Regulation
The EU AI Act for UK organisations: scope, timing and what to do
When the EU AI Act reaches a UK organisation, what the deferred 2027 and 2028 deadlines change, and why the Article 5 prohibitions already bind you.
Two things about the EU AI Act are widely misunderstood in the UK. The first is that Brexit removed the obligation. It did not. The second is that the 2026 delay bought everyone time. It bought some people time, on some obligations, and none at all on the ones with the largest fines.
Here is the position as at September 2026.
When it reaches you
The Act applies extraterritorially. A UK-established organisation is in scope if it:
- places an AI system on the EU market, whether directly or through a distributor or reseller;
- puts an AI system into service in the EU; or
- produces AI output that is used in the EU, which is the limb that catches people, because it does not require you to sell anything in the EU at all.
That last one deserves care. If your model scores applicants and some of them are in the EU, or your assistant serves EU-based customers of a UK product, you are likely in scope regardless of where your servers, your staff or your company registration sit.
Where you land in the obligations depends on your role: provider (you develop it and put it on the market), deployer (you use it), importer or distributor. Providers carry the heaviest load by a wide margin. Many UK organisations assume they are only deployers, then discover that materially fine-tuning or rebranding a model can make them a provider.
The timeline, after the delay
Following the EU AI Act Omnibus (provisional political agreement reached on 7 May 2026), the high-risk deadlines were deferred:
| Obligation | Status |
|---|---|
| Article 5 prohibited practices | In force and enforceable now |
| General-purpose AI model obligations | In application |
| Annex III high-risk: recruitment, credit scoring, education, law enforcement, essential public and private services | 2 December 2027 |
| Annex I high-risk: AI as a safety component of regulated products | 2 August 2028 |
The deferral applies to the high-risk regime only. It does not touch the prohibitions.
The prohibitions are the urgent part
Article 5 bans a defined set of practices outright. Breach carries the Act’s heaviest penalty: up to €35 million or 7% of global annual turnover, whichever is greater. That is a larger maximum than GDPR.
The prohibited categories include social scoring, manipulative or exploitative techniques that materially distort behaviour, certain emotion inference in workplace and education settings, untargeted scraping of facial images to build recognition databases, and biometric categorisation inferring sensitive attributes such as race, political opinion, religion or sexual orientation.
Most organisations reading this are not deliberately doing any of that. The exposure is almost never in what you built. It is in what you bought. Emotion analysis quietly bundled into a contact-centre platform. Candidate scoring in a recruitment tool that infers more than it should. Workforce analytics that categorise people in ways nobody reviewed.
The single highest-value hour of AI compliance work available to a UK organisation right now is to list every third-party tool with an AI feature and check it against Article 5. Not the tools you built. The ones that arrived in a subscription.
What to do before December 2027
If you have systems that will fall into Annex III, the deadline sounds distant and is not. High-risk obligations require artefacts that have to be built alongside the system, not bolted on:
- A risk management system running across the lifecycle, not a one-off assessment.
- Data governance covering training, validation and test data, including examination for bias, and documented provenance.
- Technical documentation sufficient for an authority to assess conformity.
- Automatic logging with retention appropriate to the purpose.
- Transparency and instructions for use enabling deployers to use the system correctly.
- Human oversight designed in, and genuinely capable of preventing or mitigating harm.
- Accuracy, robustness and cybersecurity appropriate to the intended purpose, and stated.
Reconstructing training-data provenance for a model built eighteen months ago, by people who have since left, is the kind of task that quietly turns into a rebuild. If a system is likely to be Annex III, the documentation obligation should shape the build now.
The interaction with UK obligations
UK organisations in EU scope face a dual picture: the UK’s principles-based, regulator-led framework domestically, and the EU’s prescriptive regime for anything touching the EU market.
This is more tractable than it sounds, because the underlying evidence overlaps heavily. A properly conducted DPIA, an accurate AI inventory, a real risk register, documented human oversight and an evaluation record satisfy the bulk of what both regimes want to see. The difference is in formality and thresholds, not in kind.
The practical mistake is running two separate compliance programmes. Build one evidence base, structured to the stricter requirement, and map it to each regime.
A proportionate sequence
1. Determine scope, in writing. For each AI system: are any outputs used in the EU? Are you provider or deployer? Record the reasoning, not just the conclusion; the reasoning is what you will need when someone asks in two years.
2. Screen everything against Article 5. Third-party tools first. This is the live liability.
3. Classify against Annex III. Recruitment, credit, education, essential services and law enforcement are the categories that catch commercial organisations. If a system is close to the line, treat it as in scope; the cost of over-preparing is documentation you would benefit from anyway.
4. Check your GPAI position. If you fine-tune or substantially modify a general-purpose model, you may inherit provider obligations you did not expect.
5. Build the evidence alongside the system. For anything heading for Annex III, the documentation is not a phase at the end. It is a property of how the thing is built.
The reasonable summary
For most UK organisations, the EU AI Act is not an immediate crisis, and the delay to December 2027 is real relief for high-risk systems.
But it is not permission to stop, for two reasons. The prohibitions bind now, at the highest penalty tier, and the exposure is in procured tooling that nobody has reviewed. And the high-risk documentation burden is structural: it changes how a system must be built, which means the deadline that matters is the start of your next build, not the end of 2027.
General information as at September 2026, not legal advice; the position continues to move. For obligations specific to your systems, consult a qualified solicitor.
See also: UK AI regulation in 2026. If you need the evidence base built, that is our AI governance and compliance service.