On 26 February 2026, ESMA published a supervisory briefing on algorithmic trading in the EU. It is nonbinding and addressed to national competent authorities, which makes it easy to misfile as supervisory housekeeping. That would be a mistake. Supervisory briefings are the instrument ESMA uses to converge NCA practice, and this one operationalises existing MiFID II expectations on governance, testing, and outsourcing while explicitly noting the growing use of AI in algorithmic trading. Its content will shape the examination of every desk running automated or AI-assisted execution.
The briefing creates no new law. What it changes is the standard of proof. The focus areas it names, pre-trade controls, governance arrangements, testing frameworks, and outsourcing, are precisely the ones where supervisory practice had diverged across member states, and convergence means the lenient interpretations disappear. Read as a whole, the briefing moves supervision from formal compliance toward demonstrable control effectiveness: the question is no longer whether a policy document exists, but whether the firm can show the controls working in live trading.
The perimeter: ML gives no exit
Under MiFID II Article 4(1)(39), algorithmic trading is trading where a computer algorithm automatically determines individual order parameters. The briefing reiterates the broad reading: if an algorithm determines any aspect of an order beyond mere routing or post-trade processing, the activity can be treated as algorithmic trading. Machine learning does not create a carve-out. If an ML model determines an order parameter or an element of the trading decision, the activity sits inside the perimeter, whatever the model architecture.
This matters because firms adding ML for sizing, timing, or venue selection often assume a “decision support” framing keeps them outside scope. It does not, unless a human materially re-decides before the order forms. The first governance task is therefore an inventory question: which models touch order formation, directly or through a chain of downstream logic. Firms are regularly surprised by their own answer, because the perimeter grew one convenience feature at a time.
What supervisors will ask to see
Summaries of the briefing indicate expectations around testing before deployment, retesting after material changes, real-time monitoring, hard and soft blocks, and responsibility for direct electronic access clients; the underlying ESMA document is the authoritative reference for the detail. The governance layer this implies is familiar from model risk management: an inventory of models and algorithms with named ownership, change control, testing and approval gates, incident management, and fallback procedures for when an algorithm must be pulled.
The shift is in the evidence layer. ESMA’s briefing is explicitly aimed at practical supervisory use, which means firms must be able to show an NCA how controls behave in production, not on paper. That looks like records of pre-trade blocks actually firing, retest artefacts tied to the specific change that triggered them, and monitoring alerts with the human responses attached. A control that works but leaves no trace will be treated, at examination time, much like a control that does not work.
Surveillance models need the same governance
There is no equivalent 2026 primary source for AI in market abuse surveillance, so this part is synthesis rather than citation, but the logic transfers directly. AI-driven surveillance is typically deployed to improve alert generation, pattern detection, and escalation quality. The governance obligations do not soften because the model sits in compliance rather than the front office: model risk, explainability, calibration, false positive and false negative behaviour, auditability, and human review all need owners.
The quiet failure mode is recall decay. A surveillance model whose detection rate drifts downward produces no alerts about its own decline; the queue simply gets calmer, and everyone enjoys it. Guard against this with periodic calibration reviews, challenger comparisons against the incumbent, and documented threshold changes, because “we tuned the model and abuse alerts fell” is a sentence that reads very differently in a supervisory file two years later.
Outsourced algorithms are still your algorithms
ESMA identifies outsourcing of algorithmic trading systems as a key supervisory topic, and this is where many firms have the least real control. Governance must cover vendor due diligence, contractual control rights, and oversight of external model components. The structural problem is that you cannot evidence controls you cannot see: if the vendor will not provide documentation access, change notification, and testing evidence, your governance framework has a hole exactly where the supervisor is now instructed to look. Those rights get negotiated at contract time or not at all.
Making the evidence cheap to produce
The economical response to the briefing is to make evidence a by-product of normal operation rather than a project run before each inspection. Tie the algorithm registry to change control, so that deploying a change without a registry entry is mechanically impossible. Store retest artefacts with the change record that triggered them. Keep monitoring alerts, interventions, and incident postmortems in a system that can reconstruct any trading day on demand.
The failure modes are predictable: a registry that drifts from production reality within a quarter, testing evidence that exists but cannot be tied to the version actually deployed, and monitoring thresholds tuned once at go-live and never revisited as strategies and markets changed. Each is invisible day to day and expensive at examination time. The firms that will find the 2026 supervisory cycle uneventful are the ones for which the briefing describes how they already operate; for everyone else, it is a gap analysis with a deadline attached.
FAQ
Is the ESMA briefing binding on firms?
Formally no. It is a supervisory convergence tool addressed to national competent authorities, and it creates no new legislation. In practice it functions as the script for inspections, because it tells every NCA what to look for and removes the divergent, more lenient national interpretations. Treating it as optional guidance is how firms end up explaining themselves in a supervisory letter.
We only use ML for pre-trade analytics that a trader reviews before acting. Are we in scope?
The honest answer is that it depends on what the trader actually does with the output. If the model's output feeds order parameters and the human step is a rubber stamp rather than a material re-decision, supervisors can view the activity as algorithmic trading under the MiFID II definition. Document the human decision point, evidence that it changes outcomes, and revisit the analysis whenever the workflow is streamlined, because streamlining is usually where the human step quietly disappears.
Where should responsibility for algorithm governance sit: front office, risk, or compliance?
Ownership of each algorithm belongs with the desk that profits from it, because that is where the knowledge to explain it lives. Independent validation and ongoing monitoring belong outside the front office, and compliance owns the evidence trail that supervisors will consume. What matters most is a single shared inventory, so that all three functions are talking about the same versioned artefact rather than three different spreadsheets.