)
Last updated: 16 October, 2026
What does AI governance for UK financial services look like when the regulator says it will not write AI-specific rules?
It looks like your existing obligations, applied to systems that now act on their own. The FCA relies on the Consumer Duty and the Senior Managers and Certification Regime (SM&CR), the PRA relies on model risk management, and the ICO's data protection rules now cover automated decisions. Together they ask for one thing: evidence that a named person controls each AI system, and that you can show it.
TL;DR: Key Takeaways
- The FCA states it does not plan extra regulations for AI and points firms to the Consumer Duty and SM&CR instead (FCA).
- The Mills Review, published on 6 July 2026 after 140 written submissions, found the SM&CR model still sound for lower-autonomy AI, but flagged pressure once AI approves or observes decisions on its own (Skadden).
- HM Treasury designated the first four critical third parties on 10 July 2026, and new material third-party reporting rules apply from 18 March 2027 (MoFo).
- The Treasury Committee asked the regulators on 20 January 2026 for practical AI guidance on consumer protection and SM&CR accountability by the end of 2026 (Global Policy Watch).
- The Data (Use and Access) Act's automated decision-making rules (Articles 22A to 22D) have applied since 5 February 2026, and an ICO statutory code on AI is expected in 2027 (Bratby Law, Arnold & Porter).
At a glance: who expects what from UK financial services AI governance
Quick Answer
What is AI governance for UK financial services?
It is the set of owners, controls and records that show a regulated firm manages AI within existing rules: the Consumer Duty and SM&CR at the FCA, model risk management at the PRA, operational resilience and third-party oversight, and UK GDPR for automated decisions. No single AI rulebook exists, so the work is mapping each AI system to those obligations and keeping the evidence.
| Regulator or body | Instrument | What it asks of firms | Status at 16 October 2026 |
|---|---|---|---|
| FCA | Consumer Duty, SM&CR | Fair outcomes and named senior accountability for AI | No AI-specific rules planned |
| PRA and Bank of England | Model risk management (SS1/23), resilience | Model governance that scales to generative and agentic AI | Roundtables and a 2026 AI adoption survey |
| HM Treasury, FCA, PRA | Critical third party regime | Oversight of major providers, with firms still owning their own risk | First designations on 10 July 2026 |
| ICO | UK GDPR Articles 22A to 22D, statutory AI code | Safeguards for significant automated decisions | Articles in force, code expected 2027 |
)
What AI governance for UK financial services means in practice
The FCA's own page says it does not intend to introduce extra regulation for AI. Its approach is principles-based and outcomes-focused, and it names the Consumer Duty and senior manager accountability under SM&CR as the main anchors (FCA, 2026). The FCA's AI Lab offers support for firms developing AI use cases.
A lighter rulebook does not mean lighter supervision. According to Ropes & Gray (2026), the absence of an AI rulebook still comes with active supervisory engagement on governance, accountability and customer outcomes. The firm lists what regulators are likely to look for: a live inventory of AI use with clear ownership, useful information for boards, AI scenarios in operational resilience testing, and proportionate due diligence on third-party AI providers.
The practical reading is simple. Treat every AI system as a regulated activity with an owner. If you cannot say who approved a model, what customer outcome it affects and how you would switch it off, the Consumer Duty and SM&CR questions will find that gap first.
What the Mills Review means for senior managers
The FCA launched the Mills Review in January 2026 to examine how advanced AI may reshape retail financial services. Skadden (2026) reports that it was published on 6 July 2026, drew on 140 written submissions and a survey of more than 5,000 UK adults, and made seven recommendations to the FCA Board.
The review describes five levels of AI autonomy. The table shows where governance pressure builds.
| Autonomy level | What the AI does | Governance pressure |
|---|---|---|
| Operator | Assists when asked | Low: existing SM&CR accountability holds |
| Collaborator | Works alongside staff | Low to moderate |
| Consultant | Recommends, a human decides | Moderate: evidence the human review is real |
| Approver | Decides within limits, a human oversees | High: opacity and model drift complicate accountability |
| Observer | Executes approved actions, a human monitors | Highest: hard to show meaningful human control |
Per the same summary, no respondent argued the senior managers regime itself should change, and the review recommends no new Senior Management Functions. It does suggest clearer FCA guidance on the "reasonable steps" a senior manager should take, plus better assurance before and after deployment. The FCA Board has not yet said which recommendations it will adopt, and the FCA expects to publish good and poor practice findings later in 2026.
For a CISO, the actionable point is the autonomy level of each system. A chatbot that drafts replies for staff and an agent that approves refunds sit at different points on that scale, and your governance should show it.
PRA and Bank of England expectations: model risk and resilience
The PRA and Bank of England also take a technology-agnostic line. In their 1 April 2026 response to a government letter they reaffirmed that approach and said they will keep reviewing whether further guardrails are needed (Global Policy Watch, 2026). Their summary of earlier roundtables, published on 16 February 2026, records that firms backed the principles-based approach, including the PRA's model risk management expectations in SS1/23.
The same summary records two open questions from firms. Does model risk management scale to generative and agentic AI, and how does human-in-the-loop work as AI takes more decisions? Both are worth answering in your own policy before a supervisor asks.
On cyber resilience, a joint statement from the Bank of England, FCA and HM Treasury on 15 May 2026 urged firms to strengthen defences against frontier AI threats. Ropes & Gray (2026) lists its five areas: governance and strategy, vulnerability management, third-party risk, protection, and response and recovery. It adds that the National Cyber Security Centre recorded 204 nationally significant attacks in the 12 months to September 2025, up from 89 a year earlier.
Third-party and critical third-party AI risk
Most financial services AI is bought, not built. According to Ropes & Gray (2026), about one third of AI use cases in financial services involve third-party implementations. The FCA and ICO noted in a June 2025 joint blog that firms were unsure who is accountable when third-party AI tools make automated decisions, and some held back from deploying them.
The critical third party regime now adds oversight of the providers themselves. MoFo (2026) reports that HM Treasury designated the first four providers on 10 July 2026, effective 13 July, and that the FCA and PRA can impose conditions or prohibit a designated provider from serving authorised firms. The regime does not move your responsibility. Financial services firms remain responsible for managing their own third-party risk, and new rules on material third-party arrangements apply from 18 March 2027, including annual registers of those arrangements.
Build the register now, and include AI services that sit inside other vendors' products. A vendor that adds an agent to a tool you already use has changed your risk without a procurement event.
Data protection overlap: automated decisions and the ICO code
AI governance in a regulated firm is also data protection governance. Articles 22A to 22D of the UK GDPR replace the old Article 22 and have applied since 5 February 2026 (Bratby Law, 2026). The ICO's draft guidance defines a significant decision as one based solely on automated processing that has a legal or similarly significant effect, and treats decisions without meaningful human involvement as solely automated (ICO, 2026).
The draft guidance lists the safeguards: information about decisions, the ability to make representations, human intervention and the ability to contest. Credit, fraud and pricing decisions are the obvious cases in financial services. Special category data in automated decisions faces tighter conditions.
Two dates matter. The ICO consulted on the draft guidance from 31 March to 29 May 2026. Arnold & Porter (2026) reports that a statutory code of practice on AI and automated decision-making, created under SI 2026/425, is not expected to take effect until 2027. When it does, courts must take it into account in relevant proceedings. If your firm also sells into the EU, note that the AI Act's stand-alone high-risk obligations have been reported as moving to 2 December 2027 under the digital omnibus (Cloud Security Alliance, 2026), so check the final text with counsel.
Building an AI governance operating model
An operating model turns the sections above into routine. The table is a practical sequence, not a regulatory checklist.
| Step | What to produce | Regulatory question it answers |
|---|---|---|
| Inventory | Every AI system and agent, owner, vendor, data, autonomy level | Who is accountable, and for what? |
| Classify | Customer impact, autonomy level, automated decision status | Where does the Consumer Duty or Article 22A apply? |
| Control | Permissions, limits and approval rules per agent | Is human control meaningful? |
| Test | Pre-deployment and recurring tests, including adversarial tests | Can you evidence model risk management? |
| Third parties | Register, due diligence, exit plans | Are you ready for the March 2027 reporting rules? |
| Respond | Incident playbooks for AI failures and attacks | Does resilience planning include AI? |
| Report | Board information on use cases, incidents and dependencies | Can senior managers show reasonable steps? |
Start with the inventory, because every other step depends on it. A platform for Agent Posture Management (TrustLens) can discover agents and connections you did not list, which matters because shadow deployments are what supervisors tend to find first. For background on the wider framework choice, see our AI governance guide.
How we compared
We used the FCA's own page for its stated position, law-firm summaries of the Mills Review, the Treasury Committee report and the critical third party designations, and the ICO's draft guidance for automated decisions. Law-firm and trade summaries are secondary sources, so check the primary publication before citing a date or figure in a regulatory submission.
The evidence a supervisor can ask for
Prepare a short evidence pack per material AI system. It should hold the owner and approval record, the data used, the test results, the monitoring in place, the human oversight design and the exit plan for any vendor. Ropes & Gray (2026) points to approvals, risk assessments, data protection impact assessments, testing records and incident playbooks as the records firms should be ready to show.
Keep it current. A pack that was accurate at launch and never updated is weaker than a shorter one reviewed every quarter.
A 90-day plan for a UK firm
In the first 30 days, build the inventory and name an owner for each system. In days 31 to 60, classify by autonomy and customer impact, and map automated decisions against Articles 22A to 22D. In days 61 to 90, test the highest-risk systems, draft incident playbooks and give the board a first report. Revisit the plan when the FCA publishes its good and poor practice findings and when the ICO finalises its guidance.
Security and governance: enterprise risks of ungoverned AI agents
Agents add risks that classic model governance misses. Prompt injection can redirect an agent through content it reads, data can leak through prompts and tool outputs, and an agent with broad tool access can misuse it. For regulated firms these are also governance failures, because they weaken the meaningful human control that firms must be able to evidence.
NeuralTrust approaches this through the Runtime Security Mesh. TrustGate sets per-agent permissions and limits, ships 200+ pre-built MCP servers and acts as a policy and logging point between agents and tools. TrustLens inventories agents and connections. TrustGuard monitors agent behaviour at runtime. TrustTest red teams agents before they reach production. NeuralTrust deploys in your own infrastructure or on-premises, with a managed service hosted in the EU or the US. The platform holds four Gartner Hype Cycle 2026 recognitions, including AI Runtime Defense. See also our guidance on the NCSC's AI guidelines and on AI gateway governance.
Which should you choose?
| Firm type | First priority | Why |
|---|---|---|
| Banks and building societies applying SS1/23 | Extend the model inventory to generative and agentic systems | Firms questioned whether model risk management scales |
| Insurers and asset managers buying AI from vendors | Third-party register and exit plans | Reporting rules apply from 18 March 2027 |
| Retail-facing fintechs and advisers | Consumer Duty outcome testing and named SM&CR owners | The Mills Review focuses on consumer-facing AI |
| Firms with automated credit or fraud decisions | Map each decision to Articles 22A to 22D | The ADM safeguards have applied since 5 February 2026 |
Conclusion
AI governance for UK financial services is not waiting for a new rulebook, because the rules that matter already apply. The FCA expects the Consumer Duty and SM&CR to cover AI, the PRA expects model risk management to stretch, and the ICO expects safeguards on automated decisions. Build an inventory, name an owner for every system, test the risky ones and keep evidence that a human can really intervene. Firms that do this now will answer the FCA's good and poor practice findings from a position of record, not recollection.
Secure Financial Services AI in Production with NeuralTrust
Discover agents, set per-agent limits and monitor every tool call before a supervisor asks.
)
Related Comparisons
- AI Governance 2026: Frameworks & Tools
- NCSC AI Guidelines for UK Enterprises 2026
- AI Gateway Governance for Enterprises
- The UK CISO's Complete Guide to Enterprise AI Security (2026)
FAQs about AI Governance for UK Financial Services
1. Does the FCA have specific AI rules?
No. The FCA says it does not plan extra regulations for AI and relies on existing frameworks, mainly the Consumer Duty and senior manager accountability under SM&CR. Firms are still expected to evidence governance, testing and customer outcomes for AI they deploy.
2. Who is accountable for AI under SM&CR?
Senior managers remain accountable under the existing regime. The Mills Review found the model sound for lower-autonomy AI but flagged difficulty identifying a responsible manager when AI approves or observes decisions. It suggests clearer guidance on reasonable steps.
3. Do UK banks need a separate AI model risk framework?
Not necessarily. The PRA's roundtable summary shows firms backing SS1/23 on model risk management while questioning whether it scales to generative and agentic AI. Most firms extend their existing framework with agent-specific controls and human oversight evidence.
4. What do the critical third party rules mean for AI vendors?
HM Treasury designated the first four critical third parties in July 2026, and regulators can impose conditions on them. Firms remain responsible for their own third-party risk, and material third-party reporting rules apply from 18 March 2027. Keep a register that includes AI services.
5. Does UK GDPR restrict automated decisions in financial services?
Yes, for significant decisions made solely by automated means. Articles 22A to 22D, in force since 5 February 2026, require safeguards such as information, representations, human intervention and a way to contest. Special category data faces tighter conditions.
6. When will the ICO's AI code of practice apply?
The statutory code on AI and automated decision-making is not expected to take effect until 2027. The ICO consulted on draft automated decision-making guidance between 31 March and 29 May 2026, and final guidance is expected later in 2026.
About the Author
Roger Howroyd is Head of Global SEO and AI at NeuralTrust, where he leads the company's search strategy across SEO, AEO, GEO, and LLM optimization. He specialises in AI-powered search, content strategy, and SEM. Connect on LinkedIn.
NeuralTrust is the leading platform for securing and scaling AI agents. Named a Pioneer in the Gartner Emerging Market Quadrant for AI Application Security 2026, recognised across four Gartner Hype Cycle reports in 2026, and featured in the Gartner Market Guide for Guardian Agents 2026, the Gartner Market Guide for AI Gateways 2025 and the KuppingerCole Leadership Compass for Generative AI Defense 2025. Headquartered in Barcelona with offices in London and New York. ISO 27001 certified.
Sources
- FCA, AI and the FCA: our approach, page updated 2 October 2026.
- Skadden, FCA's Mills Review Charts a Course for AI-Enabled Retail Financial Services, 21 July 2026.
- Global Policy Watch, UK Financial Services Regulators' Approach to Artificial Intelligence in 2026, April 2026.
- Ropes & Gray, From Principles to Practice: The FCA's Evolving Expectations on AI Governance, June 2026.
- MoFo, UK Announces List of First Critical Third Parties, July 2026.
- Bratby Law, Data Use and Access Act 2025 Commencement Changes, accessed 16 October 2026.
- Arnold & Porter, The ICO's New Statutory Duty to Produce an AI Code of Practice, 17 August 2026.
- ICO, Automated decision-making, including profiling (draft guidance), 31 March 2026.
- Cloud Security Alliance Labs, EU AI Act omnibus: high-risk deadline delay, 2026.
)
)
)
)