Key takeaways
- The Bank of England and FCA found that 80% of 196 surveyed firms have AI agents deployed, while only 2% report broad autonomy.
- Controls are front-loaded: more than 90% report pre-deployment controls, but only about half report ongoing monitoring.
- Third parties matter: 53% of high-materiality AI systems are third-party, and 54% of firms name third-party AI oversight as their top governance challenge.
- Nine controls close the gap: autonomy tier, permissions, tool limits, evaluation, logging, monitoring, rollback, vendor oversight and ownership. Check
- monitoring and vendor oversight first.
- The survey describes UK financial services practice. It is not a rule, but the gaps it describes apply to any enterprise running agents in sensitive workflows.
Most enterprises no longer debate whether to deploy AI agents. The Bank of England and the Financial Conduct Authority surveyed 196 firms for their fourth AI and machine learning report, published on 26 October 2026, and found that 80% of respondents already have AI agents deployed, with another 14% planning to within twelve months (Bank of England and FCA). The harder question is what evidence a firm can show once an agent is live.
This article takes the survey’s findings and turns them into a control checklist that technology and risk teams can apply to any agent in a regulated workflow. It is a practitioner view from GrowExx, not legal or regulatory advice, and the survey describes UK financial services practice rather than binding rules.
What the survey found
The report is a snapshot of self-reported practice, and a few of its numbers matter most for anyone running agents:
- Adoption is broad. 86% of firms use AI, up from 75% in 2024.
- Autonomy is limited. The report puts decision-support-only deployments at 80%, and broad autonomy at 2%. Even counting planned deployments, no more than 9% of firms expect agents to approve or decline customer outcomes, execute trades, initiate payments or change limits.
- Governance structures exist. 94% of firms have firm-wide AI policies and 84% have a central AI or model-risk governance forum.
- Controls are front-loaded. More than 90% report pre-deployment controls. About half report ongoing monitoring.
- Third parties are central. Third-party systems make up 39% of reported AI systems and 53% of high-materiality systems, up from 35% in 2024. The top three cloud and compute providers account for 79% of reported relationships, and the top three model API providers for 54%. Oversight of third-party AI was the most common governance challenge, chosen by 54% of respondents.
Policies and forums are in place. The thinner layer is evidence that an agent still behaves as approved months after launch, and visibility into the vendors it depends on.
Assess Your AI Agents Against Production Governance Requirements
Review agent permissions, human oversight, auditability, monitoring, and risk controls to identify gaps before expanding enterprise deployments.
Why launch-time review is not enough
An agent is more than a model. It is a model, a prompt, a set of tools, permissions, and the data it can reach. Any of those can change after approval: a vendor updates a model version, a team adds a tool, a data source changes its schema, a prompt is edited to fix one complaint and shifts behavior elsewhere. A pre-deployment review captures one moment. Monitoring is how a firm shows the approved configuration is still the running configuration.
For a risk owner, the practical test is simple. If a supervisor, auditor or board member asked how you know a given agent is behaving as approved today, could you answer with logs and metrics instead of a policy document?
A control checklist for agents in regulated workflows
| Control area | What good looks like | Evidence to keep |
|---|---|---|
| Scope and autonomy tier | Each agent is classified: suggests only, acts with human approval, or acts alone within limits. Customer-outcome actions default to human approval. | Autonomy register with owner and approval date |
| Permissions | Each agent has its own credentials, limited to the systems and actions its task needs. No shared service accounts. | Permission matrix; access reviews on a set schedule |
| Tool and data boundaries | A list of tools the agent can call and data it can read. New tools or data sources trigger review. | Tool inventory; change log |
| Evaluation | A test set drawn from real cases, run before release and again after any model, prompt or tool change. Pass thresholds defined in advance. | Evaluation results by version |
| Logging | Inputs, tool calls, outputs and human overrides are recorded in a form reviewers can search. | Retention policy; sample audit trail |
| Ongoing monitoring | Defined metrics (error rate, override rate, escalation rate, drift against the test set) with alert thresholds and a named reviewer. | Dashboards; alert history; review minutes |
| Rollback and stop | A tested way to disable the agent or revert to a prior version, with a clear owner. | Rollback test record |
| Third-party oversight | For each model and cloud dependency: contract terms on change notification, a fallback or exit plan, and a way to detect upstream model changes. | Vendor register; exit plan; change notices |
| Ownership | A named accountable executive per agent and a defined path for incidents. | Owner register; incident log |
The ongoing monitoring and third-party oversight rows address the survey’s two weakest signals, so check those first. For a deeper look at ownership, see our post on building an AI agent inventory.
Illustrative examples
These are hypothetical scenarios to show how the checklist applies, not descriptions of client work.
A claims-triage assistant. The agent reads incoming claims and drafts a priority recommendation. It has read access to the claims system and no ability to approve or deny. Its tier is “suggests only.” The monitoring metric that matters is how often adjusters override its priority, and a rising override rate triggers review of the latest model or prompt change.
A know-your-customer review assistant. The agent gathers documents and summarizes discrepancies for an analyst. Permissions are read-only on case systems. Every summary stores the source documents it used, so an analyst or auditor can trace any statement back to evidence.
In both cases the value is faster handling, and the control design is what keeps a human accountable for the decision.
What this means by role
CRO and model risk: Treat ongoing monitoring evidence as a requirement equal to pre-deployment validation, and ask for it by agent.
CIO and enterprise architecture: Make permissions, tool inventory and logging platform defaults so every new agent starts with them.
CISO: Review agent credentials and tool access like any privileged integration, and test the stop procedure.
COO and business owners: Name the accountable owner and agree what level of override rate triggers a pause.
Prepare AI Agent Deployments for Regulatory Scrutiny
Evaluate your AI systems against relevant regulatory expectations and establish evidence of risk management, human oversight, and operational accountability.
Conclusion
The Bank of England and FCA survey shows that AI agents are already in production at most surveyed firms, and that controls are concentrated before launch rather than after it. For enterprise teams, the practical response is to treat each agent as a system with an owner, an autonomy tier, scoped permissions, logging, ongoing monitoring and a rollback path. Start with the two weakest areas in the survey, monitoring after launch and third-party oversight, and collect evidence you can show an auditor or regulator. If you want help turning this checklist into working controls, our AI agent development team and AI consulting team can support that work.
GrowExx helps enterprises design, build and deploy AI agents with governance and monitoring built in. If you want an outside view of how your current agents would stand up against a checklist like this one, talk to our team about an AI agent governance and production-readiness review.
Common questions
Does the survey set rules for AI agents?
No. It is a regulator survey of firm practice, not a rule or guidance document.
Do these findings apply outside the UK or outside banking?
They are not legal requirements elsewhere. The control gaps they describe, monitoring after launch and third-party oversight, apply to any enterprise running agents in sensitive workflows.
Are fully autonomous agents common?
Not in this survey. Broad autonomy was reported by 2% of respondents.
Where should a team start?
Build an inventory of agents, assign an autonomy tier to each, and pick the two or three monitoring metrics you can start collecting this month.
Move Enterprise AI Agents Into Production With Confidence
Book an AI Governance Assessment
