What the EU AI Act asks of a deployer

An agent that matches supplier invoices and an agent that filters job applicants may run on the same platform. That does not give them the same obligations under the EU AI Act. The purpose of the work matters, as does the part your company plays in providing or using the system.

For an operations team, the starting questions are practical. Are we a deployer, a provider, or both? Is this use high-risk? What must we do now, and what evidence will we need when the remaining rules apply?

We covered the broader picture in the rulebooks arriving around AI. This article takes the deployer’s side: the organisation using an AI system under its authority. It focuses on recurring business work, not the specialist rules for law enforcement or every product-safety regime.

01Are you a deployer or a provider?

A deployer uses an AI system under its authority, outside purely personal, non-professional activity. Buying a vendor’s system and using it for its documented purpose will often put your company in that role.

A provider develops a system, or has one developed, and places it on the market or puts it into service under its own name or trademark. You do not have to train the underlying model or sell software to other companies to qualify. Commissioning a system for your own use can also matter. These are the definitions in Article 3.

The roles attach to the system and the activity. A company can be a deployer of one system and the provider of another. Buying an agent platform does not, by itself, settle the status of everything built on it.

Article 25 identifies three circumstances in which a deployer or another party takes on the obligations of a high-risk system provider:

• Putting its own name or trademark on an existing high-risk system, subject to the provision concerning contractual allocation of obligations.
• Substantially modifying an existing high-risk system in a way that leaves it high-risk.
• Changing the intended purpose of a system that was not high-risk so that it becomes high-risk.

Routine configuration is not automatically a substantial modification. Equally, a change described internally as “just a new workflow” may change the intended purpose. Extending a document assistant into applicant screening deserves a fresh assessment before it goes live.

02Is the work high-risk?

An agent is not high-risk simply because it acts autonomously. Nor is it outside the high-risk rules because someone approves its output.

There are two main routes under Article 6. One concerns certain products and safety components covered by the legislation in Annex I, where the relevant third-party conformity assessment condition is met. The other concerns the uses listed in Annex III, including particular applications in employment, education, essential services, biometrics and critical infrastructure.

For many operations teams, Annex III is the more immediate question. Screening job applicants, evaluating the creditworthiness of individuals, and risk assessment or pricing for life and health insurance are listed uses. Matching supplier invoices or chasing a delivery date will not normally fall within those categories on that description alone.

There are limited exceptions for some Annex III systems under Article 6(3), including certain narrow procedural or preparatory tasks that do not pose a significant risk. These are conditions to assess, not labels to apply casually. Annex III systems that profile natural persons are always treated as high-risk under that provision.

Ask the provider for the intended purpose and classification, then compare those with what you actually plan to do. Record the decision and the reasons. If the work changes, revisit it.

This is related to, but separate from, whether the work is suitable for an AI agent. A process can be legally permissible and still be too expensive to check or too difficult to supervise.

03What applies now, and what was delayed?

The Digital Omnibus on AI, Regulation (EU) 2026/1744, amended the AI Act in July 2026. It postponed specified high-risk provisions, not the whole Act. The Council’s adopted text sets out the changes to Articles 111 and 113.

These are general dates. Amended Article 111 contains transitional rules for high-risk systems already placed on the market or put into service before the relevant Chapter III date, including conditions concerning significant design changes. Establish which rules apply to the particular system rather than treating December 2027 as a universal deadline for every existing deployment.

AI literacy still requires action

The amended Article 4 requires measures to support the development of AI literacy. It does not require a guaranteed level for every individual. Training and guidance should reflect the people, the systems and the context of use. Keep a record of what was provided and why it was appropriate. The Commission’s AI literacy questions and answers explains the amended approach.

Transparency depends on the use and the role

Article 50 separates provider duties from deployer duties. Providers must design systems intended to interact directly with people so that people know they are dealing with AI, unless that is already obvious in the circumstances. Providers also carry the machine-readable marking obligation for synthetic content.

Deployers have specific duties for emotion recognition and biometric categorisation, deepfakes, and AI-generated or manipulated text published to inform the public on matters of public interest. The public-interest text duty has an exception where there is human review or editorial control and a person or organisation holds editorial responsibility. Special rules also apply to evidently artistic, creative, satirical or fictional deepfakes. See Article 50.

For an existing generative system placed on the market before 2 August 2026, the provider’s Article 50(2) marking deadline is 2 December 2026. That transition does not postpone the deployer’s own disclosure duties.

The Omnibus also adds prohibitions concerning specified non-consensual sexual or intimate content and child sexual abuse material from 2 December 2026. A disclosure cannot make a prohibited use lawful.

The GDPR, employment rules and applicable sector requirements continue alongside the AI Act. The high-risk deferral does not suspend them.

04What does Article 26 require of a high-risk deployer?

Once Article 26 applies to your system, its main requirements concern how the work is used, supervised and monitored. The following is a practical summary, not a replacement for the full provision or its sector-specific conditions.

The full wording is in Article 26. Having somebody available to approve an output is not enough if they cannot understand the limitations, obtain the relevant information or stop the work.

Article 27 adds a fundamental rights impact assessment for specified deployers and uses. It covers bodies governed by public law, private entities providing public services, and deployers of the listed creditworthiness and life or health insurance systems. The provision excludes the Annex III critical-infrastructure category. It is not an assessment required of every private company using high-risk AI.

Where required, the assessment precedes first use, must be kept current, and its results must be notified as specified. The Omnibus permits cross-references to relevant parts of an existing data protection impact assessment, or their inclusion in the fundamental rights assessment. It does not remove the assessment. See Article 27 and its amendments.

Six months is a retention period

Article 26(6) concerns how long the relevant logs must be retained once the duty applies. It does not require six months of historical logs to exist on the first day.

Starting earlier can help you test monitoring and investigate failures, but it should be a deliberate operating decision. Set the retention period, access permissions and deletion rules together, taking account of applicable personal-data requirements.

05What happens when an incident is found?

When Article 26(5) applies, a deployer identifying a serious incident must immediately inform the provider, then the importer or distributor and the relevant market surveillance authorities. If the provider cannot be reached, Article 73 applies to the deployer with the necessary adaptations.

For a risk of the kind specified in Article 26(5), the duty includes informing the relevant parties without undue delay and suspending use. Not every incorrect output is a legally defined serious incident, but the team needs a route for assessing one without waiting for a monthly review.

Article 73 primarily sets the provider’s reporting requirements. Its general outer limit is 15 days after awareness, with shorter limits of 10 days for a death and two days for a widespread infringement or the specified serious disruption of critical infrastructure. Those limits sit alongside requirements to report immediately at the relevant trigger. They are not permission for a deployer to wait before notifying under Article 26.

The Commission’s incident-reporting material includes draft guidance. Base the incident procedure on the statutory triggers and check the status of any guidance before relying on it.

Before launch, agree who receives an alert, who can suspend the system, how the provider is contacted and who makes any required notification. Test that route. Logs may establish what happened, but they do not contact the right people on their own.

06Can a person ask why the AI influenced a decision?

Article 86 gives affected people a right to a clear and meaningful explanation in defined circumstances. It concerns decisions based on outputs from Annex III high-risk systems, except the critical-infrastructure category, which have legal or similarly significant effects and are considered by the person to adversely affect their health, safety or fundamental rights.

The explanation must cover the AI system’s role and the main elements of the decision. The provision includes exceptions and applies only to the extent that the right is not already provided under other EU law.

Article 86 was not expressly included in the postponement of Chapter III, Sections 1 to 3. A reading of that amendment alone therefore leaves the general 2 August 2026 date in place. How the provision interacts with deferred classification rules and transitional treatment needs to be assessed for the particular system. Do not assume that a later Article 26 date also postpones every explanation duty.

Operationally, retain the information needed to explain the decision: the relevant facts, criteria, system output and human involvement. A technical event log may help locate that evidence. It is not necessarily a meaningful explanation to the person affected.

07What should the operating record contain?

Take one agent role and ask whether someone outside the team could establish the following from the records you keep:

This is an operating checklist, not a claim that the Act requires one standard file containing all six items. Some records belong with the agent platform; others belong with the process owner, HR, privacy or legal team.

This is where Aitonomy’s work fits. The Scan and our team establish the intended work, the systems involved and the gaps in the available evidence. Where classification or a change in legal role needs assessment, we prepare the technical and operational information for the customer and its advisers.

In production, Aitonomy Control holds the operating evidence: agent activity, approvals, checks and the history of changes and supervision decisions. Our team uses that record to investigate problems and bring decisions to the customer’s accountable process owner. The supervision levels set out how authority is earned and reduced. They support human oversight; they are not a substitute for all of Article 26.

Control does not, by itself, complete an impact assessment, notify workers or determine whether a configuration makes someone a provider. The required logs, access and retention also need to be checked for the actual deployment. Those responsibilities have to be assigned, not assumed to be covered by buying a platform.

Start with the agent you already run or are considering. Record its purpose and the parties’ roles, establish which duties apply, and test whether the responsible people can retrieve the evidence and act on it. That work is useful before a deadline because it is also what allows you to run an agent responsibly in production.

Common questions