https://www.noxia.co.uk/field-notes/the-ai-deadline-that-moved · printed from noxia.co.uk · sources checked 24 September 2026
Field note · Buying AI
The AI deadline moved to December 2027. What it does not buy you.
Every compliance calendar in Europe had 2 August 2026 on it. In May the institutions agreed to move it by sixteen months, and in July it became law. The right response for a UK firm is not relief, because the obligation that was most likely to touch you is one of the ones that stayed exactly where it was.
- Buying AI
- EU AI Act
- Vendor risk
- Regulation
The short answer
Agreed on 7 May 2026 and adopted as Regulation (EU) 2026/1744, in force since 27 July 2026, the Digital Omnibus postpones key AI Act deadlines. Obligations for Annex III high-risk systems move from 2 August 2026 to 2 December 2027; Annex I high-risk systems move from 2 August 2027 to 2 August 2028; regulatory sandboxes move to 2 August 2027. The Article 50 transparency obligations did not move and still apply from 2 August 2026, with a grace period to 2 December 2026 for watermarking systems already on the market.
On this page · 5 sections
A delay is the easiest news in regulation to misread, because the headline is a date and the substance is a list. Here is the list.
Why a UK broker should read this at all
Three reasons, in descending order of likelihood.
Your vendor is in scope even if you are not. The AI Act binds providers placing systems on the EU market. A UK firm buying a tool from a European provider, or from an American one that sells into Europe, is buying a product whose roadmap is now shaped by December 2027 rather than August 2026. Features you were told would arrive with the compliance release have just been given sixteen more months of runway. Ask.
Some of you do sell into the EU. A UK firm with Irish clients, or a product with European users, can be a deployer of a high-risk system under the Act. Recruitment screening and credit scoring are both named Annex III uses, and both are things small professional firms now do with software they did not write.
And the transparency rules are the ones that bite in public. Telling a person they are talking to a machine, and labelling synthetic content, are the obligations a customer can observe. They did not move.
What we would actually do with sixteen months
Not wait. The classification work — deciding which of your systems, if any, are high-risk, and writing down why — is the part that takes calendar time rather than effort, because it needs decisions from people who are busy. It also does not expire. A system inventory written this autumn is still correct in 2027, whereas a compliance sprint started in October 2027 will discover in week two that nobody knows which model the recruitment tool uses.
- List the systems, not the vendors. One vendor can supply three systems with three different risk profiles. The stack check exists because most firms cannot answer "what is actually running" without looking.
- Write down the use, in a sentence, for each. Annex III is a list of uses, not of technologies. "Ranks job applicants" is in scope; "drafts a job advert" is not.
- Ask each provider, in writing, what changed for them when the Omnibus became law. The answer tells you whether they have a compliance function or a marketing page.
- Do the transparency work now, because it is small and it is due. If a customer can talk to something of yours, it should say what it is.
The status, stated honestly
It is law. The provisional agreement reached in the early hours of 7 May 2026 and confirmed by Member State representatives on 13 May was adopted as Regulation (EU) 2026/1744 of 8 July 2026, published in the Official Journal on 24 July and in force since 27 July 2026, a week before the 2 August deadline it was designed to relieve. The dates above are the regulation’s.
This note was first published describing it as a provisional agreement, from a law firm’s reading in May, two months after it had become law — which is the kind of thing the register exists to record, and it does.
If you want the general version of this argument — that the thing to buy is not the model but the record of what it did — it is in the audit trail note, and the commercial version is in how AI vendors price. Our own view of which claims from AI vendors survive checking is the same exercise on five published AI claims.
Questions people actually ask
Has the EU AI Act been delayed?
Parts of it. Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026, moves the Annex III high-risk obligations from 2 August 2026 to 2 December 2027, and the Annex I high-risk obligations from 2 August 2027 to 2 August 2028. Regulatory sandboxes moved to 2 August 2027. Obligations already in force since 2025, and the Article 50 transparency duties, were not delayed.
Do the AI Act transparency rules still apply from August 2026?
Yes. The Article 50 transparency and disclosure obligations — including telling people they are interacting with an AI system and disclosing synthetic content — still apply from 2 August 2026. A four-month grace period to 2 December 2026 applies to the watermarking requirement for systems placed on the market before that date.
Does the EU AI Act apply to a UK company?
It can. The Act binds providers placing AI systems on the EU market and deployers established in the EU, and it can reach a UK firm whose system output is used in the EU. More commonly, it reaches UK firms indirectly: the vendor you buy from is in scope, and its product roadmap and contractual terms follow the Act’s dates rather than yours.
What counts as a high-risk AI system?
Annex III lists uses rather than technologies — including employment and worker management such as ranking job applicants, creditworthiness assessment, education, law enforcement and border control. Annex I covers AI as a safety component of products already regulated under EU product legislation, such as machinery and medical devices. The test is what the system is used for.
Sources
- Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI); in force 27 July 2026. eur-lex.europa.eu ↗ — primary; the regulation itself
- European Commission, “AI Omnibus enters into force”, 27 July 2026: “On 27 July 2026, the AI Omnibus enters into force across the EU”; rules for high-risk AI systems in Annex III apply from 2 December 2027, and for high-risk AI in products under Annex I from 2 August 2028. digital-strategy.ec.europa.eu ↗ — primary; the Commission
- Council of the EU, “Artificial Intelligence: Council and Parliament agree to simplify and streamline rules”, 7 May 2026: “Today, the Council presidency and European Parliament negotiators reached a provisional agreement”. consilium.europa.eu ↗ — primary; the Council
- Gibson Dunn, “EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes”: the provisional agreement, which it dates 6 May 2026 (the Council’s release is of 7 May), confirmed by Member State representatives on 13 May. Annex III high-risk from 2 August 2026 to 2 December 2027; Annex I high-risk from 2 August 2027 to 2 August 2028; AI regulatory sandboxes from 2 August 2026 to 2 August 2027; a new grace period to 2 December 2026 for the Article 50(2) watermarking requirement. Formal adoption and Official Journal publication expected before 2 August 2026, which happened. gibsondunn.com ↗ — secondary; a law firm’s reading of the provisional agreement, now superseded by the regulation above
- EU Artificial Intelligence Act implementation timeline, as published at artificialintelligenceact.eu and read on 18 September 2026: 2 February 2025 prohibitions and AI literacy; 2 August 2025 general-purpose AI model rules, governance and confidentiality; 2 August 2026 the remainder of the Act begins to apply; systems generating synthetic audio, image, video or text that “have been placed on the market before this date need to be compliant with Article 50(2) by 2 December 2026”; 2 December 2027 Annex III high-risk; 2 August 2028 Annex I high-risk. artificialintelligenceact.eu ↗ — secondary; an independent tracker, useful because it shows the dates as they now stand rather than as they were agreed. Read on one date and reported as of that date re-checked yearly
- The three reasons a UK firm should care, the four things to do with the extra time, and the observation that classification work does not expire are ours. — our own argument, labelled as such
Checked 24 September 2026. Next scheduled check 24 September 2027. Numbers that move — leaderboards, live indices — are re-checked every 30 days; annual datasets and rules in force every six months; dated research once a year. If something here has gone stale before we got to it, tell us and we will correct it and say what changed.
Cite this note
Noxia, “The AI deadline moved to December 2027. What it does not buy you”, Field notes, 18 September 2026; sources checked 24 September 2026. https://www.noxia.co.uk/field-notes/the-ai-deadline-that-moved
Ask your vendor what changed in July. We will read the answer.
Most AI procurement questions are answered by a document nobody reads closely. We read them: what the system does, what it is classified as, what the provider has committed to and by when, and what falls to you as deployer. It is a short engagement and it usually ends with two emails you should send.
Talk to us about thisRead next
Advice & compliance
In a regulated firm, the audit trail is the product.
An unregulated business buys the output. You are buying the ability to explain the output eighteen months later to someone who is not impressed.
What we found
We checked five AI statistics we were about to publish. Three had no source.
One was on our own homepage, credited to Gartner. We first said Gartner never published it. Gartner had — as something smaller than we claimed. What happened when we looked, twice.
Buying AI
Per-minute AI pricing charges you for the thing working.
A price that rises with volume charges you for success. Worse, it makes the quote uncomparable against the only alternative you were seriously considering — a person.