GBAF Logo
Global Banking & Finance Awards® 2026 Nominations open, free to enter Nominate now →
The FCA Has Highlighted Good AI Governance. But Would Yours Work in Practice? - Technology news and analysis from Global Banking & Finance Review
Technology

The FCA Has Highlighted Good AI Governance. But Would Yours Work in Practice?

Published by Barnali Pal Sinha

Posted on November 19, 2026

6 min read
Add as preferred source on Google

By Findlay Whitelaw, Field CISO, Exabeam

Before giving an AI agent authority to act, I would want to know three things. Can we see what it is doing? Who is accountable for its actions and outcomes? And who has the authority to intervene when something does not look right?

Try answering those questions for a process that crosses business teams, internal systems and third-party platforms. The answers are not always straightforward.

The Financial Conduct Authority’s publication, AI in financial services: good practice and areas for improvement, published on 19 October 2026, provides a useful starting point. It offers examples of good practice and areas for improvement within the existing regulatory framework, rather than introducing new requirements. In particular, it discusses meaningful human oversight and examples of firms centrally registering AI agents, including their ownership, access and controls.

For me, the question is not simply whether an organisation has an AI governance framework. It is whether that framework works when AI is operating in the business. Increasingly, conversations I am having with organisations are about visibility, accountability, oversight and the ability to intervene.

Visibility: do we know what is operating, and what it is doing?

Organisations are becoming more structured in how they approve and deploy AI. But much of that governance applies to the AI they know about. What about AI embedded in existing applications, introduced through third parties or adopted outside established approval processes? And what about agents whose tools, permissions or responsibilities have changed since approval?

An inventory is an important starting point, not proof of complete coverage. Discovery and governance capabilities help establish what is present. Identity and behavioural monitoring can help explain what is happening, provided the relevant activity is observable. These are complementary questions, and neither capability solves the other.

For example, behavioural analytics may identify first-seen activity, unusual access or a change in established patterns. That does not, on its own, prove an AI agent is involved or that anything malicious has occurred. It gives a team a reason to investigate, using identity, business and technical context.

Knowing an agent exists is not the same as knowing what it actually does.

Ownership and oversight: who can judge whether the outcome is right?

The FCA’s discussion of meaningful human oversight is particularly relevant here. Effective oversight requires more than a named owner or a person clicking an approval button. People need appropriate information, capability, authority and escalation routes.

Consider a hypothetical insurer using an approved AI agent to amend customer policies. The agent authenticates, accesses the right systems and completes a change. But suppose it also retrieves unrelated customer records. Or it produces an incorrect policy outcome while operating within its permissions.

Security may identify unusual access without being able to judge whether the policy outcome was fair or correct. The business may understand the customer outcome without seeing the underlying data access. Both perspectives matter, and someone must be accountable for bringing them together.

My background in insider risk shapes this view. Legitimate access does not automatically mean appropriate activity. The same principle is relevant to agents, although an agent is not a human insider and an anomaly is not evidence of intent. We need to connect identity, activity, purpose and outcome before drawing conclusions.

Intervention: can the organisation act when it matters?

This is where governance is tested. If an agent is supporting an important customer process and its activity raises concern, can security suspend it? Does the decision belong to the business owner? What happens to transactions already queued, or to customers if the process stops?

Technical capability and decision authority are not the same thing. Nor does stopping an agent reverse actions it has already taken. Firms need defined escalation thresholds, containment options, recovery arrangements and clear decision rights, including outside normal working hours.

I prefer the term “human in the lead”. It means people define the purpose and boundaries, remain accountable for outcomes and retain a workable means of intervention. It does not mean a human must review every automated action. The oversight model should reflect the consequences and reversibility of the actions being delegated.

Five questions to take into your next governance meeting

Rather than commission another policy review, choose one material AI-enabled process. Bring its business owner, technology team, security and risk colleagues together, and ask:

1. What AI is actually operating? Compare approved inventories with embedded, third-party and locally introduced capabilities. Record where discovery or attribution is incomplete.

2. Can we reconstruct a real transaction? Trace the initiating request, agent or service identity, tools and systems used, information accessed, actions taken and business outcome. Be explicit about missing evidence and protect sensitive logs.

3. Who owns the activity and the outcome? Identify the accountable business owner, technical owner and operational responder. Test whether each understands the agent’s delegated authority and when to escalate.

4. Can we intervene safely and in time? Introduce an agreed deviation in a controlled environment. Check prevention, detection, escalation, containment and recovery, including any actions already in progress.

5. Are the controls still appropriate? Review changes to permissions, models, tools, integrations and business purpose. Ask when the controls were last tested and whether customer outcomes remain acceptable.

These are my proposed operational tests, not an FCA checklist. The output should be evidence: what was visible, what could not be established, who made the decision, how long intervention took and what needs fixing before the agent’s authority expands.

From good practice to evidence

Agent behavioural analytics (ABA) can contribute by corrolating observable identity and agent activity across available telemetry, highlighting first seen behaviour, deviations from established patterns and activity that warrants further investigation. However, It is not a substitute for AI discovery, access enforcement, model evaluation, business assurance or accountable decision-making. Those capabilities must work together.

The FCA publication gives firms a useful prompt to examine their arrangements. The practitioner challenge is to demonstrate that those arrangements work, not simply that they are documented.

Pick one agent already operating in your business. Demonstrate visibility, ownership and oversight, and intervention. Then ask whether the evidence justifies the authority you have given it.

Because approving an AI agent is a point-in-time decision. Being accountable for what it does is an ongoing responsibility.

Source - https://www.fca.org.uk/publications/good-poor-practice/ai-financial-services

*The insurance example is hypothetical. The practical tests are the author’s recommendations, not FCA requirements

Related Articles

More from Technology

Explore more articles in the Technology category