EU AI Act in practice14 min read

EU AI Act compliance for a classification agent: what the high-risk rules require

Contents

EU AI Act compliance for a classification agent starts with one question: is it a high-risk AI system? Under Article 6 of Regulation (EU) 2024/1689, EUR-Lex, that depends on its intended purpose: a safety component of a regulated product (Annex I) or a use case listed in Annex III. Four questions settle what the Act asks of your organisation:

  1. Is it an AI system? Article 3(1) means a machine-based system, designed to operate with varying levels of autonomy, that infers from its input how to generate outputs such as predictions, content or recommendations.
  2. What is your organisation's role? Providers (Article 3(3)) and deployers (Article 3(4)) carry different obligations.
  3. Does Annex III name the purpose, and does an exception under Article 6(3) apply? This is assessed and documented for each Vorgang (one intake, one task, one destination) separately.
  4. What applies when? According to Article 113 as amended, EUR-Lex, the obligations for high-risk AI systems under Annex III apply from 2 December 2027.

Intended purpose, not technology, settles the classification

Article 6 has two routes to high-risk status. The first runs through Annex I: the AI system is a safety component of a product, or is itself a product, that needs a third-party conformity assessment under the Union harmonisation legislation listed there, such as a toy, a lift or a medical device (Article 6(1)). The second runs through Annex III: the system is intended for one of the use cases listed there (Article 6(2)).

For an agent that reads incoming post and assigns it to a department, the second route is the one to examine. Germany's Bundesnetzagentur, which runs the national AI Service Desk, puts the test plainly (in German): what counts is the intended purpose. Article 3(12) defines it as the use the provider intends, as described in the instructions for use, sales materials and technical documentation.

The same technique can therefore be high-risk in one Vorgang and not in another, so its purpose should be written down precisely enough for a third party to check.

Annex III lists use cases, not techniques

Annex III of the AI Act, EUR-Lex lists eight areas, each with specific purposes:

Area in Annex IIIExample of a listed purpose
1 Biometricsremote biometric identification, emotion recognition
2 Critical infrastructuresafety components in critical digital infrastructure, road traffic, or the supply of water, gas, heating or electricity
3 Education and vocational trainingadmission to institutions, evaluation of learning outcomes
4 Employment and workers' managementanalysing and filtering job applications, evaluating candidates
5 Essential private and public serviceseligibility for public assistance benefits, creditworthiness, evaluating and classifying emergency calls
6 Law enforcementevaluating the reliability of evidence
7 Migration, asylum and border controlassisting the examination of asylum and visa applications
8 Administration of justice and democratic processesassisting a judicial authority in researching and interpreting facts and the law

Two things follow for a classification agent from the text of Annex III. First, classification itself appears in it: emergency calls that are evaluated and classified (point 5(d)), and job applications that are analysed and filtered (point 4(a)). The technique that routes a delivery note to the warehouse is the same one that could sort out a job application. Second, what matters is what the assignment means for natural persons. An agent that distributes applications for public assistance benefits to case teams in a public authority works close to point 5(a). Whether it falls under it depends on what the assignment does to the assessment of eligibility, and that is assessed case by case.

Article 6(3) exempts systems that pose no significant risk

An Annex III system is not considered high-risk under Article 6(3), EUR-Lex where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making. That applies where any one of four conditions is met:

  1. The system is intended to perform a narrow procedural task.
  2. It is intended to improve the result of a previously completed human activity.
  3. It is intended to detect decision-making patterns or deviations from prior patterns, and not to replace or influence the previously completed human assessment without proper human review.
  4. It is intended to perform a preparatory task to an assessment relevant to an Annex III use case.

One rule overrides all four: an Annex III system that performs profiling of natural persons is always considered high-risk. Under Article 6(5) the Commission is to provide guidelines with practical examples of high-risk and non-high-risk use cases; the Bundesnetzagentur tracks their status in its AI Act timeline (in German).

Relying on the exception means documenting the assessment first

Under Article 6(4), EUR-Lex, a provider who considers that an Annex III system is not high-risk documents that assessment before placing the system on the market or putting it into service. The provider registers the system in the EU database under Article 49(2) and gives the documentation to national competent authorities on request. The AI Omnibus kept this registration and reduced the information it requires (Regulation (EU) 2026/1744, recital 22, EUR-Lex).

Such an assessment needs facts about the Vorgang. For a classification agent, they are recorded in these places:

Question for the assessment under Article 6(3)Where the fact is recorded in operation
What exactly does the system do?The agreed Vorgang: one intake, one task, one destination
What does it explicitly not do?The boundary: classify, extract metadata, route, create a task; never post, approve, pay, delete or trigger deadlines
Who works on the matter afterwards?The responsible Fachbereich (department), where the task arrives in the target system
What happens when the system is unsure?Below the confidence threshold the case goes to a named person
Are profiles of natural persons built?The classes from the week-1 process conversation and the extracted metadata
Can every assignment be traced?The decision log: every assignment with timestamp and reasoning

These records do not replace the assessment; they are its material. This article does not anticipate the result for any Vorgang.

A deployer is whoever uses the system under its own authority

When Agentivo says we operate an AI employee, we mean that we monitor it, correct it and take responsibility for the result. The AI Act uses its own terms. Under Article 3, EUR-Lex, "operator" is an umbrella term for provider, product manufacturer, deployer, authorised representative, importer and distributor (point 8). The deployer is a person, authority or body using an AI system under its authority (point 4). The provider develops an AI system, or has it developed, and places it on the market or puts it into service under its own name or trademark (point 3).

The Bundesnetzagentur explains (in German) that a legal person remains the deployer when it brings contractors into the operation of a system, as long as they act in its name and under its responsibility and control. Conversely, under Article 25 a deployer can become a provider, for instance by changing a system's intended purpose so that it becomes high-risk.

Which roles your organisation and Agentivo hold for your Vorgang follows from these definitions and from the contract, and the obligations below depend on it. The full set of EU AI Act obligations for companies as deployers has its own article.

If an agent is high-risk, Article 26 asks the deployer for oversight, monitoring and logs

Article 26 of the AI Act, EUR-Lex sets out what deployers of high-risk AI systems do. For a classification agent, these paragraphs are closest to daily work:

Deployer obligationParagraphWhat operation provides
Use in line with the instructions for use, with technical and organisational measures26(1)The agreed boundary of the Vorgang and the handover document
Human oversight by people with the necessary competence, training and authority26(2)Named people to whom the agent hands unsure cases
Relevant and sufficiently representative input data, where the deployer controls it26(4)Classes and examples from the process conversation
Monitoring operation, informing provider and authority of a risk and suspending use26(5)We monitor and correct; the intervention path stops the agent
Keeping automatically generated logs for at least six months, where they are under the deployer's control26(6)The decision log, with timestamp and reasoning
Informing workers' representatives and affected workers beforehand26(7)Your organisation's task
Informing natural persons subject to decisions the system makes or assists with26(11)Your organisation's task
Cooperating with competent authorities26(12)Log and records on request

The right-hand column describes what exists in operation. It does not say that an obligation is met; your review assesses that. For Annex III systems, Article 26 applies from 2 December 2027 (Article 113(c), EUR-Lex).

Public bodies assess the impact on fundamental rights first

Deployers that are bodies governed by public law, or private entities providing public services, carry out a fundamental rights impact assessment before deploying an Annex III high-risk AI system; so do deployers of certain Annex III point 5 systems (creditworthiness, life and health insurance). Systems for critical infrastructure under Annex III point 2 are excluded (Article 27, EUR-Lex).

The assessment covers, among other things, how and how often the system is used, who may be affected, the specific risks, human oversight and the steps to take if risks materialise. Since the AI Omnibus it can refer to the relevant sections of a data protection impact assessment (Article 27(4), consolidated text, EUR-Lex).

Public authorities that use an Annex III high-risk system (other than point 2) also register themselves in the EU database beforehand, select the system there and register its use (Article 49(3)).

Articles 12 and 14 ask for traceability and a way to intervene at any time

Both requirements are addressed to the provider who builds a high-risk system, and both shape what the deployer works with.

Article 12, EUR-Lex requires a high-risk AI system to record events automatically over its lifetime. The logs make its functioning traceable to a degree appropriate to its purpose: to identify risks, to support post-market monitoring and to let the deployer monitor operation under Article 26(5).

Article 14(4) requires that the people assigned to oversight can understand the system's capacities and limitations, remain aware of automation bias, interpret its output correctly, disregard, override or reverse that output, and interrupt the system through a stop button or a similar procedure.

In the operation of a classification agent, these are three mechanisms:

  • The decision log is the AI audit trail of the agent: it records every assignment with timestamp and reasoning.
  • The confidence threshold is the certainty below which the agent stops and hands the case to a named person.
  • The human intervention path is documented and can be triggered at any time: a person can stop the agent, correct an assignment or take over a case.

Whether these mechanisms meet the requirements in a given case is part of the review, not of this article.

The AI Omnibus moved the high-risk dates to December 2027

Regulation (EU) 2026/1744 of 8 July 2026, EUR-Lex, the AI Omnibus, amended Article 113. It was published in the Official Journal on 24 July 2026 and, according to the European Commission, entered into force on 27 July 2026. Originally, the high-risk obligations for Annex III were to apply from 2 August 2026. According to the consolidated text, EUR-Lex, the dates are now:

FromWhat appliesBasis
1 August 2024The Act is in forceArticle 113, first paragraph; recital 1 of Regulation (EU) 2026/1744
2 February 2025Chapters I and II: AI literacy (Article 4) and prohibited practices (Article 5)Article 113(a)
2 August 2026General date of application, including the transparency obligations in Article 50Article 113, second paragraph
2 December 2026The prohibitions added by the AI OmnibusArticle 113(a)
2 December 2027Chapter III, Sections 1 to 3, for Annex III high-risk systems, including Articles 6, 26 and 27Article 113(c)(i)
2 August 2028The same sections for Annex I high-risk systemsArticle 113(c)(ii)
2 August 2030High-risk systems for public authorities that were already on the marketArticle 111(2)

High-risk systems placed on the market or put into service before Chapter III applies are covered under Article 111(2) only if their design changes significantly afterwards. Providers and deployers of high-risk systems intended for use by public authorities comply in any case by 2 August 2030 (Article 111, consolidated text, EUR-Lex). If you work from material written before July 2026, check its dates against the consolidated text.

Articles 4 and 50 apply without a high-risk classification

AI literacy under Article 4 as amended, EUR-Lex concerns every provider and deployer, not only those of high-risk systems. Since the AI Omnibus, it asks for measures that support the AI literacy of staff and others who operate and use AI systems on their behalf, without requiring a specific level for any individual. For an operated agent, that means the people who handle escalated cases, read the decision log and can trigger the intervention path; see EU AI Act Article 4.

According to the European Commission, the transparency rules have applied since August 2026. Article 50 covers certain systems, such as those intended to interact directly with people. Whether it concerns a classification agent is a separate check, for example where the agent sends receipt confirmations to people.

A Pilotbetrieb produces the records a classification rests on

A classification needs facts, and those come from operation. The Pilotbetrieb, our entry engagement, takes seven weeks, four of them productive operation on your real documents:

  1. Week 1: the process conversation with your Fachbereich defines the Vorgang, the classes and the access.
  2. Week 3: acceptance against the hit rate agreed beforehand, go-live and handover.
  3. Weeks 4 to 7: the agent runs on real volume, every assignment is in the decision log, and we monitor and correct.
  4. End of week 7: the written evaluation of hit rate achieved, cases processed and time saved.

Every operation includes:

  • Operation entirely within the EU, responsibility from Düsseldorf.
  • GDPR-compliant processing with an AVV (data processing agreement) included.
  • A complete decision log available as evidence for EU AI Act audits.
  • A defined human intervention path, documented and triggerable at any time.

Agentivo is not a law firm. This article explains the criteria of the EU AI Act, Regulation (EU) 2024/1689; whether your Vorgang is high-risk and who holds which role is for you to assess with your legal advisers. Personal data remain subject to the GDPR, which the AI Act leaves unaffected under Article 2(7); see GDPR compliant AI. Who is answerable for a running agent day to day is the subject of AI agent governance.

Frequently asked questions

Is an AI agent that classifies incoming post a high-risk AI system?

There is no general answer. What counts is the purpose: if Annex III names it, the system is high-risk unless an exception under Article 6(3) applies. Profiling of natural persons is always high-risk. The assessment is documented for each Vorgang separately; this article does not anticipate it, and Agentivo is not a law firm.

When do the obligations for high-risk AI systems apply?

Under Article 113 as amended by the AI Omnibus, from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems. AI literacy (Article 4) and the prohibitions (Article 5) have applied since 2 February 2025, the transparency obligations (Article 50) since 2 August 2026. The source is the consolidated text.

Who is the provider and who is the deployer of an operated agent?

That follows from Article 3(3) and (4) and from the contract. The deployer uses the system under its own authority; the provider develops it and places it on the market or puts it into service under its own name. According to the Bundesnetzagentur, an organisation remains the deployer when contractors work in its name and under its control.

What should a decision log do under the EU AI Act?

Article 12 requires high-risk systems to log events automatically, so that their functioning is traceable to a degree appropriate to the purpose. Under Article 26(6), deployers keep the logs under their control for at least six months unless other law says otherwise. Our decision log records every assignment with timestamp and reasoning.

No. The log shows what happened to each document and why. Whether your Vorgang falls under Annex III, whether an exception applies and who holds which role is a legal assessment for you and your advisers. We provide the facts for it: the Vorgang and its boundary, the log, the intervention path and the written evaluation.

Does the EU AI Act apply if the agent is not high-risk?

Partly. Since 2 February 2025, Article 4 has asked every provider and deployer for measures that support the AI literacy of the people involved. Since 2 August 2026, Article 50 has covered certain systems, such as those interacting directly with people. Articles 8 to 15 and Article 26 apply only to high-risk systems.

The next step

If someone in your organisation assigns the same incoming documents by hand every day, and you need records for an AI employee that a classification can rest on, that is the Vorgang for a Pilotbetrieb: one intake, one task, one destination. Fixed price, agreed before we start. Seven weeks, four of them productive operation on your real documents, every assignment in the decision log.

Request a Pilotbetrieb

← Back to insights