As an ex-Londoner now living in Sydney I always look on with interest at my ex-clients and one of those the Bank of England and FCA have issued the results of a survey which should be read by all Australian Boards
OK take a seat. If you and I were sitting down for a flat white this morning, this is the thing I would want to talk to you about.
I have lived in Sydney for a eight years, but as an ex-Londoner I still keep one eye on Threadneedle Street, a stone throw from my offices whilst in London. Not out of nostalgia (well, not only), but the UK regulators tend to publish hard data on where financial services is heading, and in my experience the questions they ask in London have a habit of turning up in the Australian Boardrooms a year or two later, so as you can expect they can be very helpful to me.
On 26 October 2026, the Bank of England and the FCA published their fourth survey of AI in UK financial services. 196 firms responded. I’ve read it end to end, and the headline everyone is quoting is not the one that worries me.
The headline, lots of agents and very little autonomy
The numbers first, because they’re striking.
86% of firms now use AI, up from 75% in 2024.
Generative AI has gone from a sideshow to the majority, accounting for 56% of all reported AI systems.
So, agents are everywhere. 80% of firms in the full survey have AI agents in production, and another 14% plan to deploy them in the next twelve months.
But here’s the interesting bit. The survey breaks agents down by how much freedom they’re given, and it reads like a ladder:
- Decision support (the agent recommends, a human decides and acts): 80% of firms
- Constrained execution (the agent acts within predefined rules and limits): 46%
- Conditional autonomy (the agent acts independently, but needs approval for certain actions): 22%
- Broad autonomy (the agent initiates and executes with limited real-time human involvement): 2%
Even looking twelve months ahead, no more than 9% of firms expect agents to approve or decline customer outcomes, execute trades, move payments, or change limits and controls.
Honestly? Sure, that’s sensible, so UK firms are letting agents help, not decide. If the conversation stopped there, I’d say the sector is behaving itself.
But, it doesn’t stop there.
The real story is the most important AI is being bought, not built
This is the bit I would lean across the table for.
Third-party systems now make up 39% of all AI in UK financial services (17% in 2022, 33% in 2024).
More importantly, they now account for 53% of high-materiality AI, up from 35% two years ago. The systems that matter most are increasingly someone else’s. Now we can sit less comfortably in our seats.
And look at how firms say they understand those systems. For AI they built themselves, 77% rely on deep knowledge of how it works. For third-party AI, that drops to 24%. Instead, 81% rely mainly on vendor assurance and contracts.
Put plainly, for more than half of their most material AI, most firms are governing on the vendor’s word.
Then add concentration. The three largest cloud and compute providers account for 79% of all reported relationships, and 98% of firms use at least one of them. Of the cloud relationships firms themselves rate as highly critical and hard to substitute, only about a third have tested exit plans.
Now, I spent a big part of my career in and around Data Centres, including running a colocation facility. I know the difference between a customer who reads your assurance pack and one who comes in and tests your failover. When something goes wrong, only one of them can show a regulator what they actually checked.
The gap, governance that stops at go-live
The second thing that jumped out is when the controls happen.
Before deployment, UK firms look strong. Human approval, pre-deployment review and vendor due diligence are each reported by more than 90% of firms.
After deployment, it falls away. Ongoing monitoring, data leakage testing and periodic review or recertification are each reported by only around half. Fewer than half have AI incident management and response playbooks.
What firms monitor is telling too. 85% track user feedback, but only 39% track customer impact, and 34% track risk metrics.
Then there’s vendor change. A quarter of the (admittedly smaller) group who answered this question said they always or usually allow high-materiality vendor updates to flow straight into production.
Think about what that means.
You approved a model in March, your vendor quietly updates it in September, which impacts its behaviour changes, so your approval is now for a system you are no longer running.
It is no surprise that the most common governance headache in the survey is applying model risk management to vendor-supplied models (54% of firms), followed by overseeing systems provided or influenced by third parties (49%). Around two in five also struggle with independent validation, drift monitoring and change control.
Governance that stops at deployment, and relies on a vendor’s word, will be fine right up until the first serious incident. Then it won’t satisfy anyone.
Why this matters in Sydney, not just London
If you’re on the Board or executive team of an Australian bank, insurer or super fund, this pattern should look familiar. Most of the AI coming into Australian financial services is arriving inside platforms, SaaS products and cloud services, not built in-house.
We also already have the hooks. APRA’s CPS 230 has put material service providers, operational resilience and credible plans for disruption squarely on the Board’s agenda.
CPS 234 covers information security across third parties. ASIC’s 2024 review of AI use by financial services licensees warned that governance was lagging behind adoption.
None of that was written with AI agents in mind, but all of it applies to them. I’ woul treat the UK survey as a preview of the questions APRA and ASIC are likely to ask, and I’d much rather a Board had the answers before the questions arrive.
Five questions I would ask at the next Board meeting
If I were sitting on your Board, these are the questions I would put on the table:
- Do we have a complete AI inventory? Including the AI embedded in vendor products we never bought as “AI”. In the UK, only 68% of firms have a central inventory in place.
- Which systems are high-materiality, and who owns each one? A named owner, not a committee.
- When a vendor changes a model, how do we find out, and what testing happens before it reaches customers? If the answer is “the contract says they’ll tell us”, that’s a gap.
- Can we roll back, and have we ever actually tried? A rollback plan that’s never been exercised is a hope, not a control.
- What is our exit plan from our key cloud and model providers, and when did we last test it? Concentration risk doesn’t go away because it’s inconvenient to plan for.
And one bonus question: what are we monitoring after go-live beyond whether staff like the tool?
Where BAIGS® fits
This is exactly why we built BAIGS®, the Bushey AI Governance Spine. The idea is simple: governance shouldn’t be a gate you pass through once. It should run the full length of an AI system’s life, from intake and materiality assessment, through approval and deployment, into monitoring, change control, incident response and eventual retirement.
Crucially, BAIGS® treats third-party models as first-class citizens. If a vendor’s model is making or shaping material decisions, it needs an owner, evidence of how it behaves, controls over updates, and a way out. A signed contract is part of that evidence, not the whole of it.
The takeaway
The UK data tells a reassuring story about agent autonomy and a less reassuring one about everything else. Firms are cautious about what agents are allowed to do, while quietly outsourcing their most material AI and relying on vendors to keep it honest after go-live.
Australia has the advantage of seeing this a step ahead, and we can use it.
If you’re working through any of this in your own organisation, I’m always happy to have that coffee for real. Drop me a message.
Summary
The latest Bank of England and FCA survey shows UK financial firms are cautious about AI agent autonomy, with 80% running agents but only 2% giving them broad autonomy.
The bigger risk is that 53% of their most material AI now comes from third parties, governed mainly on vendor assurances and with weak oversight after go-live. Australian Boards should treat this as a preview of the questions APRA and ASIC are likely to ask.
About Bushey’s AI Services
Bushey is a specialist boutique consultancy, helping Corporate and Government organisations across APAC use AI with control, confidence and measurable business benefit. Our Artificial Intelligence Practice delivers:
- BAIGS® – Bushey AI Governance Spine. Clear decision ownership, approvals, risk evidence and a safe path to production for every AI use case and agent.
- BAIGS® AI Governance Assessment. An independent view of where your AI governance stands today, with the gaps prioritised and a practical roadmap to close them.
- AI Readiness and Implementation. Taking AI from experiment to production, delivered through AssureChange®, our Transformation delivery framework that brings together Project Management, Change Management and Governance to deliver real business benefits.
- Fractional Chief AI Officer. Senior AI leadership for your executive team and board, without the cost of a full-time hire.
- AI and Cybersecurity. Through our Bushey Cybersecurity Practice, we help secure the systems, data and agents your AI touches, and prepare your incident response for mandatory AI reporting.
Ready to put a chain of accountability around your AI agents? Visit bushey.com.au to book a conversation.
#AIGovernance #FinancialServices #APRA #ThirdPartyRisk #BAIGS

Comments are closed