AI Agent Compliance in the UK: A Guide for Fintechs and Banks

There is no AI-specific law for UK financial services. AI agents at UK payment firms, e-money institutions, fintechs, and banks fall under the rules those firms already follow: the FCA Principles and Consumer Duty, SM&CR accountability, SYSC systems and controls, outsourcing and operational resilience, PRA model risk and third-party expectations for banks, and UK GDPR. The regulators' message is consistent: an agent acting for you is you acting, and you answer for it.

This is a guide to what those rules ask of AI agents and how teams meet them on Runtime. It is the UK companion to AI agent risk management, which covers the controls in depth. For other regions, see the US and EU guides.

What it covers. The rules that apply, what each one asks, and where the evidence lives when the FCA, your auditor, or a partner bank asks.

What it is not. Legal advice. Runtime enforces and records the rules your team writes; your compliance team and counsel decide how each rule applies to you.

At a glance

Is there an AI law in the UK?

No. The government's AI regulation white paper response (6 February 2024) set five cross-sectoral principles for existing regulators to apply within their remits, on a non-statutory basis: safety, security and robustness; appropriate transparency and explainability; fairness; accountability and governance; and contestability and redress. As of October 2026, no general AI bill has been passed.

The FCA follows the same line. Its AI approach page says: "We do not plan to introduce extra regulations for AI." The Mills Review (6 July 2026) looked at how autonomous AI could reshape retail finance and recommended the FCA "monitor the transition to autonomous models and adapt regulatory frameworks," without changing the current approach. The FCA's AI Live Testing second cohort (announced 21 April 2026) includes agentic AI use cases such as agentic payments, and the FCA says it will publish a good and poor practice report on AI later in 2026.

So the work is mapping existing rules onto agents.

How do FCA rules apply to AI agents?

Principles and Consumer Duty

What it asks. The Principles for Businesses require firms to act with due skill, care and diligence and to have adequate risk management systems. The Consumer Duty (PRIN 2A, in force since 31 July 2023) requires good outcomes for retail customers across products and services, price and value, consumer understanding, and consumer support, and avoiding foreseeable harm. The FCA's AI Update (22 April 2024) says these rules apply to AI use. An agent drafting dispute replies or support answers is part of how you meet those outcomes, and you need to be able to show it.

How Runtime helps.

• Approvals. Anything going to a customer can wait for a person to approve it in the dashboard or Slack.
• Record of every run. Each reply is kept with the request, the steps, and the data the agent used, so you can review outcomes for groups of customers.
• Evals on change. Agents are re-checked against cases your team already resolved before a model, skill, or tool change goes live.

SM&CR accountability

What it asks. Under the Senior Managers and Certification Regime, each senior manager has a Statement of Responsibilities and a duty to take reasonable steps to control their area. The FCA has not created a dedicated AI senior manager; the AI Update points to existing SM&CR accountability. The PS26/6 reforms (21 April 2026) streamlined approvals but did not add AI-specific rules. If an agent works in payments operations, the senior manager for payments operations answers for it. SM&CR does not apply to payment or e-money institutions, but the same logic holds: a named person should own each agent.

How Runtime helps.

• Agent inventory with owners. Every agent has its own identity and a named owner.
• RBAC and scoped identity. Scoped API keys and a list of who can call each agent, so access matches the job.
• Risk scoring. Every run is scored Low, Medium, or High by what it touched, so the accountable senior manager sees where exposure sits.

SYSC systems, controls, and outsourcing

What it asks. SYSC requires adequate systems and controls and orderly records. SYSC 8 says a firm that outsources critical or important functions keeps responsibility and must oversee the provider. Payment and e-money firms have equivalent outsourcing conditions under the Payment Services Regulations 2017 and Electronic Money Regulations 2011. An AI vendor running agents for you is a third party to assess.

How Runtime helps.

• Guardrails. Org-wide rules, egress allowlists, command allow and deny lists, and approval gates.
• Audit export. The inventory, runs, approvals, and risk scores, exported for your outsourcing file or a review.
• SOC 2. Runtime is SOC 2 audited, which supports your third-party due diligence.

Operational resilience

What it asks. Under operational resilience rules (PS21/3, in force 31 March 2022), firms in scope, which under SYSC 15A include payment and e-money institutions, had until 31 March 2025 to show they can stay within impact tolerances for their important business services, including mapping the people, processes, technology, and third parties behind them. If an agent supports an important business service, it belongs in that map.

How Runtime helps.

• Isolated sandboxes, own cloud. Agents run in isolated sandboxes wiped at session end, in your own AWS, GCP, or Azure account, or self-hosted.
• Agent inventory. The systems each agent can reach, ready for your service mapping.

Safeguarding for payment and e-money firms

What it asks. PS25/12 brought in a stronger safeguarding regime for payment and e-money institutions from 7 May 2026, with more consistent record keeping and stronger reporting and monitoring. Agents that help with reconciliations or safeguarding breaks touch this directly.

How Runtime helps.

• Approvals. Money movement waits for a named person.
• Record of every run. Each reconciliation run is kept with its steps and the data it used, so a break can be traced.

What do the PRA rules mean for banks?

SS1/23 model risk management

What it asks. SS1/23 (effective 17 May 2024, amended 23 April 2026 under LIAF01/26 to clarify its link to internal model approval) applies to banks, building societies, and PRA-designated investment firms with internal model approval. Its five principles are model identification and risk classification, governance, development and use, independent validation, and risk mitigants. It does not name AI agents; your model risk function decides what is in scope.

How Runtime helps.

• Agent inventory. Model identification starts with knowing which agents exist and which models they use.
• Evals on change. Re-checks after every model, skill, or tool change, with results kept alongside the change history.
• Risk scoring. Per-run scores for ongoing monitoring.

SS2/21 outsourcing and critical third parties

What it asks. SS2/21 (effective 31 March 2022) sets expectations for outsourcing and third-party risk, including data security and exit planning. Separately, HM Treasury designated the first critical third parties (AWS, Google Cloud, Microsoft, and Oracle entities) with effect from 13 July 2026. That oversight covers the providers; it does not move responsibility off the firms that use them.

How Runtime helps.

• Own cloud. Deploy in your own cloud account or self-host, so agents sit inside the third-party arrangements you already manage.
• Masking. Card numbers and secrets are masked before any prompt or log, and Runtime never trains on your data.

What does UK GDPR require of AI agents?

What it asks. UK GDPR and the Data Protection Act 2018 require a lawful basis, data minimisation, and appropriate security. Section 80 of the Data (Use and Access) Act 2025 came into force on 5 February 2026 and replaced the old Article 22 restriction on solely automated decisions. Significant automated decisions are now allowed on wider lawful bases, with safeguards: telling people, letting them make representations, offering human intervention, and letting them contest the decision. Special category data stays more protected. The ICO says all DUAA data protection provisions were in force by 19 June 2026, consulted on updated ADM guidance in spring 2026, and opened an agentic AI call for evidence on 8 October 2026, closing 20 November 2026. In its words, "the fact AI agents act with autonomy is not an excuse for poor compliance."

How Runtime helps.

• Masking. Sensitive values are masked before they reach a prompt or a log.
• Approvals. A person can review any decision about a customer before it takes effect, which supports human intervention.
• Record of every run. What the agent saw and did, so you can explain and revisit a decision when someone contests it.

Standards you'll be asked about

Runtime supports your assessment against these. It does not certify your compliance with any of them.

What the FCA, your auditor, or a partner bank will ask

A 30-day plan

You can start with a prompt like this:

List every agent in our workspace with its owner, the systems it can reach, and whether it can contact customers or move money. Flag any agent without a named owner and any agent that makes decisions about customers without an approval step.

A forward-deployed AI engineer helps your team set up the first agents and guardrails. The AI agent risk management guide covers risk scoring, evals, and audits in more detail.

The point

The UK chose to regulate AI through the rules firms already follow, so agents inherit all of them. Give every agent an owner who maps to a senior manager, put a person in front of money and customer decisions, and keep the record. Then the questions from the FCA, the ICO, or your partner bank already have answers.