Boards keep asking when the UK will pass its AI Act, as if the obligations arrive with the statute. Britain has chosen a different route, and the effect is that your AI systems are already regulated today, by the same bodies that regulate the rest of your business.
The short answer
- The UK has no horizontal AI law and, on current policy, is not planning one. It regulates AI through existing sector regulators applying existing law.
- A 2023 white paper set five cross-sectoral principles. The regulators, not a new AI regulator, apply them: the ICO for data protection, the FCA for financial conduct, Ofcom, Ofgem, the CMA and others in their patches.
- Your live obligations already sit in data protection, equality, consumer, product safety and financial services law. None of them waited for an AI Act.
- If you place AI systems or their outputs on the EU market, the EU AI Act reaches you regardless of where you are based.
- The board task is to map each material AI use to the regulator and the law that already bite on it, rather than to scan the horizon for a bill.
What Britain decided to do instead of an Act
In March 2023 the government published A pro-innovation approach to AI regulation. It set out five principles that regulators are expected to apply within their existing remits: safety, security and robustness; appropriate transparency and explainability; fairness; accountability and governance; and contestability and redress. The principles are guidance rather than statute, a common frame that each regulator interprets for its own sector.
The consultation response in February 2024 confirmed the model and asked the main regulators to publish their strategic approach to AI by 30 April 2024. The ICO, the FCA, Ofcom, the CMA, Ofgem and others did so, each describing how the five principles map onto the risks in its sector and what it will do with the powers it already holds. That request, and those publications, are the closest thing the UK has to an AI rulebook, and they point back into existing legislation rather than forward to a new one.
The direction has held since. The government has continued to favour regulation at the point of use by expert regulators, and its more recent legislative work has concentrated on structures such as statutory regulatory sandboxes rather than a single AI statute. For a board, the practical reading is that the obligations are distributed across the regulators you already answer to, and are in force now.
Which regulator already governs your AI
Most enterprise AI touches at least one regulated activity. The system is governed by whichever regime that activity already falls under.
| If your AI does this | The regime that already applies | The regulator |
|---|---|---|
| Processes personal data, including for training or automated decisions | UK GDPR and the Data Protection Act 2018 | ICO |
| Informs lending, pricing, advice or a financial decision | FCA rules and the Consumer Duty | FCA, with the Bank of England and PRA for prudential firms |
| Affects a hiring, promotion or service decision about a person | Equality Act 2010 | EHRC, and the courts |
| Shapes what a consumer sees, is offered or is charged | Consumer protection and, for large platforms, digital markets rules | CMA |
| Operates inside a communications or media service | Communications and online safety regimes | Ofcom |
| Sits in an essential service such as energy, water or transport | Sector safety and security duties, including the NIS regulations | The sector competent authority, for example Ofgem |
Read that table as a routing exercise. A recruitment screening model is an Equality Act and data protection matter before it is anything else. A credit model is an FCA and data protection matter. A demand-forecasting model on a distribution network is a safety, security and data protection matter under the sector regime. In each case the question a board should ask is not what a future AI law will require, but what the existing regulator already requires of a decision made this way.
The EU AI Act still reaches you
Britain not passing a law does not put a UK company outside the EU one. The EU AI Act entered into force in 2024 and phases in through 2026 and 2027. Its scope is defined by market, not by geography: a provider or deployer whose AI systems, or whose AI-generated output, are placed on or used in the EU market is caught wherever it is based. A UK insurer serving EU customers, a UK software firm selling an AI feature into the EU, or a group with EU subsidiaries should assume the Act applies to those activities and should classify the relevant systems against its risk tiers on the EU timeline.
This is the part boards most often miss. The mental model of "no UK Act, so nothing yet" is wrong twice over: the domestic obligations are already live under sector regulators, and the EU obligations arrive on a fixed calendar for anything touching the EU market.
What good governance looks like without a statute
The absence of a single Act removes the temptation to treat compliance as one project with one deadline. It puts the work where it belongs, inside each function that owns a regulated decision.
Keep an inventory that names the regime, not just the system
A register of AI uses is common. A register that records, against each use, which regulator and which existing law govern it is rarer and far more useful. It converts an abstract worry about AI regulation into a list of specific, testable obligations that already have owners.
Assign each material use to the accountable executive for that regime
Data protection accountability already sits with a named owner. So does financial conduct, and product safety. AI lands inside the accountabilities you already have rather than creating a new one. Where a use spans two regimes, say a model that makes a pricing decision using personal data, both owners are engaged rather than a new committee invented to sit above them.
Adopt a recognised standard as the common control set
Because the five principles are deliberately high level, many boards adopt an external framework to make them operational. ISO/IEC 42001 and the NIST AI Risk Management Framework both give a structured control set that satisfies the principles and travels across regimes, which matters when the same system answers to two regulators.
The constraint
For most boards the binding limit is the absence of a mapping from each material AI use to the regulator and the existing law that already governs it. Commissioning horizon-scanning for a UK AI Act, or waiting for one, spends attention without reducing exposure, because the obligations that can be enforced against the company today already exist in data protection, financial, equality and sector law, and the decision that has to be taken is who owns each of them.
The board question that replaces "when is the Act coming"
A more useful standing question at the risk committee is this: for each AI system that informs a regulated decision, who is the accountable executive, which regulator governs it, and can we evidence that the existing obligations are met. A system that cannot answer those three is exposed now, under law already in force, whatever happens to future legislation. A company that can answer them for its material uses is compliant in the only sense that is enforceable, and is also ready for a UK Act or an EU deadline should either change the picture.
Sources
- Department for Science, Innovation and Technology, A pro-innovation approach to AI regulation, March 2023.
- DSIT, A pro-innovation approach to AI regulation: government response, February 2024, including the request that regulators publish their strategic approach to AI by 30 April 2024.
- Information Commissioner's Office, guidance on AI and data protection.
- Regulation (EU) 2024/1689 (the EU AI Act), in force from 1 August 2024, phasing in through 2026 and 2027.