WARRANTWhat Article 12 actually requires.
Article 12(1) binds providers of high-risk AI systems to automatic event logging over the lifetime of the system. The application date is 2 August 2026, subject to a provisional deferral to 2 December 2027 for Annex III standalone systems under the May 2026 Digital Omnibus. Non-compliance is reachable under Article 99(4) at up to EUR 15 million or 3 percent of global annual turnover. This is a dated, penalty-backed obligation, not a hypothetical. The line-by-line read is in Article 12, line by line.
What the sentence does not do is fix the shape of the record. It scopes the horizon to the lifetime of the system. It names the events as relevant under paragraph 2. It leaves open the unit of the record itself. A running log and a per-action record both satisfy the verb technically allow for the automatic recording on a plain reading. The text does not choose between them.
What paragraph 75 changed for agentic systems.
The draft Commission Guidelines on Article 6, published 19 May 2026, name agentic AI directly. Paragraph 75 treats a complex or agentic system as a whole. The unit of assessment is the system, not the individual step. Paragraph 76 then carves out a component only when it is genuinely separable and deployed independently.
The Guidelines illustrate both endpoints. A whole-system case sits in under paragraph 75 and the related paragraph 90. A genuinely-separable, independently-deployed component sits out under paragraph 76. What the draft does not do is work the example through to a component that runs after the high-risk decision to produce records of what the agent did. The classification mechanics are covered in the high-risk classification Guidelines, read in full.
Where the boundary is undrawn.
Put the two texts together and a gap opens in the middle. An autonomous agent in an Annex III deployment takes many consequential actions in one run. Article 12 says the system must log its events. Paragraph 75 says the agentic system is assessed as one unit. So the question stands: does Article 12 record-keeping mean one running log per system, or a discrete, attributable record for each autonomous action?
Set the two endpoints next to the live case and the gap is exact. The grid below is the boundary.
The shape of the answer is the one abstraction Warrant has placed on the public record: an agentic system emits a structured execution trace, and a separable component consumes it to produce Article 12 records after the high-risk decision. Where that component lands on the paragraph 75 / paragraph 76 line is what the draft leaves open.
Why this gets decided now, not later.
The classification boundary is the only boundary a provider can act on today. The draft Guidelines cover classification. The obligations under Articles 11 to 15 and Annex IV are left to a future instrument. The harmonised standards that would settle system-internal logging against downstream record-keeping are unfinished.
With the obligations-side guidance pending and the standards both unfinished, teams shipping into the greater-than-one-year application window are making the system-versus-record-layer decision now. They are hard-coding it into production before any of that lands. Re-architecting the record layer after the standards fix the boundary is more expensive than choosing the boundary deliberately at the start.
What a deployer should do before it hardens.
A deployer that runs a high-risk agentic system carries the Article 26 obligations and signs its name to what leaves the building. The detail is in the Article 26 deployer obligations, line by line.
The structural choice is whether the record of what the agent did lives inside the system or in a separable record layer downstream of the decision. Paragraph 75 is what determines how that stack is allowed to be drawn. Deciding it deliberately now is cheaper than re-architecting once the standards fix the boundary.
The practical posture for a deployer building before the standards land: write down which side of the paragraph 75 / paragraph 76 line the record layer is intended to sit on, keep the agent's execution trace intact and addressable, and keep the record layer separable enough that a later boundary ruling does not force a rebuild of the agent itself.
The question, on the public record.
Warrant filed this exact question into the EU AI Office's targeted consultation on 2 June 2026. Contribution ID a39b0f7e-0e8b-48b7-ad3a-d4b50741b30f. The filing asks the Office to settle whether an after-the-fact separable component that consumes an agentic execution trace to produce Article 12 records is inside the classified system or genuinely separable.
The consultation closes 23 June 2026, 22:00 CET. The question is now a dated, public position, not a marketing claim. Whatever the Office decides, the boundary it draws will determine how every Annex III deployer is allowed to structure its records under Article 12.
Two terms, defined.
Questions a deployer asks first.
Read the source directly.
- Regulation (EU) 2024/1689 · EUR-Lex CELEX:32024R1689
- Article 12 record-keeping · annotated text
- Article 6 classification rules for high-risk AI systems
- Article 99 penalties
- Article 12, line by line · the obligation read in full
- The high-risk classification Guidelines · paragraph 75 in context
- Article 26 deployer obligations · who signs for the agent
Authored by Warrant Compliance, the regulatory-analysis function at Warrant. [email protected]. Editorial commentary on regulatory text. Not legal advice. The verbatim quotation of Article 12(1) reflects the official English-language text of Regulation (EU) 2024/1689 as published in the Official Journal of the European Union on 12 July 2024. Paragraph references are to the draft Commission Guidelines on the classification of high-risk AI systems under Article 6, published 19 May 2026.