EU AI Act Compliance Checklist for Deployers: Deadlines Mapped Through 2028

If your organization uses AI systems under its own authority in the EU, whether that means screening candidates, scoring credit or publishing AI-drafted content, you are a deployer under Regulation (EU) 2024/1689. As of 30 September 2026, three sets of duties already bind you: the Article 5 prohibitions and the Article 4 AI literacy duty (since 2 February 2025), the penalty regime (since 2 August 2025), and the Article 50 transparency duties (since 2 August 2026). Two new prohibitions follow on 2 December 2026. The Article 26 and 27 high-risk deployer obligations were postponed by the Digital Omnibus on AI, Regulation (EU) 2026/1744, to 2 December 2027 for Annex III systems and 2 August 2028 for AI embedded in regulated products.

Both the "the AI Act is delayed, relax" reading and the "everything hits at once" panic are wrong. The omnibus is not a compliance holiday for deployers. Three sets of duties bind you now, one more arrives in two months, and the fourteen months of runway to December 2027 is roughly what the slow work in Articles 26 and 27 takes: system inventory, oversight design, log-retention plumbing and fundamental rights impact assessments do not compress into one quarter.

The checklist below has five deadline buckets and twenty obligations, each with a Monday-morning action and an evidence column. Every date was checked on 30 September 2026 against the Official Journal text and the European Commission's own pages, both listed in the sources.

Four questions decide whether this checklist applies to you

Work through these before the tables, because the answers change which sections bind you.

1. Do we use AI systems under our authority in the course of a professional activity in the EU? If yes, you are a deployer (Article 3(4)). "Under our authority" is the operative phrase: a vendor's chatbot embedded in your hiring workflow is your deployment even though you did not build it.
2. Do any of those systems fall in an Annex III category (employment, credit scoring, education, essential services, biometrics, law enforcement, migration, justice)? If yes, the December 2027 bucket is your main event, and classification deserves its own analysis.
3. Have we rebranded, substantially modified or repurposed any AI system? If yes, Article 25 may have made you a provider. The section on Article 25 below covers the three triggers.
4. Are we a public body, a private entity providing public services, or a deployer of credit-scoring or life and health insurance risk-assessment systems? If yes, the Article 27 fundamental rights impact assessment applies to you specifically, not to all high-risk deployers.

If you answered no to question 1, you may still be in scope as a provider, importer or distributor. This article covers deployers only.

The deployer calendar has five dates behind it and six ahead

Get the omnibus status right, because getting it wrong is the fastest way to embarrass yourself in front of a board. The European Commission proposed the Digital Omnibus on AI on 19 November 2025. The co-legislators reached a trilogue agreement on 7 May 2026, Parliament's plenary approved it on 16 June 2026, the Council adopted it on 29 June 2026, and it was signed on 8 July 2026 (European Parliament Legislative Train). It was published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744 and entered into force on the third day after publication, 27 July 2026 (EUR-Lex; European Commission).

One correction to the way this is usually reported: do not say "the AI Act was delayed." The omnibus set new fixed dates for the two high-risk tranches and left everything else on schedule. It also reworded Article 4, added two prohibitions and extended the lower-of fine rule to small mid-cap companies for the two lower fine tiers.

Sources for the table: Articles 111 and 113 of Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, and the Commission's application timeline.

The deployer checklist sorts every obligation into five dated buckets

Deadlines are stated as of 30 September 2026. "Evidence to retain" is what a market surveillance authority, or your own auditor, could ask for. The canonical source throughout is Regulation (EU) 2024/1689 on EUR-Lex, read together with the amending Regulation (EU) 2026/1744.

Bucket 1: binding since 2 February 2025

Two notes for counsel. First, Article 4 was rewritten, not repealed. Deployers still "shall take measures to support the development of AI literacy," but the text now says the duty "does not require providers or deployers to guarantee any specific level of AI literacy of any individual," and the Commission and Member States take on a supporting role. Do not dismantle a training program on the strength of the softer wording. Second, treat literacy as a control: if your teams use language-model tools in regulated workflows, the training should cover their failure modes, the territory mapped in our analysis of LLM accuracy on regulatory text.

Bucket 2: binding since 2 August 2025

Article 53(1)(b) obliges model providers to make documentation available to providers of AI systems who integrate the model. A deployer that only uses a vendor's finished system should look to that vendor's instructions for use instead.

Bucket 3: binding since 2 August 2026, Article 50 transparency

Article 50 has applied since 2 August 2026, and it applies to in-scope systems regardless of when they were placed on the market, with one exception: providers of generative systems placed on the market before that date have until 2 December 2026 to meet the Article 50(2) marking duty (Article 111(4), bucket 4). Most coverage misses that Article 50(1) and 50(2), the chatbot disclosure and the machine-readable marking, are provider duties. Deployers own 50(3) and 50(4), and the timing rule in Article 50(5) is "at the latest at the time of the first interaction or exposure."

Two Commission documents now sit around the article. The Commission's guidelines on the transparency obligations were approved in content on 20 July 2026; the Commission will formally adopt them once all language versions exist, and says they apply only from then, so treat them as its stated reading in the meantime. The Code of Practice on Transparency of AI-generated Content was assessed by the Commission on 8 July 2026 as adequately covering Articles 50(2), 50(4) and 50(5), with the caveat that adherence "does not constitute conclusive evidence of compliance." The guidelines add a practical point for anyone who does not sign the code: they should be ready to show how their labeling meets the article by other means, and may face more detailed information requests.

On row 7, the guidelines are specific in ways that matter to compliance teams. Review must be substantive, with fact-checking as a minimum; spell-checking, an editorial policy on paper or cursory approval does not qualify; and a substantive AI edit after sign-off voids the exception. The person or function holding editorial responsibility should have its identity and contact details published somewhere easy to find. The examples include AI-manipulated corporate reports on a listed company's website containing investor information, which are in scope unless reviewed, while private correspondence, organization-internal communications and AI-manipulated compliance advice from a consultant to a client are not. Deep fakes generated or manipulated before 2 August 2026, and public-interest texts generated and published before that date, do not need retroactive labels, but text generated earlier and published on or after that date does.

Bucket 4: 2 December 2026, the omnibus's own deployer deadline

The omnibus is mostly read as a postponement. It also added two prohibitions to Article 5. Point (ba) covers AI systems that generate or manipulate realistic images, video or audio of an identifiable person's intimate parts or sexually explicit activities without that person's consent, and point (bb) covers material within the child sexual abuse directive. For deployers, Article 5(1a) narrows the prohibition to using a system "for the purpose of generating or manipulating" such material, while providers are caught where that is the intended purpose or the design makes it a foreseeable, reproducible outcome without adequate safeguards. Both apply from 2 December 2026.

Bucket 5: 2 December 2027, the Article 26 and 27 stack for Annex III systems

This is the bucket the omnibus moved from 2 August 2026, and it is where the real deployer workload lives: the paragraphs of Article 26, Article 27 for some deployers, and the Article 86 right to an explanation, which you should be ready to honor. Articles 26 and 27 sit in Chapter III, Section 3, so they follow the Annex III date. Article 86 sits outside Chapter III and formally applies from 2 August 2026, but it attaches to decisions based on Annex III high-risk systems, whose rules start on 2 December 2027, so it is grouped here.

Two scope notes that are easy to get wrong. The fundamental rights assessment is not a duty on all high-risk deployers: only the row 20 categories, with Annex III point 2 (critical infrastructure) excluded. And Article 26(10), the judicial-authorization rule for post-remote biometric identification, covers a narrow law-enforcement scenario, not commercial biometrics. Deployers also owe cooperation with authorities under Article 26(12), which needs no project of its own. One helpful change from the omnibus: Article 27(4) now lets you cross-reference, or copy in, the relevant sections of your data protection impact assessment inside the fundamental rights assessment, and the AI Office's questionnaire template, which already had to include an automated tool, must allow the same.

The long tail. AI embedded in Annex I regulated products follows on 2 August 2028. Under Article 111(2), high-risk systems already on the market when the high-risk rules start come into scope only on a significant design change, except systems intended for public authorities, which must comply by 2 August 2030 regardless. Components of the Annex X large-scale IT systems have until 31 December 2030. If you deploy a pre-existing system, log the version you run and watch vendor release notes: a "significant change" is what pulls you in.

December 2027 is the runway Article 26 actually needs

The tempting reading of the omnibus is "we got extra time, deprioritize." Walk through rows 11 to 20 and the arithmetic argues otherwise.

The prerequisite for everything is an AI system inventory you would sign your name to. Conformance with instructions for use means obtaining and reading provider documentation per system. Oversight design means naming and training people to intervene, which collides with real org charts. Log retention is an engineering ticket with vendor dependencies. Worker notification runs through works councils on their timetable. A fundamental rights assessment has six mandatory content elements and must be finished before first use. Each item has a lead time of months, and they serialize: you cannot design oversight for systems you have not inventoried.

For that work, across a hypothetical portfolio of ten Annex III systems in four Member States, fourteen months from today is approximately the critical path. Teams that treat December 2027 as a starting gun will find in mid-2027 that a vendor exposes no logs and the works council meets quarterly.

Article 25 quietly turns deployers into providers

The most expensive misclassification in the Act is believing you are a deployer when Article 25(1) has already made you a provider. Any distributor, importer, deployer or other third party is treated as the provider of a high-risk system, with the obligations of Article 16, in three circumstances:

1. Rebranding. You put your name or trademark on a high-risk system already on the market, without prejudice to contractual arrangements that allocate the obligations otherwise. White-labeling a vendor's screening tool as "YourCo Talent Insights" is the canonical case.
2. Substantial modification. You make a substantial modification to a high-risk system already on the market, in such a way that it remains high-risk.
3. Repurposing into high-risk. You modify the intended purpose of an AI system that is not classified as high-risk, including a general-purpose system, so that it becomes high-risk. Wiring a general-purpose model into CV scoring is the pattern to watch: the model's provider never intended employment use, your integration created a high-risk system, and you are its provider.

The consequence is the full provider stack (risk management, technical documentation, conformity assessment, registration) on the provider timeline, not the deployer one. This is why scoping question 3 sits ahead of the checklist: every internal project that customizes a vendor system deserves an Article 25 review before anyone celebrates the lighter deployer duties. One institutional footnote from the omnibus: where the same provider, or its group, supplies both a general-purpose model and the system built on it, the AI Office is exclusively competent to supervise that system, with exceptions that include AI in Annex I products and financial institutions' systems covered by Article 74(6) (Article 75(1) as amended). That competence reaches deployers only when they are also the provider or part of the same undertaking.

Penalties top out at EUR 35 million or 7%, and deployer duties sit one tier down

The fining regime in Article 99 has applied since 2 August 2025, and the AI Office and national authorities began enforcing the Act on 2 August 2026. The tiers as enacted:

SMEs pay the lower of the percentage or the fixed amount, which inverts the usual "whichever is higher" rule. Since 27 July 2026, Article 99(6a) gives small mid-cap companies the same treatment for the tiers in paragraphs 4 and 5. Enforcement began only on 2 August 2026, so precedent is thin. The mechanics are more predictable. National market surveillance authorities enforce against deployers, and first scrutiny will plausibly land on the oldest and cheapest duties to check, the Article 5 screens and the Article 50 notices, both visible from outside the organization.

Scope your AI systems this week and sequence the work before 2 December 2026

Run the four scoping questions this week and put your name on the answers. Then sequence the work:

1. Close the Article 50 gaps now, since the duties already apply: label deepfakes, decide per channel whether public-interest text gets substantive review with a named responsible party or a label, and place notices for any emotion recognition or biometric categorization.
2. Refresh the Article 5 screen with the two new items and update the acceptable-use policy before 2 December.
3. Put the December 2026 vendor confirmations on a named owner's calendar.
4. Start the Annex III inventory, because it gates everything in bucket 5.

If you also operate in crypto or US markets, the parallel clocks (our MiCA checklist covers one, and the US state privacy laws tracker another) argue for one consolidated register rather than per-regime spreadsheets. How compliance teams track this at scale is its own topic.

The checklist becomes a control when its rows turn into owned deadlines. Convert each applicable row into an entry in your obligations register (our register template works) with an owner, the deadline above and the evidence column as the artifact you attach when the row closes. RegWatch, my company, is built around that workflow: a watchlist on the AI Act and its national implementations produces findings on the schedule you set, triage turns relevant ones into alerts with cited reasoning, and an accepted alert converts to an obligation with an owner, a deadline and evidence. The mechanics are in how AI agents do regulatory compliance monitoring. For the impact-assessment discipline to hang off each change, start from our regulatory change impact assessment template. More of this regime lives in the AI compliance hub.

Whatever the tooling, check whether your deadline map changed with a dated change record when the omnibus reached the Official Journal on 24 July 2026, or silently went stale.

---

This article is general information, not legal advice.