Which AI systems
the Act actually touches.

Under Article 3(3), a provider develops an AI system or a general-purpose AI model, or has one developed, and places it on the market or puts it into service under its own name or trademark. Under Article 3(4), a deployer is anyone using an AI system under their authority, other than in the course of a personal, non-professional activity.

Read that second definition again. It has no threshold. A law firm running a contract-review tool is a deployer. A distributor using demand forecasting is a deployer. A recruitment team using a CV-screening product bought from a vendor is a deployer. There is no minimum spend, no minimum headcount, no carve-out for “we just use the off-the-shelf version”.

Article 25 sits in Chapter III, Section 3. Under the Digital Omnibus it applies from 2 December 2027 for Annex III high-risk systems and from 2 August 2028 for Annex I high-risk systems. Read the timing carefully, because it is not a reason to defer the question.

Article 25(1) converts a deployer into a provider — with the full provider obligation set — in three circumstances: you put your own name or trademark on a high-risk system already on the market; you substantially modify such a system; or you modify the intended purpose of a system, including a general-purpose AI system, in a way that makes it high-risk under Article 6. When that happens, the original provider steps out of the provider role for that system — though under Article 25(2) it must still cooperate closely with you, make the necessary information available and provide reasonably expected technical access. That is not a clean discharge, and it is not a substitute for your own record.

Here is why the deferral does not help. The conversion is triggered by what you build, brand and repurpose — and firms are doing that now, in sprint planning sessions, without a written record of having crossed the line. Wrapping a general-purpose model in your own product, giving it your brand, and pointing it at a hiring or credit decision is not procurement. It is, in law, becoming a provider. The obligations attach in December 2027. The act that attracts them happens this quarter, and by 2027 nobody will remember which release it was.

Article 2(1) applies the Regulation to providers placing AI systems on the Union market “irrespective of whether those providers are established or located within the Union or in a third country”; to deployers established or located in the Union; and — the limb that catches British firms — to providers and deployers established in a third country “where the output produced by the AI system is used in the Union”. Establishment is not the test. Output is. A London company with no EU entity, running a model on UK infrastructure, is within scope the moment the output of that system is used in the Union — in a report to an EU client, a decision affecting an EU customer, a screening outcome for an EU applicant. Brexit did not move this boundary, because the boundary was never drawn at the border.

Article 99 sets three tiers. For infringement of the Article 5 prohibitions: administrative fines of up to EUR 35 million or 7% of total worldwide annual turnover, whichever is higher for an undertaking. For most other breaches — including the deployer obligations under Article 26 and the transparency obligations under Article 50 — up to EUR 15 million or 3%. For supplying incorrect, incomplete or misleading information to notified bodies or national competent authorities, up to EUR 7.5 million or 1%.

Two moderating provisions matter, and they are not the same provision. The one most readers of this page fall under is Article 99(6): for SMEs, including start-ups, the fine is the lower of the percentage or the amount, and it applies across all three tiers. The Digital Omnibus then extended that treatment through Article 99(6a): the fine caps that applied to SMEs and start-ups — the lower of the percentage or the fixed amount — now reach small mid-cap enterprises as well. If your organisation sits in that bracket, it is worth establishing on paper which of your exposures the cap actually reaches, rather than assuming it covers all of them.

We do not lead with these numbers, and you should be sceptical of anyone who does. Fines are the tail of the distribution. The near-certain cost is elsewhere.

Every obligation in this Regulation resolves, in practice, into a question of proof. Can you show which systems you run? Can you show which of them are high-risk, and on what reasoning?

Then there is the set of questions the deferral has scheduled rather than cancelled. From 2 December 2027 for Annex III systems, and 2 August 2028 for Annex I, Article 26 will ask a deployer of a high-risk system: who was assigned oversight, and did they have the competence, training and authority to exercise it, as Article 26(2) requires? Can you produce the logs Article 26(6) requires you to retain for at least six months? Can you show workers and their representatives were informed before deployment, as Article 26(7) requires?

None of that can be manufactured after the question is asked. Log retention is not a policy you adopt in November 2027 — it is a system configuration that either was or was not set years earlier, and the six months of history the Act asks for must already exist on the day the duty bites. Oversight assignments and workforce notifications are the same: they are records of things done at the time, or they are nothing. The organisations that will struggle are not the reckless ones. They are the ones that did everything sensibly and wrote none of it down.

There is a second-order exposure that boards notice faster than regulators do, and it does not wait for 2027: your customers’ procurement teams. EU-based buyers are already putting AI Act representations into contracts. A firm that cannot answer a supplier questionnaire loses the deal long before it meets an authority.

Yes, as a deployer. Article 3(4) defines a deployer as anyone using an AI system under their authority outside personal, non-professional use — there is no size or spend threshold.

Two obligations already bind you. Article 4 literacy, which has applied since 2 February 2025. And the deployer limbs of Article 50, live since the general application date of 2 August 2026: Article 50(3) requires you to inform people exposed to emotion recognition or biometric categorisation, and Article 50(4) requires you to disclose deepfake content, and to disclose AI-generated text where it is published with the purpose of informing the public on matters of public interest — not all published AI-assisted text, and not text that has gone through human review or editorial control where a person or company holds editorial responsibility for the publication. Routine marketing copy and client documents will usually sit outside that duty. It is worth knowing which side of the line yours sits on, rather than assuming.

The third thing is not yet a duty, but it is a decision you are taking now: Article 25. If anyone in your organisation brands, substantially modifies or repurposes one of those tools in a way that would make it high-risk, the provider obligations attach when Chapter III, Section 3 applies — 2 December 2027 for Annex III systems, 2 August 2028 for Annex I. The conversion is caused by what you ship now. Being a deployer is a light obligation set. It is not an empty one, and it is not permanent.

Yes, on the Act’s own terms. Article 2(1)(c) applies the Regulation to providers and deployers established in a third country “where the output produced by the AI system is used in the Union”. Article 2(1)(a) applies it to providers placing systems on the Union market regardless of where they are established. Neither limb depends on you having an EU entity, EU infrastructure or EU staff. The practical test is where the output of the system is used — and for most UK B2B firms with European clients, at least some of it is.

No, and it is not the vendor’s to give. Provider obligations and deployer obligations are separate sets held by separate parties, and no contract moves them.

Article 26 applies to you only as a deployer of a high-risk system, and only from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I. When it applies, it will require you to use the system in accordance with the provider’s instructions for use, to assign human oversight to people with the necessary competence, training and authority, to ensure input data you control is relevant and sufficiently representative, to monitor operation and suspend use where a risk emerges, to retain logs for at least six months, and to inform workers’ representatives and affected workers before the system goes live in the workplace.

Two consequences follow. If none of your systems is high-risk, Article 26 never reaches you — which is worth establishing on paper rather than hoping. If any of them is, the evidence those duties call for has to be accumulating well before the date, and the contract you sign with the vendor this year determines whether you can produce it. A vendor’s assurance is useful evidence. It is not a transfer of responsibility.

Almost certainly not directly. The Code of Practice, published by the Commission on 10 July 2025, is a voluntary instrument for providers of general-purpose AI models, structured in three chapters — transparency, copyright, and safety and security — addressing obligations under Articles 53 and 55. If you use models rather than place them on the market, you are not a signatory and cannot become one. It is still worth reading: the transparency chapter’s model documentation form tells you exactly what your model provider should be able to give you, which makes it a useful procurement instrument even though it imposes nothing on you.

On 2 December 2027, Chapter III Sections 1–3 apply to systems classified high-risk under Article 6(2) and Annex III: the requirements on risk management, data governance, technical documentation, logging, transparency, human oversight, accuracy and robustness, together with the provider and deployer obligations that attach to them — Articles 16 to 27, including the Article 25 role conversion, the Article 26 deployer duties and the Article 27 fundamental rights impact assessment. Systems high-risk under Article 6(1) and Annex I follow on 2 August 2028. Nothing happens automatically on either date — the question is whether, by then, you can show which of your systems fall inside Annex III and which do not, and whether the records those articles will ask for have been accumulating in the meantime.

2 December 2026 is likely to be more immediately relevant. Two new Article 5 prohibitions take effect that day, at points (ba) and (bb) of Article 5(1), covering AI systems that generate or manipulate non-consensual intimate material and child sexual abuse material, with the provider and deployer conditions set out in the new Article 5(1a). Firms that build or fine-tune generative image or video capability should be reading those provisions now, not in November. The same date closes the four-month transitional window for Article 50(2) synthetic-content marking for generative systems already on the market before 2 August 2026.