Home / Questions

When does the EU AI Act apply? The timeline questions, answered

The EU AI Act does not switch on all at once: Regulation (EU) 2024/1689 entered into force in August 2024 and Article 113 then staggers application across several dates, which is why two people can both be right when one says 2025 and the other says 2027. Below are the questions that follow from that, answered by article, with pointers to the primary text rather than to a secondary summary.

Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.

What the dates actually attach to

The question "when does the EU AI Act apply" has more than one correct answer because the regulation's final article applies different chapters on different dates. Article 113 sets a general application date and then carves out earlier dates for the prohibitions and for the general-purpose AI and governance chapters, and a later date for one route into high-risk classification. So before you can answer the question for your organisation, you need two facts about each system in your inventory: which chapter's obligations reach it, and whether it is high-risk because of Annex III or because it is a safety component of a product already regulated under the Union legislation listed in Annex I.

That is also why a single internal deadline tends to mislead. A hiring screening model and an AI component inside a regulated medical device sit under the same regulation on different clocks, and a model your team shipped before the general application date may be treated differently again under the transitional provisions in Article 111. The table below is the skeleton; the answers that follow fill in the parts people get wrong.

The milestones as written in the regulation

DateWhat starts applyingWhere to read it
1 August 2024Entry into force of Regulation (EU) 2024/1689, twenty days after publication in the Official JournalArticle 113, opening paragraph
2 February 2025Chapter I general provisions and Chapter II prohibited AI practices, including the AI literacy duty in Article 4Article 113, point (a)
2 August 2025Notifying authorities and notified bodies, general-purpose AI model obligations, governance, confidentiality and the penalty provisions other than Article 101Article 113, point (b)
2 August 2026General application of the regulation, which is the date most Annex III high-risk duties are read againstArticle 113, opening paragraph
2 August 2027Article 6(1), the classification route for AI that is a safety component of a product covered by the Union legislation in Annex I, and the obligations that follow from itArticle 113, point (c)
TransitionalSeparate treatment for general-purpose AI models already placed on the market and for high-risk systems already placed on the market or put into service, with a longer horizon for systems used by publArticle 111

Questions

When does the EU AI Act apply?

Regulation (EU) 2024/1689 entered into force on 1 August 2024, and Article 113 then applies it in stages rather than on a single date. The prohibited-practices chapter and the general provisions, including the AI literacy duty, applied from 2 February 2025. A second tranche covering general-purpose AI models, notified bodies, governance and most penalty provisions applied from 2 August 2025. The general application date is 2 August 2026, with one later carve-out described in the next answer.

When do high-risk obligations actually apply?

It depends on which route into high-risk classification your system takes. Systems that are high-risk because of the use cases in Annex III are read against the general application date of 2 August 2026. Systems that are high-risk under Article 6(1), meaning AI used as a safety component of a product already covered by the Union legislation listed in Annex I, are given the later date in Article 113 point (c), 2 August 2027, together with the obligations that follow from that classification. If your inventory contains both kinds, you are managing two deadlines, not one, and the evidence you need to hold looks the same in both cases.

Has the high-risk timeline been delayed?

Amendments to the timing of parts of the high-risk framework have been publicly proposed and debated since the regulation was adopted, and we will not restate a status we cannot certify on the day you read this page. The honest working answer is to plan against the dates in Article 113 as adopted, and to check the consolidated version of the regulation on EUR-Lex plus your national supervisory authority's guidance before you move an internal milestone. A delay, if one lands, changes when an authority can act on a gap; it does not change what the record-keeping obligations ask you to have recorded. Building the log now is the part that does not become wasted work either way.

What changed on 2 February 2025?

Article 113 point (a) brought Chapters I and II into application, which covers the general provisions and the list of prohibited AI practices. This is also the date the AI literacy duty in Article 4 attaches, which is why many organisations ran training and inventory exercises first and left logging architecture for later. Practically, the work that followed was a screening exercise: confirming that nothing in the estate falls into a prohibited category before spending effort on high-risk documentation. That screening output tends to become the AI use-case registry the rest of the programme depends on.

Do systems already in production before the general application date fall in scope?

Article 111 sets transitional arrangements rather than a blanket exemption, and it treats general-purpose AI models placed on the market before the second tranche date separately from high-risk systems already placed on the market or put into service. It also gives a longer horizon for high-risk systems used by public authorities. The treatment can turn on whether the system undergoes significant changes in design after the relevant date, so the version and change history of your models is part of the analysis, not a footnote to it. Read Article 111 against your own deployment dates before concluding that a legacy system is out of scope.

How long do we have to keep the logs once the obligations apply?

Retention is set outside Article 12 itself: the provider-side duty in Article 19 and the deployer-side duty in Article 26(6) both speak to keeping automatically generated logs for a period appropriate to the intended purpose, with a floor expressed in months unless other Union or national law says otherwise. Sector rules for banking, insurance or medical devices frequently ask for longer, so most compliance officers set retention to the longest applicable period rather than the regulation's floor. What the retention period does not tell you is what each log entry has to contain, which is the Article 12 question. Our field-by-field walkthrough of Article 12 record-keeping covers that separately.

What should we be recording before the deadline rather than after it?

Records are retrospective by nature, so a log that starts on the application date cannot evidence the months before it. Auditrek is built to capture inference events as they happen, taking an input hash, model id and version, output, confidence, timestamp and human override flag into an append-only record chained with SHA-256 per AI system, and to recompute that chain server-side and return a tamper status rather than reporting a stored flag. That means the decision in front of you now is less about the deadline and more about when your first verifiable record exists. Starting earlier costs you an integration; starting later costs you a period you cannot reconstruct.

Can a tool tell us whether we meet the deadline?

No, and you should treat any tool that says otherwise with suspicion. Auditrek's report generator produces PDF and HTML output for a chosen system and date range, structured against the Article 12 and 13 fields, with one regulation and one format set at launch and a non-legal-advice disclaimer embedded in every report. That is documentation of what your system did and when, designed to support your own assessment and your counsel's; it is not an opinion on whether your practices satisfy an obligation. Auditrek is on a waitlist today, so the honest framing is what it is built to do, not what customers have proved with it.

Sources

Join the waitlist

Leave your work email and we will send the payload schema, the verification endpoint documentation and a sample Article 12 and 13 report before you talk to anyone. Pilot cohort conversations start with your own pipeline, not a demo script.

One email when it launches, and one now to confirm your address. Unsubscribe any time. Privacy.