When we launched this newsletter at the start of 2025, there was widespread scepticism that AI would ever meaningfully enter trading due to the inherent risks and regulatory hurdles. Less than a year later, that scepticism already feels outdated. AI is fast becoming an integral part of the investment and trading process.

Perhaps early resistance wasn’t born of fear but of framing. The industry’s first instinct was to look to shoehorn AI into fully automated equity CLOB execution – essentially trying to replicate, not reimagine, human trading.

The real question was never whether AI could trade, but where and how it should be involved – not replacing traders but instead powering idea generation, order creation, and cross-asset workflow orchestration. From replacing costly broker research (https://docs.bigdata.com/) or automating voice-driven workflows (https://www.vocset.net/), AI is quietly redesigning the investment process – and, by extension, the trading desk itself.

Just as early block-matching automation evolved into targeted invitations and portfolio watchlists, the underlying objective remains the same: to protect investment intent and minimise market impact. Integrating AI is simply the next logical step.

Yet, beneath that logic, the market’s structure is shifting. We are moving away from centralised exchanges toward bilateral, segmented liquidity pools – a continuously moving landscape that changes how we discover liquidity, govern intent, and retain control. Here are this week’s five key themes in AI and Trading:

1. Smarter & Faster Pre-Trade

While concerns grow over a potential AI-driven market bubble, a more enduring transformation is unfolding beneath the surface. Today, it is now increasingly possible to automate the entire pre-trade workflow – from screening and sizing to venue and strategy selection (http://bit.ly/4oieeLh). 

Retail investors now have access to pre-trade intelligence once reserved for institutional desks, thanks to open-source models and public AI tools. With retail activity accounting for roughly 20% of U.S. equity trading (SIFMA, 2024), the very definition of pre-trade research – and who has access to it – is being rewritten.

As unstructured data becomes queryable and back-testable, narratives gain quantitative grounding. Benchmarks like STAC-ML are standardizing machine learning performance, while platforms such as Bigdata.com enable retrieval-augmented research pipelines.

By turning narratives into signals – and signals into strategies, AI is propelling trading beyond short-term pricing toward adaptive, context-driven alpha.

2. Beyond Equities: Voice, Chat, and the New Front-Office Surface

We’ve previously written about the next frontier emerging in OTC markets such as illiquid fixed income and commodities ( AI in Secondary Markets Trading – Five Things Firms Need to Know in September) – but that evolution is now accelerating – “What is your enterprise chat data worth?” – LinkedIn post by Matthew Cheung).If AI can finally capture and interpret unstructured voice and chat data, it would close one of the market’s biggest historical blind spots and unlock a range of new asset classes for fully automatable trading. Powered by speech recognition, domain-tuned natural language processing, and API-driven communications platforms, conversations can now be diarised, timestamped, and structured; RFQs can be automated with pre-trade checks and risk calculations; and trades can be executed electronically with full audit trails and human-in-the-loop approvals. This development transforms AI from a static bot into an intelligent workflow engine – one that infers intent, consolidates pricing across chat rooms, triggers margin and sanctions checks, and surfaces cross-desk opportunities in real time. While fully autonomous agents remain limited by regulation, agent-to-agent handoffs in deterministic workflows are scaling quickly, deepening liquidity discovery and accelerating what can safely be automated.

3. Risks, Triggers, and the “Manchurian Model” Problem

Innovation is not without its challenges. Last week, we explored the growing difficulty of managing the explosion of trading data, revealing just how fragile fast, automated markets can become when systems are flooded with corrupted or low-quality information. Now, the Alan Turing Institute highlights a deeper threat: data poisoning – subtle “trigger phrases” implanted during model training that cause AI systems to behave unpredictably. Researchers describe these as “Manchurian behaviours.” (https://www.turing.ac.uk/blog/llms-may-be-more-vulnerable-data-poisoning-we-thought) 

Because adversaries can seed public data with poisoned content, these triggers might appear as rare strings or even common phrases that induce nonsense, language shifts, or erratic behaviour – potentially causing chaos and reputational harm rather than data theft. 

Unsupervised models can amplify these vulnerabilities, reinforcing bias and the illusion of precision without real performance gain.

The sustainable path forward lies in assistive intelligence – systems designed to support, not replace, trader judgment, and to know when to defer to human sense.

4. Surveilling the Machines: When Oversight Goes Autonomous

As AI becomes woven into trading infrastructure, human surveillance is falling behind.
This shift is driving a new era of machine-led monitoring. Eventus has introduced Frank AI (https://www.eventus.com/frank-ai-overview/) , a deterministic, audit-ready platform for trade surveillance and compliance analytics. It delivers repeatable, traceable outputs – without the hallucinations typical of generative models. Meanwhile, Aquis Exchange is using AI to detect market manipulation patterns across vast datasets in real time, at a scale no human team could match (https://www.aquis.eu/news/how-aquis-uses-ai-in-market-surveillance-from-data-to-detecting-manipulation).

But with scale comes complexity. As AI systems analyse, recommend, and execute decisions, they can create new asymmetries – enabling regulatory arbitrage, price discrimination, and systemic “wrong-way” risk.

Under stress, interconnected models could mirror each other’s behaviours, potentially compressing crises from days into minutes.

5. Regulatory Diversification: From Compliance to Continuous Supervision

As a result, regulators are adapting. The FCA’s latest feedback on AI implementation illustrates how oversight is evolving from static compliance to more continuous, data-led supervision. This covers comprehensive AI assurance frameworks that incorporate model drift detection, bias monitoring, feedback loops, and incident reporting to strengthen governance and accountability. It also emphasises the need for scenario and stress testing to assess how AI systems perform under extreme or crisis conditions, ensuring resilience and reliability in real-world deployment. In addition, cross-industry collaboration through the publication of anonymised case studies, the creation of multi-stakeholder working groups, and alignment with international standards such as those developed by NIST and Singapore’s AI frameworks are suggested to help foster global consistency and shared learning in responsible AI regulation.

As @JonDanielsson notes (https://modelsandrisk.org/blog/supervision-illusion-optimal-risk/)    the goal should not be to fine-tune risk but to prioritise resilience – building systems capable of absorbing shocks, not merely forecasting them.

This shift in oversight is crucial as markets move to 24/7 electronic trading and T+1 settlement, leaving little room for error. As AI becomes embedded in core market infrastructure rather than added on, operational resilience becomes the gatekeeper – ensuring always-on availability, data integrity, and auditable models. AI isn’t just reshaping markets: it’s rewiring them. The task now is to ensure that wiring remains secure, transparent, and guided by human judgment.

Thank you for reading. As always, I’d love to hear your thoughts – what you agree with, where you disagree, and what you’d like to see more of next time.

Best wishes,
Rebecca