MiFID II Requires Best Execution Records. AI Inference Logs Are the Missing Piece.
AI Inference Logging and Best Execution Documentation in Investment Firms
---
Markets in Financial Instruments Directive II (MiFID II) requires investment firms to document every factor considered in every execution decision. For decades, that record was a trader's decision note — manual, imperfect, and present. Today, an AI system may have identified the market opportunity, recommended the venue, and timed the order. If that inference disappears after the trade, the best execution documentation is incomplete — and the next regulatory examination will find it.
When an AI system influences an execution decision, its reasoning is part of the MiFID II record. Logging the human's action while the AI's logic vanishes is not best execution documentation. It is half of one.
---
What the Documentation Actually Requires
Think of an AI inference log as the electronic equivalent of a trader's decision note — the contemporaneous record that explains why an order was executed in a particular way, on a particular venue, at a particular time. Traders have always been required to keep those notes. AI systems that influence execution decisions carry the same requirement, extended to their inputs, their pattern analysis, and their recommendations.
Best execution compliance under MiFID II is typically framed as a reporting challenge — the right formats, the right disclosures, the right venue comparisons under RTS 27 and RTS 28 (the regulatory technical standards that specify exactly how execution quality data must be reported and disclosed). For firms using AI execution tools, the actual challenge is documentation completeness. If an AI system identified the execution opportunity, selected the venue, and timed the order, those inputs are "relevant factors" under MiFID II Article 27 — and they require documentation whether or not a human trader also reviewed them.
The European Securities and Markets Authority (ESMA) updated its MiFID II implementation Q&A in 2024 to explicitly classify AI-generated analysis as a relevant factor requiring documentation in best execution records. The same review found that 34% of investment firms using AI-assisted execution tools had not updated their best execution policies to reference AI decision inputs — meaning more than one in three firms using AI in execution had an undocumented step in a regulatory-required decision record. The update arrived as a Q&A clarification — not a new rule, not a consultation paper. Many compliance teams never flagged it because it came in a routine update channel.
---
The Three-Component Gap
Complete best execution documentation for an AI-assisted trade has three components. Most firms document only one.
First: the order details — what was executed, when, at what price, on which venue. Almost always documented, often automatically, by order management systems.
Second: the venue selection rationale — why that venue, what alternatives were considered, how the decision met the firm's best execution policy. Documented by some firms for algorithmic trades, rarely for AI-assisted decisions.
Third: the AI decision logic — which patterns triggered the recommendation, what the AI system weighted in its analysis, and what the confidence or certainty parameters were at the time of the recommendation. For most investment firms, this component does not exist in the best execution record. The AI made a contribution to the decision that is invisible to the documentation system.
JP Morgan's LOXM trading AI, deployed to execute equity trades, operates without traditional trader notes — the AI executes decisions autonomously based on market pattern analysis. MiFID II compliance for systems like LOXM requires inference logs that document each decision: which patterns were identified, which alternatives were considered, what parameters guided the execution. Without automatic inference logging built into the AI system, that documentation must be reconstructed — which is expensive, time-consuming, and inadmissible under the contemporaneous documentation requirement.
---
Why This Is Accelerating Now
Three regulatory requirements converge on the same infrastructure need: MiFID II Article 27 best execution documentation, EU AI Act Article 12 automatic logging, and the Digital Operational Resilience Act (DORA) Article 28 service chain records. All three require a contemporaneous log of every AI action that influenced a financial decision, with the input data, the system identity, and the timestamp. Investment firms that addressed only MiFID II compliance for human traders have a documentation architecture that was never designed for AI.
The EU AI Act classifies AI systems that influence investment decisions as "high-risk" — a specific regulatory designation meaning they require stricter controls, automatic logging, and documentation than lower-risk applications. This classification does not depend on how the firm categorizes the tool internally. Article 12's automatic logging requirement takes full effect in August 2026 for financial services AI. Investment firms that have not implemented inference logging by then face simultaneous non-compliance across two regulatory frameworks.
ESMA's supervisory briefing on algorithmic trading, updated in 2023, already requires firms to be able to explain the logic of algorithmic systems to supervisors — including AI-assisted execution tools. A firm that cannot explain why its AI recommended a particular execution, on a particular venue, at a particular time is in an explanation gap that supervisors are now specifically looking for.
---
What National Supervisors Are Asking
EU-based investment firms face MiFID II enforcement through their national competent authority — the Autorité des marchés financiers (AMF) in France, or BaFin, Germany's Federal Financial Supervisory Authority. Each examines against ESMA's Q&A interpretations within their national framework. AMF examination guidance published in late 2024 explicitly includes AI decision documentation as an examination criterion for firms using AI-assisted execution tools.
Supervisory trajectory is clear: once ESMA-coordinated review identifies AI inference documentation as a market-wide gap — which is the direction of current supervisory focus — the requirement becomes a simultaneous priority for national competent authorities across the EU. At that point, every investment firm receives examination requests for AI documentation at the same time, and firms without existing logging infrastructure spend months reconstructing records that should have been generated automatically.
Examination requests for AI documentation are expensive to respond to after the fact. Generating the records automatically, at the time of each inference, costs nothing incremental once the infrastructure is in place.
---
What Complete Documentation Looks Like
Fully implemented, the best execution record for an AI-assisted trade contains four elements: the order details, the venue selection rationale, the AI inference log showing inputs and recommendation, and the human review record if a human was in the loop. All four are timestamped, sourced from systems under the firm's control, and available on demand.
Investment firms that have built this architecture can respond to an examination request — "provide best execution documentation for these 50 AI-assisted trades from Q3 2025" — in hours, not weeks. The records exist. They were generated automatically. They are retrievable from systems the firm controls.
Firms that have not built this architecture face a reconstruction exercise. Which AI tool generated each recommendation? What was the version running at the time? What inputs did it use? The answers exist somewhere in vendor logs, if the vendor retains them, in the format the vendor provides, accessible through the vendor's processes.
---
Building the Documentation Infrastructure
Within two weeks: audit every AI tool that influences execution decisions. For each, determine whether it currently generates a usable inference log — meaning a record that contains the input data, the recommendation, the confidence parameters, and a timestamp, in a format retrievable for examination purposes. One important scoping note: not every AI tool used by an investment firm triggers MiFID II documentation requirements. Only tools that influence order execution decisions qualify. An AI used purely for post-trade analysis, client reporting, or risk monitoring operates outside the best execution documentation scope.
Over the following 30 days: update the firm's best execution policy to explicitly reference AI decision inputs as documented factors under MiFID II Article 27. This update signals to examiners that the firm has considered the requirement — and creates the framework that the logging infrastructure must satisfy.
Sixty days out: implement or confirm automatic inference logging for all AI execution tools. One question to put to every AI execution vendor: "Does your MiFID II compliance certification cover inference logging of each execution decision, in a format my compliance team can produce to an examiner?" A tool can be designed by a compliance team and meet regulatory specifications for execution logic while generating no inference log at all. Compliance certification on the tool is not the same as compliance logging in your best execution records.
Final step: move financial AI to sovereign infrastructure where inference logs are generated automatically, stored in systems the firm controls, and available to compliance officers and examiners without vendor intermediation. When documentation is generated at inference time and stored on your own servers, the examination response is: here are the records, they are complete, and they were generated automatically for every decision.
---
The Firms That Answer in Hours
Many compliance heads have never thought to include an AI analysis section in the best execution policy — because the tools arrived through the technology team and nobody flagged that they changed the documentation requirements. That compliance officer faces a specific choice when the next examination preparation cycle begins. Retrofitting documentation infrastructure under examination pressure is a different project than building it as routine infrastructure maintenance.
If AI helped make the execution decision, the AI's reasoning is part of the MiFID II record. That requirement has been in place since ESMA's 2024 Q&A — most compliance programs just have not caught up to it yet. The firms that catch up before the examination cycle reaches their sector will answer documentation questions in hours. The rest will answer them in months.
Automatic inference logging is not a new capability requirement. It is a documentation requirement that an AI system either satisfies at inference time or does not satisfy at all.
---
Stralevo generates complete inference logs for every AI action on financial data — timestamped, source-cited, and stored on infrastructure you control. When a regulator or auditor requests best execution documentation, your compliance team produces it from your own systems, in the format your examination framework requires.