UK financial firms have put AI agents into production almost everywhere and kept almost all of them on a short leash, so your agent autonomy policy is probably in line with your peers. Your oversight of vendor AI probably is not. That is the practical reading of the Bank of England and FCA's fourth survey of AI in UK financial services, published on 26 October 2026. Among the 105 firms that completed the full questionnaire, 80% have agents deployed and 2% let any of them act with broad autonomy. Meanwhile third-party systems now make up 53% of the AI that firms rate as high materiality, up from 35% in 2024, and 81% of firms oversee those systems mainly through vendor assurance and contracts.
If you run AI, model risk or technology at a bank, insurer or asset manager, this is the first regulator dataset you can benchmark against on both questions. Supervisors will use it the same way.
How Much Autonomy Do UK Financial Firms Give AI Agents?
Very little. The survey sorts agents into four autonomy levels, and adoption falls off steeply as you climb them. The report defines the levels like this:
- Decision support: the agent recommends, a human decides and executes.
- Constrained execution: the agent acts within predefined limits and rules.
- Conditional autonomy: the agent acts on its own but needs human approval for certain actions.
- Broad autonomy: the agent initiates and executes actions with limited real-time human involvement.
| Autonomy level | Respondents with agents at this level |
|---|---|
| Decision support | 80% |
| Constrained execution | 46% |
| Conditional autonomy | 22% |
| Broad autonomy | 2% |
The types of agent tell the same story. Enterprise agents (61%) and coding agents (52%) are the most common; client-facing agents sit at 16%, rising to 28% once planned deployments are counted. Where firms do grant more autonomy, it is mostly in operations and IT, risk and compliance, and sales and marketing. In operations and IT, 53% of respondents run decision-support agents and 1% run broadly autonomous ones.
The 2% figure has a history. The 2024 edition of the survey found that "only 2% of use cases have fully autonomous decision-making." The two numbers measure different things (share of use cases in 2024, share of respondents in 2026), so do not read them as a trend line. Both point the same way: almost every AI decision in UK finance still has a human signing it off.
Where Are Agents Allowed to Touch Customers and Money?
Customer communications are the frontier, and money movement is still off limits. Per the survey, 16% of respondents use agents for automated customer communications today, rising to 34% including planned use. Creating or changing records in core systems is at 8%, rising to 25%. For the actions that carry direct financial or conduct risk (approving or declining customer outcomes, executing trades or orders, initiating or blocking payments, changing limits and controls), no more than 9% expect to use agents even counting plans.
The steel-man for moving faster is real. The FCA is running agentic payments through its AI Live Testing programme, whose second cohort, named on 21 April 2026, includes Barclays, Lloyds Banking Group (via Scottish Widows) and UBS. Testing ends this year and an evaluation report is due in Q1 2027. Firms in the cohort will have supervised evidence on agentic payments before their peers do.
What that buys them is a sandbox result. Production autonomy over payments is still a 9%-or-less plan across the sector. The same pattern showed up at Sibos this year, where bank agents repair payments that a person still releases. The record-changing number deserves more of your attention than the payments one. An agent that writes to a core system at 25% of firms is a model risk and data integrity question your validators will have to answer within a year.
Vendor AI Now Runs Most High-Materiality Systems
The bigger shift in this edition is who builds the AI. Third-party implementations are 39% of all reported AI systems, up from 33% in 2024 and 17% in 2022, and 53% of high-materiality systems. For the systems that matter most, firms now buy more than they build.
Concentration rose with it. The three largest cloud and compute providers account for 79% of reported relationships, up from 72% in 2024 on the report's own comparison, and the top three model and API providers hold 54%. The report's chart puts the 2024 comparator for model providers at 41%; the 2024 survey itself reported 44% on a different basis. Trade coverage from Resultsense put it plainly: "the third-party figures are the ones supervisors will act on."
The oversight gap is visible in how firms say they understand these systems. For high-materiality third-party systems, 81% rely on vendor assurance and contractual arrangements. For in-house systems, 77% rely on deep internal knowledge. When firms were asked to name governance challenges, 54% picked applying model risk management to vendor models and 49% picked overseeing third-party systems.
The Bank has been flagging this concentration for over a year. The Financial Policy Committee's April 2025 report on AI in the financial system warned that "reliance on a small number of providers for a given service could lead to systemic risks." The critical third parties regime (PS16/24, in effect from 1 January 2025 and applying to a provider once HM Treasury designates it) addresses the providers. As we covered in August, that regime reaches the clouds and leaves the models inside them to you.
Vendor Model Updates Are the Thinnest Control
The weakest numbers in the report are about what happens when a vendor changes its model. Only 48% of respondents always run regression tests on high-materiality updates to vendor-provided models or their guardrails, per the survey. 73% always or usually have a defined rollback plan. And 25% always or usually let high-materiality vendor updates propagate into production automatically. Only 48 firms answered these questions, and the regulators say the results "should be treated with caution."
Read the last figure again anyway. A quarter of the firms that answered let a supplier change the behaviour of a system they rate as high materiality without a gate. Enhanced resilience arrangements such as joint incident testing or exit planning for third-party AI are reported by 27% of firms.
The broader control picture is uneven in the same way. The executive summary reports that human approval, pre-deployment review and vendor due diligence are each in place at more than 90% of respondents, while ongoing monitoring, leakage testing and periodic review or recertification are each reported by around half. The checks that run once, before deployment, are close to universal; the ones that have to keep running are at half.
This is where existing rules already bite. The PRA's model risk principles, SS1/23, apply to models "whether developed in-house or externally (including vendor models)" for the banks in scope. The outsourcing statement SS2/21 has a revised version effective 18 March 2027. A vendor update that skips regression testing is hard to square with either.
What the Survey Does Not Tell You
The data has limits you should state when you quote it. The regulators say the results are evidence from participants rather than the whole sector, and the 2026 and 2024 samples differ in composition and question scope, so year-on-year comparisons are approximate. Most headline figures come from 105 full respondents out of 196. The ten firms reporting the most systems account for 43% of all systems, so a handful of large banks shape the totals.
The survey also does not say how firms measure agent outcomes, or what incidents they have had. For that you need your own logs. US supervisors are asking similar questions: the Fed, OCC and FDIC have asked banks to explain their agent safeguards, and Singapore's MAS now requires an AI inventory that reaches into SaaS tools. A firm supervised in more than one of these places should build one evidence pack that answers all three.
What to Do Before Your Supervisor Asks
Benchmark yourself on the two numbers supervisors will read first: agent autonomy by level, and how you control vendor model changes.
This Week:
- Map every production agent to the survey's four autonomy levels. If you have anything at broad autonomy, you are in the 2%; write down who approved it and on what evidence.
- List agents that create or change records in core systems. That is 8% of firms today and 25% planned, and it is the category your validators have least practice with.
This Month:
- Pull your high-materiality AI inventory and mark each system as in-house or third-party. If vendors are more than half, you match the sector, and your model risk team needs to show how it validates models it cannot inspect.
- For each high-materiality vendor model, record whether updates reach production automatically. If any do, add a regression test and a rollback plan before the next release. Ask the vendor for advance notice of model changes in writing; our AI vendor security review questions cover the contract language.
Before Year-End:
- Move vendor oversight beyond assurance letters. Add ongoing monitoring of vendor model outputs against your own test set, which only around half of respondents run.
- Put exit planning or joint incident testing in place for your top AI providers. Only 27% of firms report it, and the revised SS2/21 takes effect in March 2027.
The Bottom Line
On agents, UK finance has chosen supervised automation: 80% of respondents use agents for decision support and 2% allow broad autonomy. On vendors, the build-versus-buy question has been answered by default. Most high-materiality AI is now bought, and most firms oversee it through vendor assurance and contracts. Banks went through the same shift with outsourced IT, and the PRA answered it with SS2/21, with its chapters on data security and exit plans. A survey this specific about vendor model updates suggests vendor AI is next in line for that kind of scrutiny.
Run the vendor update inventory first. That 25% auto-propagation figure, small sample and all, is the number most likely to come up in your next supervisory meeting.
Continue Reading
- Britain Regulated Four Clouds. Not the Models Inside.
- US Regulators Ask Banks to Explain Their AI Agent Safeguards
- MAS Makes Financial Firms Inventory AI Across Their SaaS Tools
- Barclays Picks Claude for Half Its Developers, Copilot for Desks
- At Sibos 2026, Bank AI Agents Fix Payments a Person Still Releases
- AI Vendor Security Review: 6 Questions That Change the Answer