UK regulation map

Every AI-assisted outcome has to be fair and explainable. A decline or delay needs a plain-English reason and a route to a person, and outcomes are monitored by customer group, including customers in vulnerable circumstances.

HM Treasury has designated large cloud and technology providers as critical third parties (SI 2026/777), so the Bank of England, PRA and FCA oversee the services they provide to financial firms. That doesn't move your accountability: map which important business services rely on which cloud AI services, and keep exit and fallback plans tested.

A named senior manager owns each AI-supported process. Keep a clear record of what the AI does, what it can't do and how its quality is measured.

AI and machine learning models go in the model inventory, are tiered by risk, validated independently and monitored for drift.

The AI gathers evidence and drafts. A person confirms due diligence outcomes and possible sanctions matches, and the AI never files or drafts anything that could tip off a customer. The rules sit in a versioned rule set owned by financial crime, not in a prompt.

Fraud models inform a person's decision to delay a payment or reimburse a claim. Claims are normally decided within 5 business days, or up to 35 if the clock is stopped, so the design keeps evidence ready for the investigator.

Claims and renewal AI must lead to fair value and good outcomes. AI can speed claims up, but a person makes every decline.

Claims are handled promptly and fairly. Claims handling is an FCA supervisory priority in 2026, so measure time to settle and outcome quality by customer group, not just throughput.

A renewal price can't be higher than the equivalent new business price. Pricing models need monitoring and evidence for the annual attestation a senior manager signs.

Cloud and AI providers can be material outsourcing. Keep exit plans, audit rights and a tested fallback for each important business service the AI supports.

Age and disability can only shape pricing or underwriting where it's based on relevant information from a reliable source. Test models for proxies as well.

Any AI that touches patient care needs a clinical safety officer, a hazard log and a safety case, from the supplier (DCB0129) and from the organisation that deploys it (DCB0160).

If the AI is intended to support diagnosis or treatment, it may be a medical device. The MHRA said on 29 July 2026 that ambient tools that only transcribe, summarise or draft for a clinician to review are generally not.

Patients are told and can say no, and the clinician reviews every note before it goes into the record.

Suppliers meet the Data Security and Protection Toolkit and the Digital Technology Assessment Criteria. Use the minimum patient data needed and respect the national data opt-out where it applies.

Providers stay accountable for safe care and good governance. In social care, AI can draft notes and spot patterns in incidents, but staff decide on safeguarding referrals and CQC notifications.

Central government departments and in-scope arm's-length bodies publish a record for tools that significantly influence decisions or interact with the public. Councils can publish voluntarily. Design so the record is easy to keep current.

An equality impact assessment before launch, and outcomes monitored after, by protected characteristic where you can.

Meaningful human control, the right tool for the job, and openness. Service assessments treat the AI like any other part of the service.

Prompts, outputs and decision records can be requested. Keep them findable, and retain them on the same schedule as other records.

Casework, benefits and grants decisions with significant effects need the safeguards. The AI prepares the case; a caseworker decides.

Product content and agent answers must not mislead, and fake reviews are banned. You are responsible for what your agent says, so test it before launch and keep monitoring.

An agent must never talk a customer out of their statutory rights to a refund, repair or cancellation. Put the rights in rules the agent can't override; the AI explains them.

Keep card data out of the AI entirely. Payments go through a separate compliant flow, and transcripts and logs are masked.

AI-written listings take safety, allergen and energy information from the product record, never from the model, and a person approves new or changed claims.

Forecasts shared with suppliers must be prepared in good faith and with due care. Document how the forecasting model works and how a planner can override it.

Predictive maintenance informs planning but doesn't replace required inspections. A competent person decides whether equipment is safe to use.

Vision inspection is part of your quality system. Record what was inspected, by which model version, and what a person overrode.

AI that reads from the plant does it through segmented connections, one-way where possible. A cloud agent never writes directly to a control system.

Machinery with self-learning safety functions faces stricter assessment and cyber security requirements. Great Britain will keep recognising CE-marked machinery that meets it.

Cameras on a line also see people. Tell staff, minimise and blur what you keep, and run a DPIA.

Operators of essential services in energy and water must secure the systems they rely on, including AI that touches operational systems, and report incidents. For downstream gas and electricity, DESNZ and Ofgem are the joint competent authority. The Bill brings managed service providers into scope and tightens reporting.

Customers must be treated fairly and given clear information. Suppliers generally can't back-bill domestic and microbusiness customers for energy used more than 12 months before a correct bill, and since May 2026 they must take all reasonable steps to restore a smart meter to smart mode within 90 days. Billing exceptions need fixing quickly, against fixed clocks the AI can't move.

AI can spot signs that a customer needs extra help and suggest a referral. A trained person speaks to the customer and records their needs.

Leak detection models should show measurable progress against leakage commitments, with evidence you can audit. The government plans to replace Ofwat with a single water regulator; Ofwat stays in place until then.

A field engineer assistant answers from approved procedures. It never overrides a permit to work or safe isolation.

The declarant stays liable for what's declared. The AI drafts classifications and declarations from documents; a person checks and submits.

Goods stay at the trader's risk until the customer has them. Proactive delay messages help, but they don't change the customer's rights.

Re-planning to recover a delay can never push a driver over their hours. Hours rules are hard limits in the optimiser.

Pick and route optimisation uses worker data. Tell staff, minimise what you keep, and build manual handling limits into the model.

Delivery updates are service messages. Keep offers out of them unless the customer has agreed to marketing.

Competence, supervision and confidentiality apply to AI-assisted work as to any other. A solicitor stays responsible for the advice. Client information goes into AI tools only with contractual, technical and organisational safeguards, and never to train a model unless the client has authorised it.

Privileged documents stay in an environment you control, open only to the matter team and never used to train models. Log who saw what.

Document what each AI tool does, when it's appropriate to use, how it was tested and how you know it works. Audit judgement and sign-off stay with people.

Professional competence, due care and confidentiality apply whether the first draft came from a person or a model. For auditors, the FRC Ethical Standard governs independence, so AI tools and data must not compromise it.

A bid drafted with AI must still be accurate. Check every claim against evidence, and follow the buyer's rules on AI use in the tender.

Complaints stay easy to make and follow the provider's approved code, with access to free dispute resolution. AI should help identify customers in vulnerable circumstances, never steer them away from help.

End-of-contract and annual best-tariff notices come from billing on fixed dates, never from a model. Since 17 January 2025, in-contract price rises must be set out in pounds and pence. AI-drafted retention offers must match these terms.

Network AI that can change configuration is a privileged access path. Threat-model it, limit it to approved actions, and keep a person in the loop for changes to the network's management plane.

Churn models are profiling, so say so in your privacy information. Retention offers sent by email or text need marketing consent or the soft opt-in.

AI tools used by pupils and staff should meet the DfE's standards on filtering, monitoring, security and privacy. Ask suppliers to show how they meet each one.

Safeguarding is never left to an AI to judge. Any concern goes straight to the designated safeguarding lead.

Services likely to be used by children start with high privacy settings: minimal data, profiling off by default, and no nudges to share more.

Early-warning models prompt a conversation with staff; they never decide progression. Admissions and enquiry assistants give accurate, complete information, in line with consumer law. The OfS consulted in 2026 on a new condition on treating students fairly to replace C1 and C3; check which applies to you.

Treat an AI detector's score as a reason to look closer, not as evidence on its own. Concerns go through the centre's malpractice process.

For higher-risk buildings, the golden thread must be accurate, accessible and kept up to date. AI can extract and link information, but the accountable person signs off what goes in. The Building Safety Regulator has been a standalone body under MHCLG since 27 January 2026.

Social landlords must investigate and fix damp and mould and emergency hazards within fixed timescales, extending to more hazards from 30 November 2026. AI can triage and predict, but every report starts the clock, so nothing waits on a model.

Landlords need accurate data on the condition of every home and must treat tenants fairly. Risk models feed the stock condition plan and the repairs queue; they don't decide who gets help.

AI can draft complaint responses, but the landlord stays accountable for the outcome and must meet the Code's timescales.

Site safety monitoring supports the principal contractor's duties; it doesn't replace them. Cameras capture workers, so tell them and keep only what you need.

Every possible adverse event has to be captured and assessed on time. AI can extract and code reports, but it never discards one, and a qualified person reviews seriousness and causality.

Data must be attributable, legible, contemporaneous, original and accurate (ALCOA+). Validate the AI for its intended use and keep a full audit trail.

The 2025 draft of Annex 22 expects a documented AI lifecycle, performance monitoring and change control, and says large language models shouldn't be used in critical GMP applications. Keep AI in deviation and CAPA work advisory, with a person deciding.

The sponsor stays responsible for trial documents and data. AI-drafted documents are reviewed and version-controlled before use.

Trustees stay responsible for decisions made with AI. Serious incidents, including safeguarding incidents and data breaches, are reported to the Commission promptly, and a person makes that call.

AI-written appeals and supporter messages must be legal, open, honest and respectful, with extra care for people who may be in vulnerable circumstances.

Every claim needs a valid declaration. AI can match donations to declarations and flag gaps, but it never creates a declaration.

Charities can now email or text supporters who showed interest in their cause without asking for consent first, as long as people can opt out when their details are collected and in every message. Keep that opt-out in every AI-drafted message.

Impact figures drawn from case notes by AI must be accurate and traceable to the records behind them.

AI in ordering, allocation and cold chain keeps the Responsible Person in control, with full traceability and validated systems.

A pharmacist stays accountable for every clinical check. Online services need identity and suitability checks that AI can support but not replace.

Decision support follows the NHS clinical pathways and has a clinical safety case. The pharmacist decides and records.

Controlled drug orders need the right licence and paperwork. An ordering agent can't skip or waive a check.

Claims must match what was supplied. AI can reconcile and flag; people submit.

Screening tools must be tested for bias and explainable to candidates. The ICO's 2024 audit findings on AI recruitment tools set clear expectations: minimise data, don't infer protected characteristics, and let people challenge outcomes.

An onboarding assistant can guide the process, but checks follow the prescribed methods, including certified digital verification services, and a person confirms the result.

Invoice processing keeps a digital trail from the source document to the VAT return. From April 2029, VAT invoices between UK VAT-registered businesses move to e-invoices exchanged over Peppol.

Large organisations need reasonable procedures to prevent fraud, and supply chain checks. AI-assisted supplier and payment controls are part of those procedures, so keep them documented and tested.

Service desk and security agents get least-privilege access and full logging. Resetting credentials, granting access or isolating a machine needs a person's approval.