AI in trading continues to shift further from analysis toward execution. Recent developments – from Goldman Sachs and Yale University research on LLM limitations to Anthropic’s push into agentic systems, alongside Goldman’s own operating model evolution – all point to the same conclusion: AI is no longer just interpreting markets; it is starting to underpin how trading workflows are executed. However, we are still looking at potential rather than reality: for example, LLMs struggling with the core analytical precision required in trading. The more meaningful progress is occurring lower in the tech stack – across data architecture, workflow orchestration, and system integration. Agentic models that can operate across fragmented trading stacks could really begin to address one of the industry’s most persistent challenges: operational complexity. Here’s what I learned this week on AI in trading:

1. Iran to India: AI for investigating trades
The latest concerns around market manipulation from the 22 March oil market activity are likely to increase demand for greater AI capability in investigations. The evidence of large, well-timed trades placed minutes before a price-moving geopolitical signal from Donald Trump, which triggered a sharp sell-off has had commentators up in arms (https://www.reuters.com/business/energy/traders-bet-500-million-oil-price-just-before-trumps-post-delay-iran-attack-2026-03-24). The scale and timing – combined with the departure of SEC Commissioner Meg Ryan (YouTube) – echo broader concerns highlighted in cases such as the Jane Street enforcement action, where sophisticated, multi-leg strategies blurred the line between legitimate positioning and manipulation. A quantitative autopsy of the index-lifting and gamma-harvesting strategy, covering its mathematical foundations, execution lifecycle and P&L dynamics has now been published on GitHub, including enhancements to the modelling, legal interpretation, and simulations – read here https://github.com/Sandra-Cai/Jane-Street-India-Ban-Analysis

Why this matters for trading: Surveillance is shifting from addressing blatant market manipulation to studying more subtle real-time cross asset behavioural analysis. AI can now link futures, options, ETFs, and liquidity patterns into a single view of intent, while NLP tracks news and geopolitical signals alongside trading activity. This makes it possible to detect information asymmetry and coordinated behaviour earlier even when it is less obvious – but also raises the bar in terms of what is expected for market integrity and oversight.

2. Claude: the Desktop is now your API
Anthropic’s latest update (https://claude.com/blog/dispatch-and-computer-use ) shifts AI from tool-dependent to action-oriented. Instead of relying on APIs and integrations, AI can now operate directly across the desktop – apps, browser, workflows – effectively turning the interface into a universal API. Read more here – https://www.techradar.com/pro/put-claude-to-work-claude-can-now-use-your-computer-autonomously-you-just-have-to-tell-it-what-to-do

Why this matters for trading: The big bug bear for all trading firms today is the fragmented infrastructure and operational complexity. Many workflows – trade reconciliation, connectivity diagnostics, market data validation, and regulatory reporting – span legacy systems, vendor platforms, and partially integrated environments. By enabling AI to operate across these layers without requiring full standardisation, this capability reduces operational latency, manual intervention, and error rates. More strategically, it creates a programmable execution layer across the trading stack, where AI agents can coordinate data, systems, and workflows in real time – addressing one of the industry’s core challenges: stitching together disparate technologies into a coherent, efficient operating model.

3. Goldman Sachs & Yale: limits of LLMs
The Fin-RATE research from Goldman Sachs and Yale shows LLMs work for single-document extraction but fail at comparing companies and tracking changes over time. They claim they confuse entities, miss context, and fabricate trends. The key issue is not the model – it’s retrieval. How data is structured (by entity and time) matters more than the LLM itself. Read more here – https://www.linkedin.com/posts/mpfcohara_when-ai-meets-sec-filings-where-llms-deliver-share-7441831963264483328-8t-p/

Why this matters for Trading: this issue becomes far more consequential thinking of AI agents that can operate across desktop environments without relying on full system integration. While LLMs are not reliable as standalone analytical engines, they are increasingly effective as coordination and execution layers across fragmented infrastructure. This directly addresses a core industry constraint: workflows like reconciliation, market data validation, and regulatory reporting span disconnected systems and inconsistent data layers. Combined, these developments could point to a new operating model – deterministic systems provide clean, auditable data foundations, while LLM-driven agents orchestrate and execute workflows across the stack in real time. The result is a programmable execution layer that reduces operational latency, manual intervention, and error rates, and begins to turn fragmented trading environments into coherent, scalable systems.

4. Where Goldman IS choosing to use AI
Goldman Sachs’ 2025 shareholder letter outlines an AI-driven operating model under “One Goldman Sachs,” embedding AI across onboarding, KYC, regulatory reporting, lending, risk, and sales. It also highlights a structural shift in talent: nearly half the workforce is now in hubs like Warsaw, Kraków, Bengaluru, and Hyderabad rather than Wall Street or the City. The firm is still clear on the perceived risks – model errors, bias, data leakage, cyber threats, and regulatory uncertainty especially in areas requiring full auditability. Read more here – https://www.businessinsider.com/goldman-sachs-lays-out-ai-ambitions-biggest-risks-shareholder-letter-2026-3

Why this matters for trading: Trading is becoming more automated and globally distributed. Competitive advantage shifts from location and headcount to data quality, system design, and controls. Firms that can combine strong data pipelines with greater AI-driven coordination across the tech stack now have a real advantage in the ability to scale more effectively and manage complexity better.

5. Where Mustafa thinks we are heading next
Mustafa Suleyman highlights the next milestone: Artificial Capable Intelligence (ACI) – AI that can achieve real-world outcomes autonomously – his example is the ability to turn $100k into $1M. He argues that unlike the Turing Test, which measures language – ACI measures execution, meaning progress is accelerating, with agents now handling longer, more complex tasks and operating continuously across systems. Read more here – https://www.linkedin.com/posts/mustafa-suleyman_the-next-big-ai-milestone-im-watching-for-share-7441878179461632000-mOOa/

Why this matters for trading: This again raises the possibility of fully autonomous trading workflows – signal generation, execution, risk, and optimisation running continuously. But without strong controls and high-quality data – autonomy only increases risk. The key challenge becomes governance: ensuring oversight, auditability, and market integrity as trading systems evolve into continuously operating, AI-driven machines. Its still all about humans overseeing the machines for now.

Again on this week’s panel on AI – https://regtechconference.co.uk/session/who-controls-the-agents/ – the question was again asked on when AI will replace trading. But just as with electronic trading and algos – trading systems and workflows are being transformed rather than replaced. The winners will not be those with the most advanced models, but those with the cleanest data, strongest control frameworks, and the ability to orchestrate workflows across increasingly complex environments. As AI shifts from insight to execution, trading firms are moving toward programmable, continuously operating systems. The opportunity is significant – but the need for oversight, governance, and market integrity remains and becomes even more critical as autonomy increases; the traders role is just shifting again.

Thanks again for reading – more to follow next week.

As always let me know your feedback/comments

Many thanks

Rebecca