November 16, 2026 · AI Governance
The FCA's AI Good-Practice Checklist: What Regulators Will Mark Your AI Program Against
On 19 October 2026 the Financial Conduct Authority published “AI in financial services: good practice and areas for improvement”, a consolidated guide to what FCA supervisors have seen work, and where deployments keep going wrong. Jessica Rusu, the FCA’s chief data, information and intelligence officer, framed it in a speech the same day. A week later, the joint Bank of England and FCA survey set the backdrop: 86% of respondent firms now report using AI.
It adds no new rulebook and repeals none of the old one. Read properly, it is the closest thing to an FCA AI good practice checklist: the practices supervisors treat as evidence of competence, and the patterns that read as weakness. Treat it as a marking sheet, not background reading.
What the FCA actually published
The examples were drawn “from industry engagement, supervisory work and our AI Lab initiatives” and span consumer-facing and internal cases across “machine learning, generative AI or agentic AI”. The FCA is careful about status: the examples are “illustrative, non-exhaustive and will evolve as practice develops”, and they “do not create new requirements, prescribe a single model”. What the page does establish is the evidence supervisors can hold up when they ask how a deployment was governed.
The scope is wide: “regulated firms that develop, deploy, procure or use AI in connection with regulated activities, consumer journeys, operations or governance”, including tools “supplied by a third party, embedded in another system or used by staff to support decision making”. Buying the model from a vendor does not move accountability off your books. And “There is no one-size-fits-all approach” appears often enough to read as an instruction to mid-market firms: govern from your own risk profile, not someone else’s template.
Rusu’s read on the themes: “Three things stand out: consumer centricity, accountability and governance, and operational resilience.” The publication states the duty plainly: firms “remain responsible for consumer outcomes”, and one named failure is “Excluding consumers in vulnerable circumstances by default”, where an AI journey removes the human path to support.
Governance: the named under-investment area
The sharpest language is reserved for governance. Rusu: “The second principle is governance, where too many firms are still under-investing.” Her prescription fits one line: “Firms need clear accountability, risk-based controls, human oversight, escalation triggers, and senior managers personally accountable for AI systems.”
The picture of a governed firm: “Proportionate, risk-based AI governance throughout the AI lifecycle”, plus coverage from scoping to decommissioning with formal challenge in the middle: “Formal evaluation, clear accountability and executive approval helped identify and mitigate risks before launching new AI systems.”
For mid-market firms the load-bearing paragraph: “Firms, including smaller firms, incorporated AI oversight into their existing governance frameworks. They used simpler risk tiers to increase testing, guardrails and monitoring for higher-risk use cases, overseen by their existing risk and monitoring committees.” Smaller firms in the sample scaled governance “to their size, capability, and risk profile rather than copying large firm models”. The FCA expects adaptation of what already exists, and even names the source framework: “Firms assessed whether their existing risk frameworks, eg model risk, were appropriate for AI and adapted them where needed.” Extending a model-risk spine to generative and agentic use cases is exactly the problem our model-risk carve-out piece addresses.
On accountability, senior managers are not expected to become technical specialists. They should “take reasonable steps to understand the risks arising from AI use within their area of responsibility”. Good practice maps ownership with care rather than appointing one hero per model, with accountability attaching “from the design stage onwards for each AI use case” and auditable records behind it.
The failure side mirrors it. The FCA names “Informal approval processes or generic governance”. The repeat offender is sequence: “Firms designed and deployed AI use cases without first considering how to govern them.” One case study is blunt: “One firm deployed AI tools before it had established clear governance arrangements”, and the attempted retrofit left gaps in accountability. Rusu’s verdict on retrofitting: “That is the wrong way around.” The other named failures: ownership that was “unclear, fragmented or overly diffused, particularly across technology and business teams”, and decisions documented too thinly to show the steps taken to mitigate risk.
Agent inventories and honest monitoring
The agentic pattern held up as good practice should sound familiar to readers of our shadow-AI roadmap: “Firms developing agentic use cases registered their agents in a central inventory so they were easier to oversee”, with a maintained “clear view of which agents are live, who owns them, what systems and data they can access, and what controls apply”.
Monitoring gets the same outcome-first treatment: “Forward-looking monitoring. Firms adopted monitoring approaches that tracked real-world outcomes and impacts, and adapted to evolving risks, such as model drift.” The toolkit is unglamorous: dashboards “with thresholds and benchmarks used to oversee model performance and escalate issues”, and “Logging and audit trails of AI interactions to help firms monitor outcomes and deal with complaints”. One case study grades claims information on completeness, faithfulness, reasoning and safety, combining “checks by staff with an automated assessment, including LLM-as-a-judge” while watching for “biases shared by models from the same family”.
The deficiencies are as pointed. “Where firms monitored technical metrics rather than outcomes, they had less chance of meaningful oversight.” The competence gap sits with the checker: “Oversight functions lacking AI-specific capability” left firms short of “expertise in their second- and third-line oversight functions to identify or challenge emerging risks”.
Human oversight gets its own warning, and it is the one most likely to sting. In one case study, “the human reviewer role was unclear and became a tick-box exercise over time”. Effective oversight, the FCA says, is about authority, not eyeballs: ensuring people have “the right authority, information, capability and escalation routes” to challenge or intervene. A reviewer who cannot act is a rubber stamp, and rubber stamps now read as a governance failure.
Dependencies belong in the resilience plan
“Firms can outsource AI services, but remain responsible for them.” The resilience chapter folds AI dependency into ordinary resilience scope; for firms in scope of SYSC 15A the mapping is “a requirement”. Some firms still “did not appropriately consider their use of AI, or their dependence on AI models, in their operational resilience plans”.
Vendor oversight is where checklist and survey agree hardest. The FCA recommends vendor-oversight processes up front, “contractual requirements to obtain notice of, and further information about, significant model changes”, and architecture that runs on more than one provider. Named failures: “Overreliance on untested vendor assurances” and “Single-platform dependence without exit options”. The cost case is vivid: “One firm built an AI-enabled product around a particular model without establishing a viable alternative or transition plan. When the model was retired, the firm had to redesign the product.”
The survey puts numbers on the exposure: third-party implementations “accounted for 39% of all reported AI systems” and “53% of high-materiality AI systems”, while the “top three cloud/compute providers” hold 79% of reported relationships. Assurance you have not tested is not assurance; our third-party risk overhaul piece carries the discipline.
Rusu turns it operational: “Make sure you have proper cyber insurance in place.” And as autonomy raises the speed and scale of failure: “That is why firms need to invest now in cyber and operational resilience”. Coverage readiness is an evidence problem, and our cyber-insurance renewal piece is the practical companion.
The adoption backdrop, with the caveat that matters
The joint survey, published 26 October 2026, supplies the context: “86% of respondents reported using artificial intelligence (AI), up from 75% in 2024”, a series that also ran “75% in 2024 and 58% in 2022”. 196 firms responded, 101 larger and 95 smaller. Generative AI now dominates the mix: it “accounted for 56% of all reported AI systems”. Rusu had already named the shift: “Across financial services, firms are moving from experimentation to deployment.”
One caveat travels with any comparison, in the source’s own words: “Differences in respondent composition and question scope mean that the 2026 and 2024 percentages do not relate to the same population of firms and should therefore be compared with caution.”
The most useful split for governance is intent versus operation: “94% of respondents reported firmwide AI policies, principles or standards”, while “ongoing monitoring, leakage testing and periodic review or recertification were each reported by around half of respondents”. Policies pass easily; proving the controls operate between updates is the half the checklist targets.
Where supervision heads next
The forward path is a change in evidence expectations. Existing incident-reporting rules “must extend to significant AI failures and near-misses”, and the speech, citing the AI Minister’s own recognition, warns that near-miss reporting for agentic AI “is becoming a national security priority”. A one-time sign-off is not what agentic systems demand: “Agentic systems therefore require continuous, conditional assurance.” The supervisory agenda goes further, toward “a trusted framework for agentic finance, including clear approaches to agent identity, delegated authority, accountability and execution”. Agent identity and delegated authority are the groundwork our agent-identity tenets piece already covers; firms with that grounding will read the framework as a checklist, not a scramble. On the Mills Review, “We are embedding 6 of the 7 recommendations into live work immediately.”
What to do this quarter
-
Inventory agents and AI use cases first, with owner, systems, data access and controls, keeping the distinction some firms already make “in their inventories between experimental tools and production tools”. Third-party-built AI belongs on the register too.
-
Scale governance through what you already run: simpler risk tiers, heavier testing for the risky tier, existing committees doing the oversight. Nobody is being graded on committee count.
-
Name the accountable owner and keep the receipts: one senior owner per material use case from design stage onward, decisions and escalations auditable. The documentation finding is the one senior managers cannot delegate away.
-
Rebuild monitoring to measure outcomes, not internals: dashboards with thresholds and escalation, drift spot checks, audit trails, and a periodic check on whether human review still catches anything.
-
Stress-test, then test the vendor’s story: map AI dependencies into resilience scenarios, require notice of significant model changes, and drill the fallback. An exit plan that has never been exercised is a paragraph, not a plan.
-
Bring incident reporting into AI scope now: agree what counts as a significant AI failure or near miss, who declares it, and how it flows into existing reporting obligations. The lane is expanding; agreeing the definition is the cheap early step.
None of this requires new technology. It is ordinary control discipline applied to AI: know what you run, prove someone owns it, watch what it does, keep the receipts, rehearse the exit. Firms that can show operating evidence will read each FCA update as guidance; firms that cannot will read it as a finding on a delay.
Related reading:
- The shadow-AI roadmap behind the FCA’s central agent inventory: https://ntdciso.com/insights/ai-inventory-shadow-ai-roadmap/
- Adapting model risk frameworks for generative and agentic AI: https://ntdciso.com/insights/genai-model-risk-carve-out/
- Vendor assurance testing, concentration risk and exit plans: https://ntdciso.com/insights/bank-tprm-guidance-overhaul/
- Governing agentic AI before the supervisor does: https://ntdciso.com/insights/agentic-ai-governance-trust-gap/
- Agent identity as the groundwork for agentic finance: https://ntdciso.com/insights/nist-agent-identity-tenets/
- Cyber insurance readiness behind the “proper cyber insurance” line: https://ntdciso.com/insights/cyber-insurance-renewal-readiness/
Related reading
Agentic AI governance: what 86% deployment and 34% trust mean for your next audit
AI Inventory & Shadow AI Governance Roadmap
The GenAI Carve-Out in Bank Model Risk Guidance
US Bank Regulators Just Rewrote the Vendor Rules
Cyber Insurance Requirements Rise While Rates Fall: What Mid-Market Fintech Should Prove at Renewal