AI isn’t just “enhancing trading.” It is changing how liquidity forms, how volatility propagates, and where operational risk sits. This week, three things were in focus: (1) new AI-enabled services hitting production workflows, (2) market repricing that looks increasingly “AI-synchronised”, and (3) supervisors upgrading both expectations and tooling. Here’s what I learnt this week on AI in Trading:

1. AI moves from Lifting all Boats to Sinking all Ships

Volatility is increasingly driven by AI-led repricing and conditional liquidity. As Reuters noted, AI-driven expectations can reprice business models quickly, and trading systems can express that repricing at speed and scale. (https://www.reuters.com/business/stock-market-ai-turns-lifting-all-boats-sinking-ships-2026-02-12/) A case in point was the recent sell-off in software companies on the back of news from Anthropic and then U.S. brokerages following an AI-enabled tax planning launch (https://www.reuters.com/markets/europe/european-shares-slip-ai-disruption-worries-mixed-earnings-weigh-2026-02-13/) – both of which illustrate live examples of how new AI services trigger fast cross-asset positioning shifts (equities, options, vol), which then feed into liquidity provision behaviour. (https://www.ft.com/content/1c63983f-5ff0-4b70-ad55-7e4334f59422).

Why this matters for secondary markets: under stress, liquidity doesn’t just “thin” – it becomes increasingly conditional and synchronised (threshold-driven withdrawal, correlated hedging, dynamic spread widening). The risk is less “bad models” and more feedback loops between signals, hedging, and liquidity controls.

2. Risk management is migrating from the order book to funding and operational resilience

The clearest signal that liquidity fragility is increasingly a balance-sheet and funding problem, and not just a venue problem. We wrote last week on the FSB’s repo report here (https://mindfulmarketsai.beehiiv.com/p/ai-in-trading-2026-liquidity-market-structure-and-adaptive-price-discovery) which highlights rising vulnerabilities in leverage, collateral reuse, concentration, procyclical margining. The subsequent move by the ECB’s move to broaden its euro backstop facility is similar: in stress, funding mechanics can force sales and impair intermediation so stabilising euro funding conditions becomes market-structure relevant. (https://www.reuters.com/business/finance/ecb-broaden-access-euro-backstop-board-member-says-2026-02-12).

Why this matters for secondary markets: as AI-driven execution and risk engines become more dominant, liquidity can evaporate non-linearly across products because funding and margin constraints bite faster than humans can intervene. Read more here on how this has impacted US Market Structure to date from @PaulRowady @Alphacution here https://alphacution.com/case-study-d-e-shaws-strategy-frame/

3. New services are moving AI from analytics into the execution stack

It’s not just execution but market infrastructure that is repositioning for AI-native workflows. LSEG’s planned on-chain settlement capability (Digital Securities Depository) reflects how exchanges are adapting infrastructure to support next-generation trading by redesigning post-trade rails for tokenised instruments and interoperable settlement 24/7, programmatic execution and AI-orchestrated workflows (https://www.lseg.com/en/media-centre/press-releases/2026/lseg-advances-next-generation-of-digital-markets-infrastructure-with-on-chain-settlement).

The UK’s digital gilt pilot selecting HSBC Orion shows sovereign debt issuance testing the same direction of travel: programmable instruments, new settlement and operational models that will ultimately pull execution and risk controls closer to the asset lifecycle. (https://www.hsbc.com/news-and-views/news/media-releases/2026/hsbc-orion-awarded-digit-platform-mandate).

Why this matters for secondary markets: when settlement becomes more programmable and interoperable, execution strategy design shifts from venue microstructure optimisation to end-to-end workflow optimisation – including collateral, margin, settlement, funding, and controls.  This is potentially where agentic systems could add significant leverage.

4. Regulators are responding by upgrading “how” they supervise, not just “what” they regulate

The regulatory direction is converging on three themes:

4.1 Supervisors are building their own AI capability (SupTech)
The FCA’s reported step-change in internal spend on Microsoft Copilot illustrates just how regulators intend to become more data-driven, faster, and more tool-enabled – which is likely to raise the bar for what they will consider “adequate” evidence, monitoring, and control at firms. (https://www.fnlondon.com/articles/city-regulator-eyes-2m-annual-bill-for-ai-chatbot-9c927334). ESMA’s 2026–2028 digital direction explicitly includes rolling out generative AI assistants and strengthening supervisory data capabilities. (https://www.esma.europa.eu/sites/default/files/2026-01/ESMA65-955014868-12887_ESMA_Digital_Strategy_2026_-_2028.pdf)

4.2 Operational resilience and third-party concentration are now core AI issues
The UK FCA announced a forward-looking review into how AI will reshape UK financial markets with findings to inform future supervisory guidance particularly around systems and providers being resilient and controllable under stress (https://www.fca.org.uk/news/press-releases/mills-review-consider-how-ai-will-reshape-retail-financial-services). The ECB Banking Supervision has also been explicit that it will continue monitoring AI (with focus on generative AI) and deepen assessment of third-party dependencies, building on DORA and operational resilience priorities. (https://www.bankingsupervision.europa.eu/press/speeches/date/2026/html/ssm.sp260203~672ce5d5ff.en.html) and now the U.S. Federal Trade Commission has intensified its scrutiny of Microsoft, sending civil investigative demands to several enterprise software and cloud competitors to gather information about the company’s licensing and business practices (https://www.spokesman.com/stories/2026/feb/13/ftc-ramps-up-scrutiny-of-microsoft-over-ai-cloud-p/).

4.3 System governance over model governance
Along with the recent MAS Consultation Paper which has been included in multiple recent newsletters, ESMA has already signalled that firms using AI must manage risks around governance, oversight, testing, and client outcomes – including technology-neutral rules, but increasingly technology-specific supervisory expectations as well. (https://www.esma.europa.eu/press-news/esma-news/esma-provides-guidance-firms-using-artificial-intelligence-investment-services)

5 FIX addressing architecture evidence

Industry groups are continuing to move proactively to address the need for common taxonomies, improve accountability and controls, as well as align with existing regulatory regimes rather than creating entirely new rules.

In AI-enabled markets, architecture itself is becoming evidence. The bottleneck is fragmented infrastructure, inconsistent semantics, and weak decision traceability. This requires a deliberate move from model-centric thinking to system-centric design. Agentic workflows only function safely when underlying data is clean, standardised, portable, and machine-interpretable (not just readable).

Authority semantics therefore need to be 100% clear. Firms need structured records of who (human or system) was authorised to decide, who executed, who could override or escalate, and under what delegation framework. Permissions will matter more than labels. Decision trees must be reconstructable, replayable, and explainable along a defined timeline. Data quality and traceability are not compliance overhead; they will be operational prerequisites.

But the risk with agent-based markets will be the complexity – path dependence, feedback loops, and orchestration. As autonomy increases, system behaviour risk increases and the control layer will be vital – delegation, authority boundaries, and replayability depend on standardised data and consistent decision traceability. Without that foundation, governance is not scaleable.

What to watch next

There will continue to be a flood of AI-enabled workflow products but the edge is moving  to firms that can integrate and scale safely. Volatility will increasingly be shaped by coordinated model behaviour and conditional liquidity withdrawal, especially around “AI surprise” events. Supervisory expectations will harden around auditability, kill-switches, third-party risk, and evidence quality – could your firm reconstruct who authorised the decision, what the system saw, what it recommended, what was executed, and what could have stopped it – on a single timeline?

And for those who think AI in taking over – there was an interesting result from a Harvard Study (https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies-it). The research showed that generative AI increased workload rather than reducing it. Employees work faster, take on more tasks, multitask more frequently, and extend work into breaks and evenings, leading to rising expectations and workload creep. While productivity may initially improve, the long-term risks include cognitive strain, burnout, and reduced decision quality. Taiwan went a step further and have passed a new AI Basic Law establishing a rights-based framework for AI governance and protections for workers displaced by AI. It emphasizes human oversight, non-discrimination, and clear responsibility across the AI lifecycle (https://www.bakermckenzie.com/en/insight/publications/2026/01/taiwan-ai-basic-act#:~:text=On%2023%20December%202025%2C%20Taiwan’s,as%20the%20central%20competent%20authority).

The recommendations from Harvard? Organisations should establish clear norms for AI use – including structured pauses, better task sequencing, and more human collaboration and interaction – to ensure productivity gains remain sustainable. You get the impression that AI in Trading is just getting started.

As always – thank you for reading – any comments/feedback please let me know – all welcome!

Rebecca