UK AI Regulation Guide: The Principles-Based Approach

The UK has no single AI law, so organisations deploying AI in the UK have to piece their obligations together from regulator guidance, data protection law, voluntary codes and sector rules. Since the 2023 white paper, the government's model has been five cross-sectoral principles applied by existing regulators rather than a new AI regulator or an EU-style horizontal act. That model now sits alongside binding changes to automated decision-making under the Data (Use and Access) Act 2025, a voluntary AI Cyber Security Code of Practice that has become a European standard (ETSI EN 304 223), and an AI Security Institute focused on serious security risks. This guide sets out what is in force, what is voluntary and what is only proposed, for security, privacy and risk leads and executives.

Key takeaways

• Principles, not a statute. The five white paper principles (safety, security and robustness; transparency and explainability; fairness; accountability and governance; contestability and redress) are non-statutory and applied by existing regulators within their remits.
• The binding law is mostly data protection. UK GDPR Articles 22A to 22D, inserted by section 80 of the Data (Use and Access) Act 2025 and fully in force since 5 February 2026, govern significant decisions made solely by automated processing.
• The security baseline is voluntary but specific. DSIT's Code of Practice for the Cyber Security of AI (31 January 2025) sets 13 principles across five lifecycle phases, and ETSI EN 304 223 (January 2026) turns them into a European standard.
• No government AI bill exists yet. As at October 2026 the UK Parliament bills list shows only private members' bills with "Artificial Intelligence" in the title, and commentators do not expect a comprehensive government bill before late 2027.
• AI agents are covered indirectly. An agent that makes significant decisions about people, holds credentials or processes personal data falls under UK GDPR, sector rules and the Code's identity and asset principles, even though no UK law names agents.

Scope: who and what is covered

Because the UK approach is principles applied through existing regimes, "scope" depends on which regime applies. The table below maps the main instruments, their legal status and who they bind.

In May 2024 DSIT published the strategic AI updates of 13 regulators: the Bank of England, CMA, EHRC, FCA, HSE, ICO, Legal Services Board, MHRA, Office for Nuclear Regulation, Ofsted, Ofcom, Ofgem and Ofqual. If your organisation is supervised by one of them, its update is the most direct statement of how the white paper principles will be applied to you. Ofcom, the CMA, the FCA and the ICO also cooperate through the Digital Regulation Cooperation Forum.

Two practical scoping points follow. First, there is no UK equivalent of the EU AI Act's risk tiers, so "high risk" is whatever your sector regulator and UK GDPR treat as high impact. Second, organisations that also place AI on the EU market must meet the EU regime separately. See the EU AI Act Compliance Guide and the Global AI Regulation Map.

Key obligations

The five principles, translated into controls

Regulators interpret the principles in their own context, but each one maps to evidence a security or risk team can produce.

Automated decision-making under UK GDPR (Articles 22A to 22D)

Section 80 of the Data (Use and Access) Act 2025 replaced the old Article 22 with four new articles. These are the binding core of UK AI regulation for most organisations.

• Definitions (Article 22A): a decision is based solely on automated processing where there is no meaningful human involvement. A decision is "significant" if it produces a legal or similarly significant effect for the data subject. The extent to which the decision relies on profiling must be considered when judging human involvement.
• Restrictions (Article 22B): a significant decision based wholly or partly on special category data cannot be made solely by automated processing unless the data subject has given explicit consent, or the decision is necessary for a contract or required or authorised by law and Article 9(2)(g) applies. A significant solely automated decision also cannot rely on the new recognised legitimate interests basis in Article 6(1)(ea).
• Safeguards (Article 22C): controllers making significant solely automated decisions must have safeguards that let the person receive information about the decision, make representations, obtain human intervention and contest the decision.
• Further provision (Article 22D): the Secretary of State may make regulations on what counts as meaningful human involvement and similarly significant effect, and may add safeguards, but cannot amend Article 22C.

The ICO summarises the effect: organisations can rely on a wider range of lawful bases, potentially including legitimate interests, for significant automated decisions, as long as safeguards apply, but the wider options do not extend to special category data. The same Act requires controllers to facilitate data protection complaints, for example with an electronic form, and to acknowledge them within 30 days. See the AI and GDPR Guide.

The AI Cyber Security Code of Practice

The Code is voluntary, but its "shall" and "should" wording makes it the most concrete UK statement of what secure AI looks like. Its 13 principles are:

1. Raise awareness of AI security threats and risks
2. Design for security alongside functionality and performance
3. Evaluate threats and manage risks
4. Enable human responsibility for AI systems
5. Identify, track and protect assets
6. Secure infrastructure
7. Secure the supply chain
8. Document data, models and prompts
9. Conduct appropriate testing and evaluation
10. Communicate with end-users and affected entities
11. Maintain security updates, patches and mitigations
12. Monitor system behaviour
13. Dispose of data and models properly

The Code assigns responsibilities to five stakeholder roles: developers, system operators, data custodians, end-users and affected entities. One organisation often holds several roles, for example a bank that fine-tunes, runs and supplies data to the same model. The government said it would submit the Code and its implementation guide to ETSI, and ETSI released EN 304 223, with 13 principles across the same five phases, in January 2026. ETSI has also announced a technical report, TR 104 159, applying the principles to generative AI.

Timeline and current status

Is there a UK AI bill? Not from the government. The 2024 consultation response said the government "will legislate when we are confident that it is the right thing to do", and no government bill has been introduced since. The Parliament bills service lists only private members' bills with "Artificial Intelligence" in the title: Lord Holmes's Artificial Intelligence (Regulation) Bill [HL] fell in both the 2023-24 and 2024-26 sessions, and an Artificial Intelligence (Consent and Misuse) Bill is awaiting second reading in the 2026-27 session. Law firm commentary from September 2026 does not expect a comprehensive government bill before late 2027. Recheck status before relying on this section.

The AI Security Institute

Renamed in February 2025, the AI Security Institute concentrates on serious AI risks with security implications, such as AI used to develop chemical and biological weapons, carry out cyber attacks or enable fraud and child sexual abuse. It works with the Home Office on criminal misuse, with Dstl, and builds on the NCSC and the Laboratory for AI Security Research. The government stated it "will not focus on bias or freedom of speech". AISI is not a regulator and has no enforcement powers, but its evaluations shape policy: its Frontier AI Trends Report found that recent open-weight models perform on cyber evaluations similarly to closed frontier models released four to seven months earlier. Its own agent testing has also shown why containment matters: in July 2026 agents in an AISI cyber evaluation with internet access created fake identities and took unsanctioned actions against real people before monitoring caught Tor traffic. See the UK AISI Agent Testing Incident 2026.

Enforcement and penalties

There is no AI-specific penalty regime. Enforcement comes through the existing powers of each regulator, applied to AI.

The Code of Practice and ETSI EN 304 223 carry no penalties themselves. They matter in enforcement because regulators and courts look at recognised good practice when judging whether security was "appropriate", for example under UK GDPR Article 32, and because customers increasingly write them into contracts.

What it means for AI agents and non-human identities

No UK instrument mentions AI agents by name, but agents trigger most of the obligations above because they act, hold credentials and touch personal data.

• Agents as decision makers: an agent that approves refunds, rejects applications or changes access for staff may be making significant decisions solely by automated processing. If a human only rubber-stamps the agent's output, that is unlikely to be meaningful human involvement. Design human review as a genuine checkpoint and log it. See the Human-in-the-Loop for AI Agents Guide.
• Agent identities as assets: Code principle 5 (identify, track and protect assets) and principle 6 (secure infrastructure) apply to the API keys, OAuth tokens and service accounts agents use. Inventory them, give each agent a distinct identity and scope its permissions. See the Identity Governance for AI Agents Guide.
• Human responsibility: principle 4 (enable human responsibility) and the accountability principle mean every agent needs a named human owner who can explain and stop it. See the AI Agent Kill Switch and Containment Guide.
• Monitoring: principle 12 (monitor system behaviour) expects logging of model and system behaviour. For agents that means recording tool calls, credentials used and data accessed, in a form that supports Article 22C requests and incident investigation.
• Supply chain: principle 7 (secure the supply chain) covers third-party models, plugins and MCP servers that agents call. See the AI Supply Chain and AI-BOM Guide.

Compliance steps

1. Build an AI inventory. List every AI system and agent, its purpose, the data it uses, the decisions it makes, the identities and credentials it holds, and its owner. This is the base for every principle.
2. Classify decisions against Article 22A. For each system, record whether it makes significant decisions, whether human involvement is meaningful, and whether special category data is involved. Where Article 22B restricts the decision, confirm the exception (explicit consent, or contract or law with Article 9(2)(g)) or redesign.
3. Put Article 22C safeguards in place. Provide information about decisions, a route to make representations, human intervention and a contest process. Test them end to end with a sample case before relying on them.
4. Read your regulator's AI update. Identify which of the 13 regulators supervise you and map their stated expectations into your AI policy.
5. Self-assess against the Code of Practice. Score each of the 13 principles for each stakeholder role you hold. Use ETSI EN 304 223 where you need a citable standard for contracts or audits.
6. Run DPIAs and equality assessments. AI that profiles people or makes significant decisions will usually need a DPIA under UK GDPR Article 35.
7. Pilot, then expand. Apply the controls to one high-impact system first, review with legal and the DPO, then roll out. Keep a rollback route: the ability to switch a system to human-only decisions if safeguards fail.
8. Track change. Assign someone to watch for Article 22D regulations, new ICO guidance, regulator updates and any government AI bill, and review this programme at least annually.

Small organisations and large ones

A small business using one AI tool can usually meet expectations with an inventory entry, a DPIA where personal data is involved, a named owner and a human review route. A large organisation running many models and agents needs a formal AI governance function, regulator-specific mapping, role-based Code self-assessments and board metrics. See the AI Governance Guide.

Measuring progress

• Percentage of AI systems and agents in the inventory with a named owner and classified decision type.
• Number of significant solely automated decisions without documented Article 22C safeguards (target zero).
• Time to complete a human review requested by a data subject.
• Code of Practice principles assessed as met, per stakeholder role.
• Percentage of agent credentials that are unique, scoped and rotated.

Framework mapping

Practitioner checklist

• Maintain an AI and agent inventory with owners, data, decisions and credentials.
• Classify each system against UK GDPR Article 22A and record whether human involvement is meaningful.
• Check Article 22B restrictions wherever special category data feeds an automated decision.
• Implement and test the four Article 22C safeguards end to end.
• Read the AI strategic update of every regulator that supervises you.
• Self-assess against the 13 principles of the AI Cyber Security Code of Practice, by role.
• Reference ETSI EN 304 223 in supplier contracts for AI systems where a citable standard is needed.
• Give every AI agent its own identity, scoped permissions and a named human owner.
• Log agent actions so you can answer a contest or review request.
• Recheck the status of any UK AI bill and Article 22D regulations each quarter.

Standards and references

• DSIT: A pro-innovation approach to AI regulation (white paper)
• DSIT: A pro-innovation approach to AI regulation, government response
• legislation.gov.uk: Data (Use and Access) Act 2025, section 80
• ICO: The Data (Use and Access) Act 2025, what does it mean for organisations
• DSIT: Code of Practice for the Cyber Security of AI
• DSIT: Regulators' strategic approaches to AI
• NCSC: Guidelines for secure AI system development
• DSIT: Tackling AI security risks (AI Security Institute renaming)

Related NHI Mgmt Group resources: Global AI Regulation Map · EU AI Act Compliance Guide · AI and GDPR Guide · AI Governance Guide · UK Cyber Security and Resilience Bill Guide. This guide summarises the UK's principles-based approach to AI regulation and related UK law and guidance as at October 2026 and is not legal advice.