2026-08-17 - AIH 0401: Agentic AI Enters Union Law Through a Table of Codes
Article 1, point 43, of Regulation (EU) 2026/1744 adds Annex XIV to the AI Act. Its subject matter is administrative: a list of codes, categories and corresponding types of AI systems that delimits the scope of the designation of notified bodies under Article 30.
At the bottom of that list, in the final table, sits an entry worth isolating:
‘AIH 0401 — AI systems based on other emerging AI technologies not covered by other codes, including Agentic AI’
So far as I can establish, this is the first time the expression appears in a binding act of the European Union. I put it cautiously because I have not carried out an exhaustive survey of every binding act: earlier mentions exist in communications, proposals and working documents, but those are not binding instruments.
Its placement is technical and inconspicuous, and the temptation to overstate it should be resisted. It is not a legal definition, not a legal regime, not a classification criterion. In this contribution, I try to establish three distinct things: what the text says, what it reasonably permits us to infer, and what remains a hypothesis of our own.
How Annex XIV is built
The Annex is divided into four sections. Recital 43 sets out the logic: ensuring that bodies designated as notified bodies are fully competent regarding the AI systems they are required to assess, by way of a multi-dimensional typology.
Vertical codes. These identify the system by reference to the legislation governing it.
AIP codes refer back to the Union harmonization legislation listed in Section A of Annex I and run from AIP 0102 to AIP 0112, corresponding to points 2 to 12 of that section.
AIB codes are new and specific to the AI Act, identifying the biometric AI systems referred to in point 1 of Annex III: AIB 0201 for remote biometric identification, AIB 0202 for biometric categorization, AIB 0203 for emotion recognition.
Horizontal codes. AIH codes identify the underlying types and technologies and are applied in conjunction with the vertical codes. They are organized in four groups: AIH 0101 for symbolic AI, expert and knowledge-based systems and systems based on search and optimization; AIH 0201-0205 for machine learning, differentiated by the type of data processed and excluding generative and general-purpose AI; AIH 0301 for generative AI systems, including those based on general-purpose AI models; and AIH 0401 for emerging technologies.
Application for designation. Section 4 closes the loop: conformity assessment bodies use these lists when specifying the types of AI systems in the application for designation referred to in Article 29.
Two details that tell the story of the text
The first concerns numbering. AIP codes begin at 0102, not 0101. This is not a slip: Article 1, point 41(a), of the Regulation deletes point 1 of Section A of Annex I, which referred to Directive 2006/42/EC on machinery, and point 41(b) adds Regulation (EU) 2023/1230 as point 21 of Section B. The sequence of codes therefore mirrors an Annex I from which point 1 has been removed. Nothing in the text suggests that the discontinuity is destined to be repaired: it may well be permanent.
The second concerns the AIB codes, which cover only point 1 of Annex III. The other seven points have no codes, and the reason is not incompleteness but a consequence of the conformity assessment system.
Article 43(1) of the AI Act provides, for high-risk AI systems listed in point 1 of Annex III, the possibility of a procedure involving a notified body under Annex VII. Article 43(2), by contrast, provides that for systems referred to in points 2 to 8 of Annex III the provider is to follow the conformity assessment procedure based on internal control set out in Annex VI, which does not involve a notified body.
Codes are not needed where notified bodies are not needed. Annex XIV is therefore consistent with the current perimeter of third-party assessment, and it will grow if and when that perimeter changes.
The legal effects of Annex XIV
The effect is procedural and concerns conformity assessment bodies.
Article 30(2), as replaced by Article 1, point 16, provides that notifying authorities shall notify the Commission and the other Member States of each conformity assessment body ‘based on the list of codes, categories, and corresponding types of AI systems referred to in Annex XIV’, using the electronic notification tool developed and managed by the Commission.
One faulty logical step must be avoided here, and it is an easy one to take. AIH 0401 does not coincide with agentic AI. The code covers emerging technologies ’not covered by other codes’, and agentic AI appears there as an illustration, introduced by ‘including’. An agentic system built on a general-purpose AI model will normally fall under AIH 0301 and cannot be assigned automatically to AIH 0401; an agentic system based on other technologies may call for a combination of codes or a different assessment. The Annex does not, moreover, spell out the rules applicable to systems presenting more than one underlying technology, which is, in our view, a further reason to avoid absolute classifications.
The correct formulation is therefore this: a body intending to assess agentic systems falling within the residual category of emerging technologies will have to include AIH 0401 in its scope of designation and demonstrate competence corresponding to the requirements of Article 31.
That is little, but it is nothing. The first legal effect of the mention of agentic AI in Union law is a technical competence requirement on a third-party body, confined to systems that find no place elsewhere.
Three levels of reading
From here on, the ground becomes interpretive, and three planes are best kept apart.
First level: what the text says
The text says that a residual category of emerging technologies exists, that agentic AI is an instance of it, and that notified bodies must specify the codes in their application for designation. Nothing more.
In particular, the text does not use agenticity as a classification criterion. It names it inside a category defined in the negative, that is, by exclusion from the other codes. In our view, this is the most significant feature of the whole exercise, and it points the opposite way from what one might be tempted to conclude: the legislator recorded the existence of the phenomenon without giving it classificatory force.
Second level: what the text permits us to infer
One element deserves attention, with the limited force that attaches to it.
Code AIH 0205 identifies ‘AI systems that learn from their environment’. Its wording contains an express carve-out:
’excluding AI systems covered under AIH 0401’
The carve-out refers to all systems covered by AIH 0401, not specifically to agentic ones: the residual category also encompasses other emerging technologies. No position of the legislator on agentic AI can therefore be derived from it.
What remains is an observation about drafting. AIH 0401 is already defined in the negative: it covers what is not caught by other codes. A system classifiable under AIH 0205 would therefore be, by definition, outside AIH 0401, and the express carve-out is logically redundant. Redundancies in legislative texts are typically defensive: they are written when an overlap is feared.
The legitimate inference stops here: the drafter considered that something within AIH 0401 might be confused with learning from the environment, and wanted to exclude it. What that something is, the text does not say.
Third level: our hypothesis
What follows is a theoretical position I hold based on the technical literature and the work I am carrying out on agentic systems. It cannot be derived from Annex XIV, and I present it as a hypothesis, not as a reading of the text.
In our view, what makes agentic systems a distinct conformity assessment problem is not learning but operational delegation: the system’s capacity to select and chain actions on external tools in pursuit of a goal, with a causal chain between input and consequence that the operator does not determine in advance.
If this reading is correct, the drafting redundancy noted at the second level has an explanation. But the explanation remains a hypothesis, and others are possible: the redundancy may serve expository clarity, or may concern emerging technologies other than agentic ones.
One further point deserves clarification, because it is sometimes handled too loosely — including by me in the first draft of this text. Agenticity and generativity are not alternative categories. They describe different properties and routinely coexist in the same system: their placement in the table does not show that the legislator regards them as separate classes.
An unusually early date of application
One temporal feature distinguishes this part of the Regulation from most of the rest and deserves flagging.
Article 30 sits in Section 4 of Chapter III, which has applied since 2 August 2025 and was not touched by the postponement. The amendments made by Regulation (EU) 2026/1744 therefore operate from its entry into force on 27 July 2026.
Annex XIV is thus already operative, while the high-risk requirements that notified bodies will assess — Sections 1, 2 and 3 of Chapter III — are postponed to 2 December 2027 or 2 August 2028.
The
asymmetry I described when analyzing the new calendar
repeats itself here on a smaller scale: the administrative infrastructure precedes the substantive regime it is meant to serve. In the case of Annex XIV, however, the choice is defensible without reservation, because designating notified bodies is a lengthy process and must begin well before assessments fall due. The 28 January 2028 deadline set by Article 43(3), second subparagraph, for applications from bodies already notified under Section A of Annex I follows the same logic.
How the classification may evolve
The second subparagraph of Article 30(2) empowers the Commission to adopt delegated acts under Article 97 to amend Annex XIV in light of technical progress, advances in knowledge or new scientific evidence, by adding codes, categories or types of system, withdrawing them, or moving them from one category to another.
Article 97(2), as amended by point 37, specifies that this power is conferred for five years from 27 July 2026, with tacit extension unless opposed.
De lege ferenda, and with the uncertainty such an exercise entails: I consider it plausible that the first substantive recognition of agentic AI in Union law will come through a delegated act assigning it a code of its own, rather than through a new legislative act. The reason is that the power exists, is already conferred, and the residual category is expressly built to allow codes to be moved.
Other outcomes are possible, however, and should be stated: the Commission might retain AIH 0401 in its present form, break it into subcategories without extracting agentic AI, or intervene in the nomenclature in ways not foreseeable today. Nothing in the text mandates one route over another.
What can be said with confidence is that, whichever route is taken, an intervention on Annex XIV would oblige the Commission to formulate a recognition criterion. And a recognition criterion, once written, is interpretive material potentially relevant beyond its original administrative purpose.
What Annex XIV does not do
It should be said with equal clarity that on the substantive plane everything remains to be built.
There is no definition of an agentic system. There are no specific obligations for providers or deployers of such systems. There is no classification criterion bringing them, as such, within the high-risk regime: classification continues to run through Article 6, that is, through the safety component function or membership of one of the Annex III categories. An agentic system performing neither of those functions remains outside the high-risk regime, however broad the operational delegation entrusted to it.
This is the gap I have described in analyzing
operational delegation in agentic systems
. Annex XIV does not close it. It records it.
What to do now
Conformity assessment bodies. The application for designation under Article 29 must now specify the codes. Those intending to cover systems falling within the emerging technologies category must document competence for AIH 0401, bearing in mind that the code is residual and that its scope may be changed by delegated act.
Notifying authorities. Notification under Article 30(2) now proceeds based on the Annex XIV list and through the electronic tool managed by the Commission.
Providers of agentic systems. Annex XIV imposes no obligations on you. The first useful exercise is to determine which horizontal code describes your system: if it is built on a general-purpose AI model, the normal reference is AIH 0301 and not AIH 0401. Only for systems genuinely falling within the residual category could the availability of bodies competent for AIH 0401 become, where the conditions of Article 6 are met, a practical constraint on market access.
Those following the file. The channel to monitor is that of delegated acts under Article 30(2).
Closing observations
There is a discrepancy between the symbolic weight of this provision and its normative weight, and neither dimension should be overstated or understated.
Normatively, Annex XIV is what it declares itself to be: an instrument delimiting the scope of designation of notified bodies. Anyone deriving from it a legal regime for agentic AI would be reading into the text something that is not there.
Systemically, Union law has expressly acknowledged agentic AI as a phenomenon relevant to delimiting the competences of notified bodies, without clarifying its relationship to the other technological categories and without giving it classificatory force.
That is little, and it is less than a hasty reading might suggest. But it is the first textual foothold on which further legal developments may build.
Reference documents
- Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 (Digital Omnibus on AI), OJ L, 24.7.2026:
http://data.europa.eu/eli/reg/2026/1744/oj
- Regulation (EU) 2024/1689 (Artificial Intelligence Act):
http://data.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2023/1230 (machinery):
http://data.europa.eu/eli/reg/2023/1230/oj
This contribution is for information purposes only and does not constitute legal advice. Verbatim quotations are taken from the English version published in the Official Journal of 24 July 2026.
New to Gemini? This capsule is best read with a Gemini client, such as Lagrange.