Reviewed October 2026. The Act’s timeline is subject to the European Commission’s simplification proposals; the structure is not. Sources at the end.
Contents: Who is in scope · The four risk categories · The timeline · The literacy duty already in force · What each category requires · General-purpose AI and your suppliers · The UK position · Worked examples · A twelve-step plan · Common misreadings · Sources
The EU AI Act applies to UK businesses whose AI systems are placed on the EU market, used in the EU, or whose outputs are used in the EU. That covers most manufacturers selling into Europe, most businesses with EU customers, suppliers or staff, and any business whose software or service reaches EU users. For businesses entirely outside that scope the Act’s obligations do not apply directly, but its structure is the clearest available guide to what good AI governance looks like, and UK regulators point the same way. This guide sets out what it asks, of whom, and what to do.
Who is in scope
The Act defines roles, and your obligations depend on which you hold for each AI system.
| Role | Meaning | Typical UK example |
|---|---|---|
| Provider | Develops an AI system and places it on the EU market or puts it into service there under its own name | A manufacturer whose products include an AI function and sell in the EU; a software business with EU customers |
| Deployer | Uses an AI system under its own authority in the course of business | Almost every business whose staff use Copilot, Claude or ChatGPT; any business running an AI assistant or agent |
| Importer / distributor | Brings a third-country provider’s system into the EU, or supplies it there | Less common for services firms; relevant to product businesses |
Territorial scope is the key point for a UK business. The Act applies to providers and deployers outside the EU where the output of the system is used in the EU. A UK manufacturer whose predictive-maintenance model advises EU customers, a UK agency whose support agent answers EU shoppers, a UK publisher personalising content for EU readers: all in scope for those systems. A UK business with no EU market, customers, suppliers or staff is, for practical purposes, out of scope, and should still read on.
The four risk categories
The Act’s useful contribution is its structure. Every AI system falls into one of four categories, and obligations scale with the category.
Unacceptable risk: prohibited. Social scoring by public authorities, manipulative or deceptive techniques that cause harm, exploitation of vulnerabilities, untargeted scraping of facial images, emotion recognition in workplaces and education (with narrow exceptions), certain biometric categorisation and real-time remote biometric identification in public spaces. No ordinary business practice falls here; the one to watch is emotion recognition in the workplace, which rules out some HR and monitoring products.
High risk: regulated. Two routes in. First, AI that is a safety component of a product covered by existing EU product legislation, machinery, medical devices, vehicles, lifts, toys and others, which captures embedded AI in industrial products. Second, AI used in the areas listed in Annex III: biometrics, critical infrastructure, education and vocational training, employment and worker management, access to essential private and public services including credit and insurance, law enforcement, migration, and the administration of justice. For a UK business the common ones are recruitment and employee evaluation, creditworthiness and insurance pricing, admissions and assessment in education, and safety components in products.
Limited risk: transparency. Systems that interact with people, chatbots and assistants, must make clear that the person is interacting with AI. Generated or manipulated content must be disclosed in certain cases, deepfakes in particular. Emotion recognition and biometric categorisation, where permitted, require notice.
Minimal risk: no specific obligations. Everything else, which is the large majority of business use: drafting, summarising, search, analysis, forecasting, most internal automation. General law still applies, data protection above all.
A system’s category is decided by its use, not its technology. The same model is minimal risk drafting a newsletter and high risk screening CVs.
The timeline
| Date | What applied |
|---|---|
| 1 August 2024 | The Act entered into force |
| 2 February 2025 | Prohibitions and the AI literacy duty (Article 4) |
| 2 August 2025 | Obligations for general-purpose AI model providers; governance bodies operational; penalties applicable |
| 2 August 2026 | Most of the Act, including Annex III high-risk obligations and transparency duties |
| 2 August 2027 | High-risk obligations for AI embedded in regulated products (the product-safety route) |
The European Commission has proposed simplifications to parts of the Act, and some high-risk deadlines may move. The categories, the literacy duty and the transparency duties are settled; plan on the structure and check the dates.
The literacy duty already in force
Article 4 requires providers and deployers to take measures to ensure a sufficient level of AI literacy among their staff and other people operating AI systems on their behalf, taking account of their technical knowledge, experience, education and training, the context the systems are used in, and the people the systems affect. It has applied since 2 February 2025, and it applies to every deployer in scope regardless of risk category: a business whose staff use an enterprise assistant is deploying an AI system.
“Sufficient” is proportionate to role and risk. A leadership team deciding on AI investment needs to understand what the technology can and cannot do; a customer-service team using an assistant needs to know when to check its output and how to escalate; an engineer building an agent needs the technical understanding to make it safe; roles near high-risk uses need the specific obligations. One generic e-learning module for everyone does not meet the spirit of it. Keep a record of who was trained in what, and when: the record is the evidence.
What each category requires
Minimal risk. Nothing specific. Data protection, consumer law, employment law and sector regulation apply as they always did.
Limited risk. Disclosure. People interacting with an AI system must be told, unless it is obvious. Synthetic audio, image, video and text must be marked as generated in a machine-readable way by providers; deployers of deepfakes must disclose; deployers of AI-generated text on matters of public interest must disclose unless there is human editorial review. For a business this means: the assistant on the website says it is AI and offers a person; generated content is labelled where the audience would expect it.
High risk. For providers: a risk-management system across the lifecycle; data governance for training, validation and testing data; technical documentation; logging; transparency and instructions to deployers; human oversight designed in; accuracy, robustness and cybersecurity; a quality-management system; conformity assessment and CE marking; registration in the EU database; post-market monitoring and incident reporting. For deployers: use the system according to its instructions; assign human oversight to competent people; monitor; keep logs; inform workers and their representatives where the system affects them; in some cases a fundamental-rights impact assessment; inform people subject to decisions.
The deployer obligations are the ones most UK businesses will meet, because most will buy high-risk capability from a provider rather than build it. They are also the ones that map directly onto the four governance fields on a business case: risk category, data protection, human oversight, transparency.
General-purpose AI and your suppliers
Providers of general-purpose models, the foundation models behind Claude, ChatGPT, Copilot and the rest, carry their own obligations: documentation, copyright policy, training-data summaries, and more for models with systemic risk. These are the provider’s obligations, not yours. What matters to a deployer is that the provider’s enterprise terms give you what you need to meet your own obligations: documentation of capabilities and limits, no-training commitments for your data, and the logs and controls to run human oversight. Ask for them at contracting.
The UK position
The UK has no equivalent statute. Its approach is principles-based and sector-led: existing regulators, the ICO, the FCA, the CMA, the MHRA, Ofcom, apply five cross-sector principles, safety, transparency, fairness, accountability and contestability, within their remits. The ICO’s guidance on AI and data protection is the most developed and applies to any UK business processing personal data with AI. The Data (Use and Access) Act 2025 adjusted the rules on automated decision-making. A UK-only business therefore has no AI Act to comply with, but has data-protection law, sector regulation and, increasingly, customer and tender requirements that read like the Act. Treating the Act’s structure as the standard is the practical choice.
Worked examples
A York manufacturer selling drives into Germany. An on-device diagnostic function that affects safety is a high-risk safety component under the product route: provider obligations, from August 2027. A knowledge assistant for its engineers over standards and SOPs is minimal risk. A customer-facing troubleshooting assistant is limited risk: it discloses that it is AI. Its staff using an enterprise assistant triggers the literacy duty now.
A UK publisher with EU readers. Personalisation and recommendations are minimal risk under the Act, with data protection applying. An editorial assistant is minimal risk. AI-generated text published on matters of public interest requires disclosure unless there is human editorial review, which a publisher has. A reader-facing assistant is limited risk.
An ecommerce agency with EU shoppers on its clients’ stores. A support agent answering shoppers is limited risk: disclosure and a route to a person. The agency’s internal use of the same agent is minimal risk. If a client’s store uses AI to decide who gets credit at checkout, that is the client’s high-risk system, and the agency as a supplier should know it.
A charity screening grant applications with AI. Access to essential services and benefits is an Annex III area. If the system materially influences decisions about individuals’ access to support, it is high risk; if it only sorts and summarises for a person who decides, it is likely not, and the human-oversight design is what makes the difference. This is the kind of case to classify carefully and document.
A twelve-step plan
- Name an owner for AI policy and the Act: a director or a working group.
- Inventory every AI system in use or planned, including unofficial use, with its role (provider or deployer) and its EU exposure.
- Classify each system: prohibited, high, limited, minimal, by use.
- Add four fields to every AI business case: risk category, data protection, human oversight, transparency.
- Write the usage policy and the sanctioned-tool list, with the enterprise terms that protect data.
- Train proportionately, leadership, everyone with access, functions, engineers, roles near high risk, and record it. This is Article 4.
- Disclose: every customer-facing assistant says it is AI and offers a person; label generated content where expected.
- Design oversight into any agent: confidence scoring, thresholds, escalation, review.
- For anything high risk, obtain the provider’s documentation, assign competent oversight, keep logs, inform affected workers, and assess fundamental-rights impact where required.
- Check suppliers: enterprise terms, documentation, no-training commitments, what they will give you to meet your obligations.
- Monitor for drift, bias and incidents, and keep the record.
- Review annually, and when the Commission’s simplification changes the dates.
Most of this is what a well-run AI programme does anyway. The Act’s gift is a vocabulary that makes the governance conversation short: “limited risk, transparency duty” is a sentence a board can act on.
Common misreadings
“It does not apply to us, we are in the UK.” It applies where the output is used in the EU. Check the exposure per system before concluding.
“We should block AI tools until we understand the Act.” The Act asks for classification, oversight and literacy, not prohibition. A block pushes use into personal accounts the business cannot see, which satisfies nothing.
“Everything we do is high risk.” Almost nothing most businesses do is. The high-risk list is specific; read it.
“Literacy is a nice-to-have.” It has been a legal duty since February 2025 for every deployer in scope, and it is the cheapest obligation to meet.
“Our supplier is responsible.” For their obligations, yes. Deployer obligations are yours.
Sources
- EU AI Act, Article 4 (AI literacy)
- EU AI Act implementation timeline
- EU AI Act, Annex III (high-risk areas)
- ICO guidance on AI and data protection
- UK government: A pro-innovation approach to AI regulation
This guide is general information, not legal advice; for a specific system in or near the high-risk categories, take legal advice. Our AI governance service covers classification, the governed-access model and the literacy training; the governed access and AI literacy posts go deeper on two of the steps.