What does the EU AI Act require of deployers?
Under the EU AI Act, a deployer is an organization that uses an AI system under its own authority. A provider develops a system and places it on the market under its own name. Most software companies deploy third-party AI, such as coding agents and chat assistants, and provide the agents they ship. Deployer duties are heaviest for high-risk systems.
Every deployer, whatever the risk level, has two duties that already apply: take measures to support the AI literacy of the people who use AI on its behalf (Article 4), and stay clear of the prohibited practices (Article 5). Transparency duties in Article 50 apply to some deployers, such as those that publish deep fakes. The long list of deployer obligations in Article 26 applies only to high-risk AI systems, which the Act defines in Article 6 and Annex III.
Whether a system is high-risk, and which role you hold for it, is a legal judgment. This guide explains how the obligations are structured so that you can prepare the question for your counsel. It isn't legal advice. For the wider program these duties sit in, see what AI governance is. For the obligations as a list, see the EU AI Act checklist.
The Act is Regulation (EU) 2024/1689. It was amended in July 2026 by Regulation (EU) 2026/1744, the Digital Omnibus on AI, which moved the high-risk dates. Article references in this guide are to the Act as amended, read on EUR-Lex on November 16, 2026.
Are you a provider or a deployer under the EU AI Act?
You're a provider of an AI system if you develop it, or have it developed, and place it on the market or put it into service under your own name or trademark (Article 3(3)). You're a deployer if you use an AI system under your authority in a professional context (Article 3(4)). Personal, non-professional use makes nobody a deployer.
Many companies hold both roles at once, for different systems. The following table shows the usual answer for common situations.
| Situation | Usual role | Why |
|---|---|---|
| Engineers use Claude Code, GitHub Copilot or Cursor | Deployer | You use a system that someone else provides |
| Staff use ChatGPT or another chat assistant at work | Deployer | Professional use of a third-party system |
| You ship a support agent inside your product | Provider | You put an AI system into service under your own name |
You run that same support agent for your own customers | Provider and deployer | You build it and you also use it under your authority |
| You call a model API to build an internal tool | Provider of the tool, deployer of it too | The model vendor provides the model; the tool you build is your system |
You put your brand on, or substantially modify, a high-risk system | Provider | Article 25(1) moves the provider duties to you |
Article 25 matters for deployers. If you change the intended purpose of a system that wasn't high-risk, and the change makes it high-risk, you take on the provider's obligations for it. A general-purpose chat assistant that you start using to screen job applicants is the standard example.
The Act's definition of an AI system in Article 3(1) is legal and broad: a machine-based system that infers from its input how to produce outputs such as predictions, content, recommendations or decisions. For governance work, a narrower working definition helps. An AI system is an application or agent together with the model providers and tools it uses. Grouping things that way lets you assign one role, one owner and one risk tier to each system.
What does Article 26 require of deployers of high-risk AI systems?
Article 26 requires deployers of high-risk AI systems to use them as instructed, put competent people in charge of oversight, monitor them, keep their logs, and inform workers and affected people. The following table paraphrases each paragraph. Read the Official Journal text before you rely on a row.
| Paragraph | What it asks of a deployer (paraphrased) |
|---|---|
| Article 26(1) | Take technical and organizational measures to use the system according to the provider's instructions for use |
| Article 26(2) | Assign human oversight to people with the competence, training, authority and support to carry it out |
| Article 26(3) | Paragraphs 1 and 2 don't limit other legal duties or how you organize your own resources |
| Article 26(4) | Where you control the input data, make sure it's relevant and sufficiently representative for the intended purpose |
| Article 26(5) | Monitor operation against the instructions for use; tell the provider and the market surveillance authority about risks and serious incidents, and suspend use where the Act requires it |
| Article 26(6) | Keep the logs the system generates automatically, to the extent they're under your control, for at least six months unless other law says otherwise |
| Article 26(7) | As an employer, inform workers' representatives and affected workers before you use the system at the workplace |
| Article 26(8) | Public authorities and EU bodies register their use and don't use a system missing from the EU database |
| Article 26(9) | Use the provider's information to carry out any data protection impact assessment that the General Data Protection Regulation (GDPR) requires |
| Article 26(10) | Specific rules for post-remote biometric identification in law enforcement |
| Article 26(11) | Tell people that an Annex III system is used to make, or help make, decisions about them |
| Article 26(12) | Cooperate with the competent authorities |
Article 27 adds a fundamental rights impact assessment for some deployers: public bodies, private entities that provide public services, and deployers of the credit-scoring and life and health insurance pricing systems in Annex III point 5. Since the July 2026 amendment, that assessment can cross-reference a data protection impact assessment that already covers the same ground.
Four of these paragraphs describe work that produces records: oversight by named people (26(2)), input data that matches the purpose (26(4)), monitoring (26(5)) and logs (26(6)). Those are the paragraphs where tooling can supply evidence. The rest are notices, assessments and cooperation, which are your organization's work.
Does the EU AI Act apply to coding agents and chat assistants?
Yes, but usually not as high-risk systems. Writing code and drafting text aren't in the Annex III list, so a deployer of a coding agent or a chat assistant normally carries the duties that apply to every deployer, not Article 26.
Those general duties still matter:
- AI literacy (Article 4). Take measures to support the AI literacy of staff and others who use AI on your behalf, taking their knowledge and the context into account. The July 2026 amendment reworded this duty: it no longer requires you to ensure a particular level of literacy for each person.
- Prohibited practices (Article 5). Don't use AI for the practices that Article 5 bans, such as manipulative techniques that cause significant harm, or emotion recognition at the workplace outside medical and safety uses.
- Transparency (Article 50). If you publish deep fakes, or AI-generated text that informs the public on matters of public interest without human editorial control, disclose it.
Use changes the answer. The same model becomes high-risk when you use it for a purpose that Annex III lists. If you use an assistant to rank job applicants, or use an agent's activity to evaluate engineers for promotion or to allocate their work by behavior, that falls under the employment area of Annex III point 4. Article 25 can then make you the provider as well.
If you ship an agent that talks to people, you're its provider. Article 50(1) asks providers to design such systems so that people know they're interacting with AI, unless that's obvious from the context.
What are the key EU AI Act dates for deployers?
The Act entered into force on August 1, 2024, and applies in stages. Article 113 sets the dates. Regulation (EU) 2026/1744, published in the Official Journal on July 24, 2026 and in force since July 27, 2026, amended them. The following table reflects the amended text as read on EUR-Lex on November 16, 2026.
| Date | What applies | Source |
|---|---|---|
| February 2, 2025 | Chapters I and II: definitions, AI literacy, prohibited practices | Article 113(3)(a) |
| August 2, 2025 | General-purpose AI model rules, governance, penalties and notified bodies | Article 113(3)(b) |
| August 2, 2026 | The rest of the Act, including Article 50 transparency, except the dates that follow | Article 113, second paragraph |
| December 2, 2026 | Two added prohibitions on sexual deep fakes and child sexual abuse material; Article 50(2) marking for generative systems already on the market | Article 113(3)(a) and Article 111(4), as amended |
| December 2, 2027 | High-risk obligations in Chapter III, Sections 1 to 3, for Annex III systems, including Articles 26 and 27 | Article 113(3)(c)(i), as amended |
| August 2, 2028 | The same obligations for high-risk systems under Article 6(1) and Annex I (regulated products) | Article 113(3)(c)(ii), as amended |
| August 2, 2030 | Providers and deployers of high-risk systems used by public authorities, placed on the market before the high-risk dates | Article 111(2), as amended |
Before the amendment, Annex III obligations were due to apply from August 2, 2026, and Annex I obligations from August 2, 2027. Older guides still give those dates. Dates in the Act have been changed once and could change again, so confirm the consolidated text on EUR-Lex on the day you rely on a date.
How long must deployers keep AI system logs?
Deployers of high-risk AI systems must keep the logs the system generates automatically for at least six months, under Article 26(6). The duty covers logs under your control, for a period that suits the system's intended purpose. Other EU or national law can set a different period, and data protection law is the case the Act names.
In practice, a deployer of a third-party high-risk system often holds only part of the logs. The rest sit with the provider, which has its own six-month duty under Article 19. Agree in the contract who keeps what. For the full comparison with NIST AI RMF and HIPAA, and for what a log should hold, see how long to keep AI system logs.
What Arcjet does not cover
Arcjet supplies evidence for some deployer obligations. It isn't a finding that you meet the EU AI Act, and the following work stays with you:
- Classification and role. Deciding whether a system is high-risk and whether you're its provider or deployer is a legal judgment. Arcjet records the answer you reach.
- AI literacy (Article 4). Training and the measures behind it are organizational.
- Notices. Informing workers and their representatives (26(7)), people subject to decisions (26(11)) and people who see generated content (Article 50).
- Assessments. The fundamental rights impact assessment (Article 27) and any data protection impact assessment (26(9)).
- Reports to providers and authorities. Arcjet's records can show what happened; telling the provider or the market surveillance authority is your decision and your action.
- Competence. Arcjet names an owner. It can't show that the owner has the training and authority that 26(2) asks for.
- Systems Arcjet can't see. The inventory covers enrolled macOS and Windows devices, agents and applications that report to Arcjet, and coding agents with hooks installed. A third-party high-risk system that runs entirely in a vendor's cloud keeps its own logs, which Arcjet doesn't hold.
Where Arcjet fits
Article 26 assumes you know which AI systems you run and who is responsible for each. NIST AI RMF asks for the same base in its own terms: an inventory of AI systems (GOVERN 1.6), documented roles (GOVERN 2.1), a documented purpose (MAP 1.1) and policies for third-party AI risk (GOVERN 6.1). Arcjet builds that record from what it observes rather than from a survey.
Discovery finds the AI providers people use on managed macOS and Windows devices. The SDK and coding-agent hooks for Claude Code, GitHub Copilot, Cursor, OpenAI Codex and Muse Code report the agents you build and run. Arcjet groups them into AI systems in the AI inventory, and people confirm the grouping. Each AI system carries an owner, a purpose, a review date, a risk tier and its EU AI Act role.
For a deployer, those records line up with Article 26 as follows:
- 26(1), use as instructed. The documented purpose sits next to the Guard policies that enforce it. Guard decides tool calls, runs rules in dry run first, and won't put a rule live until it passes a test.
- 26(2), human oversight. Each system has a named owner who must be a member of the team; an owner who leaves reads as no owner. Hold decisions route to that owner.
- 26(4), relevant input data. Purpose drift compares the data classes a system declares with what sensitive-information detection finds in its traffic (PII, secrets and PHI) and notifies the owner of differences.
- 26(5), monitoring. Guard records every decision it makes. Provider decisions approve or prohibit each AI provider, and Discovery blocks prohibited ones on managed devices. A use after a prohibition is a violation, attributed to a person.
- 26(6), logs. Console decision history follows your plan's log retention. Opt-in long-term retention keeps decisions for longer, and on the Enterprise plan a log drain exports them to your SIEM or log store.
The evidence page maps these records to the EU AI Act, lists every requirement as supported or not covered, and exports as JSON or CSV or pushes to Vanta and Drata. To work through the obligations one at a time, use the EU AI Act checklist. To decide which systems need Article 26 at all, start with AI risk tiers.
Frequently asked questions
Is a company that uses ChatGPT or Claude Code a deployer under the EU AI Act?
Yes. Using an AI system under your authority for professional purposes makes you a deployer (Article 3(4)). For coding agents and chat assistants used to write code or text, that usually means only the duties every deployer has, such as supporting AI literacy and avoiding prohibited practices, because those uses aren't high-risk.
When do the Article 26 deployer obligations apply?
For high-risk systems listed in Annex III, from December 2, 2027, and for high-risk systems under Annex I, from August 2, 2028. Regulation (EU) 2026/1744 set those dates in July 2026, replacing August 2, 2026 and August 2, 2027. Confirm the consolidated text on EUR-Lex before relying on a date.
Can a deployer become a provider under the EU AI Act?
Yes. Under Article 25, a deployer takes on provider obligations if it puts its name or trademark on a high-risk system, makes a substantial modification to one, or changes the intended purpose of a system so that it becomes high-risk, such as using a general chat assistant to screen job applicants.
Do all deployers need a fundamental rights impact assessment?
No. Article 27 requires it from public bodies, private entities that provide public services, and deployers of the Annex III credit-scoring and life and health insurance pricing systems. Since July 2026, the assessment can cross-reference a data protection impact assessment that covers the same ground.
Did the Digital Omnibus remove the AI literacy duty?
No. It reworded Article 4. Providers and deployers must still take measures to support the AI literacy of staff and others who use AI on their behalf, but the text no longer requires them to ensure a particular level of literacy for each person.
AI runtime security in your code
Protect your AI agent workflows with Arcjet
Coding agents are the AI systems most engineering teams deploy. Arcjet's hooks decide and record each agent action, so the monitoring record exists before anyone asks for it.