It is hard to believe that it is not yet even a year since the #AIinTrading debate in New York – hosted by @SydneyBadman and @SamBeck of DBT with @BTON, @Imandra, @Sigma against a backdrop of genuine uncertainty. At the time, the industry was still asking a fundamental question: whether AI should even play any role at all in trading (read more here.) In the months since, that debate has shifted decisively. The focus has moved away from generative AI hype towards workflow architecture and accountability – defining where intelligence belongs in the trading stack, when determinism remains absolute and why infrastructure resilience and continuous oversight have become non-negotiable.

As AI has flattened traditional information asymmetries through tools such as @Bigdata.com and @Perplexity, internal governance, integration, and resilience are emerging as the true differentiators. Industry consensus is converging on automation that prioritises intelligence over speed, agentic rather than fully autonomous systems, and live oversight in place of static, rules-based compliance.

AI will reshape trading not by replacing humans, but by redistributing broader, deeper intelligence at machine speed within tightly controlled, auditable, and accountable architectures – making infrastructure itself a form of market structure. Here is what we have learnt over the past year on AI in Trading:

1. From “AI in Trading” to “AI as the Operating Model”

One of the most important shifts in 2025 was the move away from viewing AI as a set of predictive trading models to an end-to-end operating layer spanning research, pre-trade, execution, post-trade, surveillance, and governance. Early in the year, AI was still discussed largely as an incremental enhancement – AI-powered trading tools, AI-driven index strategies, and smarter analytics layered onto existing workflows (February 23, 2025).

By mid-year, particularly around the June AI Summit (https://bit.ly/3HKDjhL) the debate moved decisively toward workflow orchestration. Agentic systems were increasingly positioned to coordinate multi-step processes across regulated functions such as KYC, compliance, onboarding, and investment support, alongside growing recognition of the governance challenges this introduces: auditability, cascading failures, and opaque decision chains (June 8, 2025; June 16, 2025).

By year-end, consensus had converged on a clear architectural principle: intelligence belongs in idea generation, orchestration, and optimisation, while determinism must remain absolute in execution and risk. Strategy can adapt; control cannot.

2. 2025 introduced the start of the Agent Decade

2025 will be remembered as the year agents entered trading workflows, while the industry simultaneously drew a firm line against full autonomy. As Andrej Karpathy noted, this is “the decade of agents, not the year” (October 25, 2025) a view that resonated with operational experience.  While early hype cast agents as productivity multipliers for research, monitoring, and coordination, it underestimated the implications for accountability and the basic fundamentals of data quality, governance, security, and training. Agents as near-term replacements for human decision-makers are still a long way off (May 18, 2025).

This recalibration crystallised at the June AI Summit (https://bit.ly/3HKDjhL) where agentic AI was reframed from “assistive tools” into a systems-level orchestration layer for regulated workflows – immediately raising hard questions around auditability, escalation, and governance (June 8, 2025; June 16, 2025). It also exposed a second constraint: without industry protocols, interoperability standards, and continuous supervision, agents cannot be safely composed across the trade lifecycle, rendering static controls and one-off validation insufficient. While A2A (Agent to Agent) protocols continued to mature over the year (https://a2a-protocol.org/latest/) – they still do not account for industry interaction with existing FIX industry protocol (https://www.fixtrading.org/what-is-fix/) contributing to stalled initiatives, unclear ROI, weak data foundations, and insufficient controls (June 29, 2025).

By autumn, the debate narrowed to configuration discipline, reliability engineering, and the reality that moving impressive demos into resilient production systems is a multi-year – if not decade-long – effort. Parallel discussions on continuous learning reinforced a similar insight: improved learning does not automatically translate into safe behaviour in live markets (October 25, 2025; December 1st, 2025 and December 14, 2025). By year-end, a clear boundary had been established, consistent with broader AI adoption in Trading – agent intelligence may adapt at the strategy layer, but determinism remains absolute in execution and risk management – with human oversight and control paramount (December 21, 2025).

3. The Real AI Bottleneck: Data, Evaluation, and Determinism

Another clear lesson from 2025 was that AI success in trading is constrained far less by model capability than by data provenance, evaluation, and determinism. Without clean inputs, labelled ground truth, permissioned access, and auditable lineage, even the most advanced models fail in production. This was made explicit at @TradeTech2025 where @JamieOvenden, CTO at @Schroders described “data quality as non-negotiable” (https://bit.ly/45rYv56) with the need to involve governance and security from the outset as a mainstream adoption principle (May 18, 2025).

As the year progressed, attention shifted toward evaluation and verification, with growing emphasis on challenge datasets, shadow-mode testing, and trackable decision steps. By September, determinism was increasingly described as the plumbing for trust: “reproducibility” – ensuring the same input yields the same output – was recognised as essential for compliance, auditability, and testing, even if it does not guarantee success (September 14, 2025). By October, consensus moved decisively away from static model sign-off toward continuous assurance, with ongoing monitoring, traceability tags, replayable decision logs, and shadow deployment becoming the benchmark for production readiness (October 25, 2025).

FIX Working Group discussions reinforced this shift, highlighting the need for standardised metadata and audit tags capturing whether, when, and how AI influenced trading decisions (September 5, 2025; September 27, 2025) – more to follow on this in 2026.

4. From Models to Systems: LRMs, Infrastructure-Level AI, and Quantum

Technically, the debate evolved from early LLM hype towards more reasoning-based models and learning systems, before confronting a much harder problem for trading: continuous learning over long sequences and real-time optimisation, and why this approach collides with basic determinism requirements in regulated markets ( June 8, 2025; December 14, 2025). Particular concern emerged around AI moving “down the stack,” from decision support into infrastructure-level optimisation, where the risk is no longer poor prediction but machine-discovered strategies that are difficult to audit, hard to constrain, and capable of creating systemic feedback loops (December 14, 2025).

Quantum computing entered the discussion late in the year, but cautiously: not as a source of near-term trading edge, but as a potential efficiency layer for training, exploration, scenario analysis, and systems-level optimisation (September 27, 2025; December 21, 2025). These themes surfaced repeatedly in Paris-based discussions, including around the ACI World Congress and FIXFrance25. Increasingly quantum and infrastructure-level AI are being framed less as imminent disruptors and more as extensions of a systems-thinking approach to market resilience (June 16, 2025; September 27, 2025).

5. From Alpha to Market Structure in Always-On, Tokenised Markets

Another major shift was the recognition that AI is not just an alpha tool but now a catalyst for changes in market structure, particularly in the context of rising tokenisation and 24/7 trading. In always-on markets, human workflows will not be able to operate at the required pace or continuity, making AI a necessary execution and decision co-pilot rather than an optional enhancement.  Liquidity formation, venue selection and post-trade evaluation continue to be redefined as markets shift towards always on global interaction, increasing reliance on sophisticated pre-trade analytics, volatility profiling, dynamic routing and real-time controls across asset classes  (June 29, 2025; July 7, 2025; September 14, 2025). As markets become more globally integrated and infrastructure-dependent, sustainable advantage is increasingly seen as coming not from marginally better models, but from resilient, auditable intelligence infrastructure capable of operating deterministically and safely at scale (September 14, 2025), effectively making infrastructure itself a component of market structure (October 25, 2025; December 14, 2025).

6. AI Is Physical: Infrastructure, Compute, and Energy Constraints

Finally, 2025 reinforced a new key risk – AI in trading is inseparable from physical infrastructure. Performance and resilience are increasingly shaped by access to compute, cloud architecture, data centres, and energy grids, not just low-latency networks, and algorithms. As AI embedded itself deeper into trading and control workflows, attention shifted to hyperscaler concentration risk, operational resilience, third-party auditability, and the credibility of exit and failover strategies.

These concerns moved into the core industry narrative at the June AI Summit (https://bit.ly/3HKDjhL), where AI adoption intersected with energy constraints, sovereign compute, and geopolitical dependencies (June 8, 2025; June 16, 2025). By autumn, high-profile cloud disruptions reframed resilience itself as a competitive asset rather than a compliance requirement (October 25, 2025). By year-end, discussions around data-centre buildout, power availability, and regulatory response explicitly linked infrastructure dependency to market stability – underscoring that compute, energy, and governance are now integral components of market structure, not just AI (December 7, 2025; December 14, 2025).

7.  Regulation: From Principles to Operational Control

Regulatory thinking in 2025 seemed to shift from a debate over the use of AI in Trading to operational expectations around its use. Across jurisdictions, firms are now expected to identify and inventory AI systems, tier them by materiality, validate them continuously, and demonstrate end-to-end governance in live environments. In the EU, trading-related AI was increasingly framed as potentially high-risk, bringing explicit obligations around explainability, validation, auditability, and supervisory oversight (June 29, 2025; November 8, 2025).

The UK moved towards a more pragmatic “test-and-learn” approach, using sandboxes, live testing, and supervisory collaboration to assess AI in production rather than on paper (June 16, 2025; October 25, 2025; December 7, 2025).  In Singapore, MAS proposals went further by embedding AI risk within formal lifecycle governance, board-level accountability, and operational-resilience frameworks – explicitly treating AI risk as operational risk, not merely a model-risk issue (November 22, 2025; December 7, 2025).

The direction of travel is clear: oversight is moving away from static, rules-based compliance toward live supervision and continuous assurance, with resilience under stress and controllability in real-world conditions taking precedence over theoretical model precision.

Looking Ahead

Taken together, the lessons of 2025 point to a single conclusion: AI in trading is no longer about automation versus humans, but about the architectural separation of intelligence and control. As a result, markets will continue to become more integrated ecosystems of interacting agents, constraints, and feedback loops, requiring the necessary bounded autonomy, deterministic risk layers, continuous supervision, and resilient infrastructure as a result.

Thank you to everyone who supported the discussion this year – not least of which the FIX AI Working Group members – next week, I will share what we are looking ahead to in 2026.

As always – thank you for reading. Best wishes for the rest of a peaceful 2025 and a very happy New Year.
Rebecca