EU AI Act checklist for providers and deployers

Every EU AI Act obligation by article, split by role, with who usually produces the evidence and which rows Arcjet records.

13 min read
In short: Every organization must support AI literacy, avoid prohibited practices and meet any Article 50 transparency duties that apply to it. Deployers of high-risk systems follow Article 26, and providers follow Articles 9 to 20, conformity assessment, registration and post-market monitoring. High-risk rows apply from December 2, 2027 for Annex III systems, as amended in July 2026.

What is on an EU AI Act checklist?

An EU AI Act checklist lists the obligations of Regulation (EU) 2024/1689 by role. Every organization must support AI literacy, avoid prohibited practices and meet any transparency duties that apply to it. Deployers of high-risk systems follow Article 26: instructed use, human oversight, monitoring and six months of logs. Providers of high-risk systems carry the rest, from risk management to incident reports.

This checklist splits those obligations into three groups: every organization, deployers of high-risk systems, and providers of high-risk systems. Each row gives the article, a short paraphrase in our own words, who usually produces the evidence, and what Arcjet records toward it. It reflects the Act as amended by Regulation (EU) 2026/1744, read on EUR-Lex on November 16, 2026.

Most rows apply only to high-risk AI systems. Whether a system is high-risk, and whether you're its provider or deployer, is a legal judgment. Get advice before you rely on a row. For the reasoning behind the deployer rows, see the EU AI Act for deployers. For where the Act sits in a governance program, see what AI governance is.

How do you use this EU AI Act checklist?

Use it per AI system, not once for the whole company. An AI system is an application or agent together with the model providers and tools it uses. Each system gets a role and a risk tier, and those two answers decide which groups of rows apply to it.

  1. List your AI systems. The AI inventory guide covers how.
  2. For each system, decide with counsel whether you're its provider, its deployer or both. Article 3(3) and 3(4) define the roles, and Article 25 says when a deployer becomes a provider.
  3. Decide whether the system is high-risk under Article 6, Annex I or Annex III. The AI risk tiers guide explains the categories.
  4. Work through the first group for every system, then the deployer or provider group for each high-risk system.
  5. Check the dates table before you treat a row as in force.

The "Who" column uses three labels. You is your organization: legal, compliance and leadership. Eng is your engineering team. Arcjet means Arcjet records the evidence. A row marked "None" in the Arcjet column has no Arcjet evidence, and that's deliberate.

What must every organization do under the EU AI Act?

Every provider and deployer, at any risk level, must support AI literacy, avoid the prohibited practices and know which of its systems are high-risk. Transparency duties apply to the providers and deployers that Article 50 names.

ArticleItemWhoArcjet evidence
Article 4

You take measures to support the AI literacy of staff and others who use AI on your behalf, suited to their knowledge and context

YouNone
Article 5

No prohibited practice is used, including the two added for sexual deep fakes and child sexual abuse material

YouNone
Articles 3 and 25

Your role for each system is decided: provider, deployer or both, and whether a change of purpose made you a provider

You, ArcjetEU AI Act role recorded per AI system
Article 6, Annexes I and III

You know which systems, if any, are high-risk, and you record the reasoning

You, ArcjetRisk tier recorded per AI system
Article 6(4)

A provider that decides an Annex III system isn't high-risk documents that assessment before release

YouNone
Article 50(1) and 50(2)

Providers design systems so people know they're talking to AI, and mark generated content as machine-readable

EngNone
Article 50(3) and 50(4)

Deployers disclose emotion recognition, biometric categorisation, deep fakes and AI-written public-interest text

YouNone

The July 2026 amendment rewrote Article 4. The duty is to take measures that support AI literacy, and it no longer requires a guaranteed level for each person. A record of who completed which training is still the usual evidence.

What must deployers of high-risk AI systems do?

Deployers of high-risk AI systems must follow Article 26: use the system as instructed, assign competent oversight, monitor it, keep its logs and inform the people it affects. Some deployers also owe a fundamental rights impact assessment under Article 27.

ArticleItemWhoArcjet evidence
Article 26(1)The system is used according to its instructions for useEng, ArcjetDocumented purpose and the Guard policies that enforce it
Article 26(2)

Human oversight is assigned to people with competence, training and authority

You, Arcjet

Named owner, Hold decisions routed to the owner. Competence is yours

Article 26(4)Input data you control is relevant to the intended purposeEng, ArcjetDeclared versus observed data classes (purpose drift)
Article 26(5)

Operation is monitored; the provider and authorities are told of risks and serious incidents

Arcjet, YouRuntime decision records and drift notices. Reporting is yours
Article 26(6)

Automatically generated logs under your control are kept for at least six months, unless other law sets a period

Arcjet

Long-term retention, or export to your SIEM through a log drain, for systems that report to Arcjet

Article 26(7)

Workers and their representatives are informed before workplace use

YouNone
Article 26(8)

Public authorities register their use and use only registered systems

YouNone
Article 26(9)

Provider information is used for any required data protection impact assessment

YouNone
Article 26(11)

People subject to decisions made or assisted by an Annex III system are informed

YouNone
Article 26(12)You cooperate with the competent authoritiesYouEvidence export supports requests
Article 27

A fundamental rights impact assessment is done where required: public bodies, public-service providers, credit scoring, life and health insurance pricing

YouNone

Article 26(10) covers post-remote biometric identification in law enforcement. It's left out of the table because it doesn't apply to software companies in the ordinary course.

For logs, read Article 26(6) together with your contract. A deployer of a third-party high-risk system keeps the logs under its own control, and the provider keeps the rest under Article 19. For periods, contents and the data protection question, see how long to keep AI system logs.

What must providers of high-risk AI systems do?

Providers of high-risk AI systems must build them to the requirements in Chapter III, Section 2, and meet the obligations in Article 16: a quality management system, documentation, logs, conformity assessment and registration. Article 72 adds post-market monitoring. If you ship an agent that falls under Annex III, these rows are yours.

ArticleItemWhoArcjet evidence
Article 9A risk management system runs across the life cycleYou, ArcjetInventory, runtime enforcement and drift as inputs
Articles 10 and 11Data governance and technical documentationEngNone
Article 12The system records events automatically over its lifetimeEng, ArcjetGuard decision records for your agents
Article 13

Deployers receive instructions for use and the information they need

EngNone
Article 14The system is designed for effective human oversightEng, ArcjetHold decisions routed to the owner
Article 15Accuracy, robustness and cybersecurity are appropriateEng, ArcjetPrompt injection and tool-call enforcement (cybersecurity only)
Article 17A quality management system is in placeYouNone
Article 18Documentation is kept for 10 years after releaseYouNone
Article 19

Logs under the provider's control are kept for at least six months

ArcjetLong-term retention, or export to your SIEM
Article 20

Non-conforming systems are corrected, withdrawn or recalled, and others are informed

You, EngNone
Articles 43, 47, 48 and 49

Conformity assessment, EU declaration of conformity, CE marking and registration

YouNone
Articles 72 and 73Post-market monitoring; serious incidents are reportedYou, ArcjetRuntime records. Reporting is yours

Providers of general-purpose AI models, such as the companies that train foundation models, have a separate set of duties in Chapter V. Those apply to the model vendor, not to a company that calls the model through an API, so they're out of scope here.

When does each part of the EU AI Act checklist apply?

Each group has its own application date, set by Article 113 as amended by Regulation (EU) 2026/1744. That regulation was published in the Official Journal on July 24, 2026 and has been in force since July 27, 2026. The following dates were read on EUR-Lex on November 16, 2026.

Checklist groupApplies fromSource
Article 4 AI literacy and Article 5 prohibitionsFebruary 2, 2025Article 113(3)(a)
Two added Article 5 prohibitionsDecember 2, 2026Article 113(3)(a), as amended
Article 50 transparency

August 2, 2026; marking under 50(2) for generative systems already on the market by December 2, 2026

Article 113 and Article 111(4), as amended
Deployer and provider rows for Annex III systemsDecember 2, 2027Article 113(3)(c)(i), as amended

Deployer and provider rows for Annex I systems (regulated products)

August 2, 2028Article 113(3)(c)(ii), as amended

High-risk systems for public authorities already on the market

August 2, 2030Article 111(2), as amended

The amended date in Article 113(3)(c) covers Chapter III, Sections 1 to 3, which hold Articles 6 to 27. Articles 43 to 49 and Articles 72 and 73 sit outside those sections, but they concern high-risk systems, so they matter in practice only once a system is subject to the high-risk rules. Confirm how each applies to your systems with counsel.

The original Act set August 2, 2026 for Annex III systems and August 2, 2027 for Annex I systems. Guides written before July 2026 still use those dates. The dates have changed once already, so check the consolidated text on EUR-Lex before you rely on any of them.

What Arcjet does not cover

Arcjet supplies evidence for some rows. It isn't a finding that you meet the EU AI Act, and most rows carry no Arcjet evidence at all. The following work stays with you:

  • Legal judgments. Your role for each system, its risk classification and any Article 6(4) assessment.
  • Training and policy. AI literacy, the quality management system and the risk management process itself.
  • Notices and disclosures. Article 50 transparency, and informing workers and affected people under Article 26.
  • Assessments and documents. Fundamental rights and data protection impact assessments, technical documentation, conformity assessment, the declaration of conformity, CE marking and registration.
  • Reports. Telling providers and authorities about risks and serious incidents.
  • Model quality. Accuracy, robustness beyond cybersecurity, and data governance for training data.
  • What Arcjet can't see. Arcjet observes enrolled macOS and Windows devices, applications and agents that report to it, and coding agents with hooks installed. A system outside those sources isn't in the record.

Where Arcjet fits

The rows with Arcjet evidence share a pattern: they ask for a record of what a system is, who is accountable for it, and what it did. Arcjet builds that record from what it observes rather than from a questionnaire.

Discovery finds the AI providers people use on managed macOS and Windows devices, and enforces provider decisions there. The SDK and coding-agent hooks report the agents you build and run, and Guard records every decision it makes on their tool calls. The AI inventory groups these into AI systems, each with an owner, a purpose, a 180-day review, a risk tier and an EU AI Act role. Hold decisions go to the owner, and a Guard policy needs a second approver to publish.

The evidence page maps those records to the EU AI Act. It lists every requirement as supported or not covered, so the "None" rows in this checklist show up there too. You can export it as JSON or CSV, or push it to Vanta and Drata. The export is the record at that moment, because the page doesn't recompute past dates.

To see how the same evidence maps to other frameworks, read NIST AI RMF vs ISO 42001 vs EU AI Act. To get the export into your GRC tool, see AI governance evidence export.

Frequently asked questions

Which EU AI Act obligations apply to a company with no high-risk AI systems?

Supporting the AI literacy of staff who use AI (Article 4), avoiding the prohibited practices (Article 5), and the transparency duties in Article 50 where they apply, such as telling people they are talking to an AI agent you provide or disclosing deep fakes you publish.

Does the EU AI Act require an inventory of AI systems?

Not in those words. But you can't decide your role, classify each system or apply Article 26 without knowing which AI systems you run, so an inventory is the practical starting point for every row of the checklist.

How long must high-risk AI system logs be kept under the EU AI Act?

At least six months, for logs under your control, unless other EU or national law sets a different period. Article 19 sets this for providers and Article 26(6) for deployers. Data protection law can limit how long personal data in those logs is kept.

Is working through this checklist enough to meet the EU AI Act?

No. It lists obligations so you can assign them. Classification and role are legal judgments, many rows are assessments, notices and documents that only your organization can produce, and Arcjet supplies evidence for some rows, not a finding that you meet the Act.

Are general-purpose AI model obligations on this checklist?

No. Chapter V obligations for general-purpose AI models apply to the companies that train and release those models. A company that builds on a model through an API isn't a model provider for that reason alone, though it may be the provider of the AI system it builds.

AI runtime security in your code

Protect your AI agent workflows with Arcjet

For the agents you build, Article 12 asks for automatic event records. A guard on each tool call records the decision, rule and reason before the action runs.