shift materially alters the risk profile. When many autonomous or semi-autonomous agents operate simultaneously, market outcomes can resemble infrastructure failures rather than isolated trading errors. Liquidity gaps, unstable price formation, and feedback loops could increase interaction effects between adaptive systems operating across venues, rather than any single decision or strategy.

2. Regulation is shifting from models to live market outcomes

As a result supervisory attention is moving decisively away from what an AI model is and toward what an AI system does in live markets. As AI development cycles outpace traditional rulemaking, regulators are focusing on observable outcomes: execution quality, price formation, resilience under stress, cross-venue effects, and accountability when autonomous actions fail. In the UK, the FCA’s AI live-testing initiative signals continued engagement with industry  – read more here https://www.fca.org.uk/news/news-stories/applications-open-second-cohort-ai-live-testing.

This marks a transition from documentation-led supervision to more operational scrutiny. Regulators are increasingly assessing deployment context, escalation logic, monitoring, and the ability to reconstruct decisions after the fact. The emerging constraint is not principles, but capability. Effective oversight depends on skills, tools, funding, and operational readiness on both sides of the supervisory boundary – read more here https://www.linkedin.com/posts/iota-kaousar-nassr-05aa66b_supervision-ai-finance-activity-7422668969234464769-JbxO/?utm_source=share&utm_medium=member_ios&rcm=ACoAAARqE5IBeeTUq6qzU4Eqccq_UO6-OMeR6Ao.

_{3. Agent systems are reshaping market microstructure}

Learning agents do not simply trade within existing market structures; they reshape them. As agents adapt to one another, liquidity becomes increasingly system driven and conditional, rather than a function of static quoting rules. Correlation risk rises as data sources and optimisation targets converge, and feedback loops shorten market response times.

As alpha generation and execution logic policies embed decision-making directly into execution, best-execution analysis and market-abuse detection becomes more complicated. Quantitative trading firms and frontier AI labs are converging around the same technical stack: large-scale compute, vertically integrated data pipelines, tight constraints, and fast feedback loops – read more here  https://www.ft.com/content/18313a5f-ae6e-44e9-a26a-4a81cd3190bf. For secondary markets, this reinforces a structural divide. Large firms can deploy AI across the full trade lifecycle, while smaller participants increasingly access AI indirectly via brokers, OMS/EMS platforms, and analytics tools, with limited transparency into decision logic. Market asymmetries are now shaped less by venue access and more by access to learning systems and technical infrastructure. The competitive edge is shifting away from venue connectivity toward access to learning systems, specialised hardware, and disciplined deployment, specialised accelerators, and system-level interpretability; it is becoming a capability race which XTX appears to be winning – https://www.linkedin.com/posts/xtx-markets-is-again-the-largest-elp-systematic-share-7422228426255765504-eSkx?utm_source=share&utm_medium=member_ios&rcm=ACoAAARqE5IBeeTUq6qzU4Eqccq_UO6-OMeR6Ao

4. Governance shifts from documentation to operational decision control

Traditional AI governance – policies, model documentation, and pre-deployment validation – appears no longer sufficient for adaptive systems operating in real time. Governance is shifting from design-time assurance and paper completeness to runtime control and decision-level defensibility, with emphasis on continuous monitoring, enforceable constraints, explicit human authority, escalation paths, kill switches, and behaviour under stress. Firms may pass checklist-based reviews yet fail when asked to explain why a specific AI-assisted decision occurred in a fast market. Regulators are likely to increasing expect audit-grade evidence showing who made or approved each decision, why it was taken, which data and constraints applied, and how controls operated at the moment of action. This is an operating-model challenge: how decisions are constrained, observed, overridden, and reconstructed while systems are live. Static documentation offers little help, and authorities will instead require a provable chain linking outputs to controls, approvals, overrides, and accountable humans, a direction reflected in Singapore’s Model AI Governance Framework for Agentic AI – read more here https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf.

5. Control, ownership, and sovereignty will matter more than raw intelligence

As AI capability commoditises, control, ownership, and sovereignty become first-order concerns because they determine whether governance is enforceable. The binding constraint is whether firms have full control of the stack: auditability, revocability, interruptibility, update governance, and jurisdictional exposure. Where critical workflows sit on vendor- or hyperscaler-controlled infrastructure, outcomes become conditional on third parties and accountability weakens. Sovereignty is therefore not about where data resides, but about enforceable responsibility – whether an organisation can audit behaviour, govern changes, and halt execution when required. If it cannot, it does not control the system; it depends on it. Value shifts from bolting powerful tools onto legacy workflows to redesigning processes with explicit control and accountability. Clear authority and oversight hierarchies will be what differentiates firms in the future.

Thanks for reading – as always feedback welcome!
Rebecca