AI regulation for UK businesses: what applies (2026).
The UK has no AI Act, but it has rules. What UK GDPR’s automated-decision regime, the ICO, the FCA and the MHRA expect from an AI system, and what is still draft.
The short answer for October 2026: the UK has no AI Act, and the May 2026 King’s Speech didn’t include an AI bill. AI is regulated through existing law and sector regulators. For a UK business building or buying an AI system, four things matter most: the new automated-decision rules in UK GDPR, which took effect on 5 February 2026; the ICO, which is writing a statutory code on AI; your sector regulator, such as the FCA or the MHRA; and the EU AI Act wherever you serve EU customers.
This guide covers what each one asks of the system you build, what is still in draft, and how to check a supplier. It is a practical summary from a team that builds these systems, not legal advice. A qualified professional should confirm how the rules apply to your case.
Is there a UK AI Act?
No. The King’s Speech of May 2026 left out a standalone AI bill,, as the IAPP noted. The government’s approach is to let existing regulators apply their own rules to AI. Its AI Opportunities Action Plan, launched in January 2025, focuses on adoption and infrastructure. In its one-year report, published on 29 January 2026, the government said it had met its commitments on 38 of the 50 actions.
For a business, that means the rules you already follow (data protection, equality, consumer protection and your sector’s regulations) apply to AI as they do to anything else, and the regulators are getting more specific about how. Sector rules matter too, such as the golden thread record for higher-risk buildings in AI in construction. Adoption is moving quickly. The ONS found that about 35% of UK businesses with ten or more employees used at least one AI technology in June 2026, up from about 12% in September 2023, and that 28% of businesses with fewer than ten employees used it against 49% of those with 250 or more employees (ONS, July 2026).
Which regulator covers which part of an AI system?
There is no single owner. The same system can sit under several regimes at once, depending on what it does and whose data it touches.
| Regime | Covers | What it asks of the system |
|---|---|---|
| UK GDPR and the ICO | Any AI that uses personal data, including prompts, retrieval and logs | Lawful basis, transparency, impact assessment for high-risk processing, safeguards on significant automated decisions |
| FCA | Regulated financial firms | Good outcomes under the Consumer Duty, named senior accountability, no separate AI rulebook |
| MHRA | Software that is a medical device | Depends on intended purpose: drafting for a clinician is treated differently from supporting a diagnosis |
| Equality Act 2010 | Decisions about people, whether made by a person, software or both | Testing and monitoring outcomes across groups |
| EU AI Act | Systems placed on the EU market, or whose output is used in the EU, wherever the builder is based | Disclosure that users are talking to AI, and later, high-risk system duties |
The four regulators in the Digital Regulation Cooperation Forum (the CMA, Ofcom, the ICO and the FCA) work together on overlaps, and its March 2026 paper on agentic AI is a useful read on how they view systems that act on their own. If you work in a sector with its own regulator, check what that regulator has said as well.
What do the new automated-decision rules require?
The Data (Use and Access) Act 2025 replaced UK GDPR’s old Article 22 with new Articles 22A to 22D. They came into force on 5 February 2026, and the commencement regulations say they don’t apply to decisions taken before that date. In short:
- What counts. A decision is “solely automated” when there is no meaningful human involvement. It is “significant” when it has a legal effect, or a similarly significant effect, on the person.
- Where the restrictions bite. A significant, solely automated decision is restricted if it relies wholly or partly on special-category data (health, ethnicity and similar), unless a narrow exception applies, such as explicit consent. It is also barred where the processing relies on “recognised legitimate interests” (Article 6(1)(ea)).
- Safeguards are required. For significant, solely automated decisions based on personal data, people must be given information about the decision, be able to make representations, be able to obtain human intervention, and be able to contest it.
The Secretary of State can also make regulations that say whether there is meaningful human involvement in specified cases and add to the safeguards, so check the current text. The safeguards are design requirements, not paperwork. A system that approves, declines, prices or prioritises people needs to record why it decided, show that reasoning to a reviewer, and offer a working route to a person. Building that in from the start is far easier than retrofitting it after a complaint.
What is the ICO doing, and what is still draft?
The Information Commissioner’s Office is the main regulator for AI that uses personal data. Since May 2026 it has been required by law (SI 2026/425) to prepare a statutory code of practice on AI and automated decision-making, including children’s data. As of October 2026, the pieces in motion are:
- Automated decision-making guidance. The ICO’s updated guidance is still a draft. It was consulted on from 31 March to 29 May 2026, and no date is given for the final version. The ICO says it will withdraw the previous guidance once the update is finalised.
- Agentic AI. The ICO opened an agentic AI call for evidence on 8 October 2026, open until 20 November 2026. It covers security, transparency, accountability, automated decision-making, fairness and lawful processing. In announcing it, the ICO also said it had made enquiries about recent agent testing and that autonomy is not an excuse for poor compliance.
- The statutory code. The evidence from both exercises will feed the forthcoming code.
Until the final versions land, the ICO’s existing expectations hold. Have a documented lawful basis, be transparent with people, collect only the data you need, and be able to explain how the system reached an outcome. The ICO’s guidance says a data protection impact assessment is needed for high-risk processing, and its list of high-risk operations includes innovative technology such as AI when combined with other risk criteria, and decisions about access to a service based to any extent on automated decision-making.
Financial services: existing rules, applied to AI
The FCA says it does not plan to introduce extra regulations for AI. Its existing regime applies instead. The Consumer Duty asks firms to design products that meet the needs of their target customers and give fair value, to communicate in ways that meet customers’ information needs, and to provide support that meets their needs, and the FCA says its senior-manager accountability rules are relevant to the safe use of AI. For a lender, insurer or wealth firm, that means monitoring outcomes by customer group, keeping an audit trail of what the system did, and being able to show both. More on use cases in our guide to AI in financial services.
Healthcare: where the MHRA draws the line
The MHRA confirmed on 29 July 2026 that ambient voice tools used only to transcribe, summarise, draft letters or suggest codes for a clinician to review aren’t regulated as medical devices under the current framework. Tools that support diagnosis, treatment or prevention, or that take automated action such as placing orders without clinician review, are, and must meet the requirements that go with that. The guidance doesn’t change the law, and clinicians remain responsible for checking AI-generated outputs.
For a clinic or health-tech company, that boundary matters at the design stage. A system that drafts for a clinician to approve is a very different project, with a different timeline, from one that recommends treatment. See also AI in healthcare.
Does the Equality Act apply to AI decisions?
Yes. The ICO’s guidance notes that the Equality Act 2010 gives people protection from direct and indirect discrimination whether it is generated by a person or an automated system, and that showing a system is not unlawfully discriminatory is separate from meeting data protection duties. Complying with one does not guarantee the other. In practice, test outcomes across groups before launch, keep monitoring them, and keep records of what you found and changed.
What if you serve EU customers too?
If people in the EU use your system, the EU AI Act can apply to you wherever your company is based: Article 2 covers providers that place systems on the EU market and providers and deployers outside the EU whose system’s output is used in the EU. The Commission says Article 50 transparency duties apply from 2 August 2026: chatbots and other interactive systems must tell people they are dealing with AI, and deepfakes and AI-generated content must be labelled or marked. Providers of systems placed on the market before that date have until 2 December 2026 to meet the marking and detection duty in Article 50(2).
Personal data can keep flowing between the EU and the UK. On 19 December 2025 the European Commission renewed its UK adequacy decisions until 27 December 2031, with a review after four years. Many UK companies design to the stricter of the two regimes, so one system serves both markets.
What is coming next?
Dates change, so treat this as a watch list as of 25 October 2026:
- 20 November 2026: the ICO’s agentic AI call for evidence closes.
- 2 December 2026: the Article 50(2) marking deadline for existing EU-market systems.
- 2 December 2027: EU AI Act rules for high-risk systems in areas such as employment, education and biometrics apply, after the AI Omnibus simplification, with products such as lifts and toys on 2 August 2028 (European Commission).
- No date yet: final ICO guidance on automated decision-making, the statutory AI code, and any regulations from the Secretary of State on meaningful human involvement.
- Regulating for Growth Bill: announced in the King’s Speech to create sandbox powers, including for AI, and to strengthen regulators’ duty to support growth (Bird & Bird summary).
How do you check a third-party AI system before you rely on it?
Most mid-size businesses buy or build on top of a model rather than train one, so a lot of the risk sits with suppliers. Put these questions to every provider, and write down the answers:
- Where is personal data processed and stored, and does the provider offer UK or European hosting?
- Who is the controller and who is the processor for each stage, including any agents or tools that call other services?
- What is logged, for how long, and can you delete it? Logs often hold more personal data than the system itself.
- What happens when the model changes? You need notice of version changes and a way to re-run your evaluation set before they reach customers.
- What evidence exists of testing, including for bias and for attacks that hide instructions in inbound content?
The government’s AI Management Essentials tool is a self-assessment aimed at SMEs. It is built on ISO/IEC 42001, the NIST framework and the EU AI Act. It checks your management processes, not the products, and completing it isn’t proof of compliance, but it is a sensible starting checklist. Our buyer’s guide covers choosing a builder.
Where should a person stay in charge?
Keep a person responsible for any decision that affects someone’s credit, job, insurance, treatment or access to a service, and for anything the system is unsure about. The legal test turns on “meaningful human involvement”, and the ICO’s guidance on it is still a draft, so design for a reviewer who has the authority and the information to change the outcome, not just a button to approve what the system proposed. AI is the wrong tool for the final call on those decisions, and a good tool for gathering the facts, drafting the explanation and flagging the cases that need attention. Tenant and buyer screening is a typical case, covered in AI agents for real estate.
Name one owner for each AI system in your company, keep a simple register of what is in use, what data it touches and who signs it off, and review it when a system or a supplier changes. The FCA’s senior-manager model is a useful pattern even outside financial services.
What to build in from the start
| Build in | Why |
|---|---|
| A record of every decision and the data behind it | Explaining outcomes to people, the ICO or the FCA |
| A working route to a person | Article 22C safeguards; the Consumer Duty |
| Clear notices that AI is involved | UK GDPR transparency; EU Article 50 if you serve EU users |
| A data protection impact assessment before launch | ICO expectation for high-risk processing, including AI-assisted decisions about access to services |
| Data minimisation in prompts, retrieval and logs | UK GDPR; logs often hold more personal data than the system itself |
| Outcome monitoring by customer group | The Equality Act and the Consumer Duty |
| An evaluation set run before every release | Evidence that the system does what you claim (see evals before features) |
How do you start, and what do you track?
- List what you run. Every AI system, who owns it, what personal data it uses and whether it makes or shapes decisions about people.
- Sort by effect. Mark any system whose output could be a significant decision. Those get the full safeguards; assistants that draft for review get lighter ones.
- Run the impact assessment for the high-risk ones, with legal or data protection advice.
- Build the route to a person and the decision record before launch, then test both.
- Review quarterly, and after each ICO or sector-regulator update.
Track a handful of numbers: the share of significant decisions with a complete record; time to a human response when someone contests a decision; the share of high-risk systems with a current impact assessment; the rate at which reviewers change the system’s proposal (near zero can mean rubber-stamping); and outcome gaps between customer groups.
If you want help scoping this, our services cover building these systems with the checks in place.
Questions UK businesses ask
Do we need to register our AI system with anyone?
There is no UK AI Act that sets up a registration duty, because the May 2026 King’s Speech included no AI bill. Sector rules still apply, such as the MHRA’s for health tools that count as medical devices.
Can we use US model providers with UK personal data?
Yes, if each restricted transfer is covered by adequacy regulations or an appropriate safeguard (with a transfer risk assessment), and with a data protection impact assessment where the processing is high-risk. Check where each provider processes data and whether it offers UK or European hosting. A qualified professional should confirm the mechanism for your case.
Does the Equality Act apply to AI decisions?
Yes. A decision doesn’t become exempt from discrimination law because software made it. Test outcomes across groups before launch and keep monitoring them after.
Is an AI assistant that drafts replies making an automated decision?
Usually not, if a person genuinely reviews and can change what it drafts and the output has no significant effect on someone by itself. If the person only clicks approve, regulators may treat it as solely automated.
Does the EU AI Act apply to a UK company?
It can, wherever you are based, if you place a system on the EU market or its output is used in the EU. Interactive systems must tell users they are dealing with AI from 2 August 2026, and the high-risk rules follow in December 2027.
Modulus Labs AI builds AI agents, assistants and automation for UK and international companies, with decision records, human review and monitoring built in. Tell us what you’re building, and we’ll tell you what we’d build first.