Firms are cautious about the risk they can see
UK financial firms answering the regulators’ survey have taken AI agents into production and then kept them on a short lead. In the Bank of England and Financial Conduct Authority (FCA) survey published on 26 October, only 2% of respondents reported agents running with broad autonomy. The same survey shows a quieter shift running the other way: vendors now supply 53% of the AI systems firms rate as high materiality, and for those systems most firms rely on the vendor’s assurances rather than deep knowledge of their own. Our news report on the survey covers the headline results. This analysis looks at what the two trends mean side by side.
Strategic Insight: In our reading, firms are applying their caution where autonomy is visible, inside their own agent deployments. The dependency that is growing sits in systems they did not build and, by their own account, know far less well.
What does the survey measure, and what does it not?
This is the fourth edition of the joint survey, after 2019, 2022 and 2024, and it drew 196 responses. Unless the report says otherwise, its figures come from the 105 firms that completed the full questionnaire. The authors are explicit that the results describe participating firms, not the entire UK sector, and that comparisons with 2024 need care where wording, definitions or the respondent mix have changed.
That caveat matters for the trend lines. It matters less for the shape of the picture within 2026, which is where the more useful findings sit.
The real story is in the follow-up questions
The headline numbers say agents are common and autonomy is rare. The follow-up questions say something more specific about how firms govern what they run. Asked how they reach a sufficient understanding of high-materiality systems, 77% of respondents cited deep knowledge of systems built in-house. For third-party systems that figure fell to 24%, and 81% leaned instead on what the vendor assured them and on their contracts.
Put that next to the materiality data and the gap widens. Third-party systems were 39% of the AI systems firms reported overall but 53% of the high-materiality ones, against 35% in 2024. High-materiality systems are more likely than the average system to be bought in, and bought-in systems are the ones firms know mainly through the supplier’s account of them.
| Metric | Value | Strategic implication |
|---|---|---|
| Respondents with agents at broad autonomy | 2% | Very few firms let agents act with little real-time human involvement |
| High-materiality systems from third parties | 53% (35% in 2024) | More than half of high-materiality AI is now bought in |
| Deep knowledge of high-materiality systems: in-house vs third-party | 77% vs 24% | Far fewer firms claim deep knowledge of vendor systems |
| Critical, hard-to-substitute cloud relationships with tested exit plans | 35% | Most of these cloud dependencies lack a tested exit plan |
Strategic Reality: The top three cloud and compute providers held 79% of reported relationships, and 98% of respondents used at least one of them. The concentration is not a feature of a few firms’ choices. It is close to the whole sample.
Where the caution sits, and where it does not
Autonomy is being rationed
The survey grades agent deployment on four steps. At the bottom, agents recommend and people decide: 80% of respondents run agents this way. Then come constrained execution within set rules (46%), conditional autonomy with human sign-off for certain actions (22%) and broad autonomy (2%).
Firms are also clear about what agents will not be doing soon. Even including what they plan over the coming year, at most 9% expect agents to decide customer outcomes, place trades, handle payments or alter limits and controls. Those are the actions with direct consequences for customers and markets, and very few firms expect to hand them to agents within a year.
Critical Context: The survey’s definition of an AI agent includes some independence of action. Its results show the most common deployment is the least autonomous form, with agents recommending and a person deciding.
The controls are strongest before launch
The second pattern is timing. More than nine in ten respondents use human sign-off, a review before launch and due diligence on vendors. Only about half keep up live monitoring, test for data leakage or periodically re-review and recertify their systems. Fewer than half have a playbook for responding to incidents.
Monitoring leans towards user feedback rather than downstream effects. User feedback was the metric most routinely tracked, by 85%. Customer-impact measures were monitored by 39%, and risk metrics by 34%.
In our view, this goes some way to explaining the autonomy figures. A firm that checks a system thoroughly at the gate, and less often once it is running, has good reason to keep a human approving each consequential action. Wider autonomy would need the controls that this survey shows are less common.
Vendor updates can bypass the gate
The front-loaded pattern has a specific weak point with bought-in systems. A quarter of respondents said they always or usually let high-materiality vendor updates flow automatically into production. Only 48 firms answered that question, so the authors warn against reading too much into it. Even so, a review carried out before deployment cannot cover a change the vendor makes afterwards.
⚠️ Warning: Firms reported 169 separate cloud and compute relationships. They rated 37% of them as both highly critical and difficult to replace, and just 35% of those had an exit plan that had been tested. Only 27% said they hold AI suppliers to tougher resilience requirements, such as testing incidents together or planning an exit.
Who carries the exposure?
The exposure is unevenly spread. Just ten firms held 62% of the high-materiality systems reported, and international banks and broker-dealers reported the highest median number of systems in production, at 43, ahead of UK banks at 22.
Smaller firms sit at the other end. In the short-form questionnaire offered to firms employing under 100 people, 73% reported using AI, and 58% of those mainly used general-purpose large language models. The survey does not say who supplies those models.
Firms expect this dependency to become a bigger concern. Asked about systemic risks, the share rating critical third-party dependencies as high or severe rose from 40% today to 63% within the next three years. Only cyber risk amplification scored higher.
| Group | What the survey shows for them | What is still unknown |
|---|---|---|
| The largest AI users | Ten firms hold 62% of high-materiality systems; international banks and broker-dealers report the highest median system count | How much of that concentration is vendor-supplied |
| Smaller regulated firms | 73% use AI, most of them mainly general-purpose language models | Whether they have any route to vendor transparency |
| Cloud, compute and model providers | Three cloud providers hold 79% of relationships; three model providers hold 54% | How regulators will treat AI services within those relationships |
What this means for firms that buy AI rather than build it
For a firm whose high-materiality AI comes from a supplier, the survey points to a specific problem rather than a general one. Extending model risk management to models supplied by vendors was the governance challenge respondents named most often, at 54%, followed by oversight of systems supplied or materially shaped by third parties, at 49%. Limited transparency into vendor systems was a major constraint for 34%.
Yet vendor assurance and contracts are the method 81% of respondents rely on to understand bought-in high-materiality systems. In our view, the challenge firms report and the method they rely on are the same issue seen from two sides.
Reality Check: More than nine in ten respondents carry out vendor due diligence, and the survey groups it with the checks made before deployment. The weaker numbers sit in what happens afterwards: monitoring, exit testing and control over updates.
The open questions
Can vendor assurance stand in for understanding?
Most firms say contracts and supplier assurances give them a sufficient understanding of third-party systems. Counting moderate as well as major constraints, firms ranked trouble securing contractual protections, such as audit rights, liability terms and warning of changes, second among third-party challenges. If the contract is doing the work of understanding, and the contract is hard to negotiate, the question of how much firms really know about these systems stays open.
Will autonomy grow before post-deployment controls do?
The 12-month plans show growth across every agent type, with the largest increase expected in multi-agent systems. Post-deployment controls, by contrast, reach only about half of respondents today. Whether firms widen autonomy first or strengthen live monitoring first is something the next edition should be able to show.
How far can trend comparisons be trusted?
The 2026 survey changed its questions and its respondent mix, and introduced a short form for smaller firms. Some changes, such as the shift towards generative AI, compare measures the authors themselves describe as not directly comparable. Readers should treat the direction of travel as more reliable than the size of any year-on-year change.
What to watch
The Bank and the FCA place this survey within wider work that includes critical third parties, operational resilience and model risk management. Three things would show whether the vendor-dependence finding is shaping that work.
- The Bank’s next system-wide exploratory scenario. Bloomberg reports, citing one person with knowledge of the discussions, that AI-driven turmoil is one candidate for the next scenario, while a second source said nothing has been decided. We covered the report in our news piece on the stress-test options. A scenario built around shared AI suppliers would indicate that supervisors see vendor concentration as a system-wide risk, not just a firm-level one.
- Third-party practice against the Financial Stability Board’s guidance. The FSB’s final Sound Practices for Responsible Adoption of Artificial Intelligence include handling dependence on third-party AI. Our coverage of the final practices sets out the full list. A rise in tested exit plans and joint incident testing would be a sign of firms taking that practice on.
- The next edition of this survey. The measures to compare are the 35% of critical, hard-to-substitute cloud relationships with tested exit plans, the 24% with deep knowledge of high-materiality vendor systems, and the 9% ceiling on agents taking consequential actions. If autonomy rises while the first two stay flat, firms will have widened the risk without widening the oversight.
The report’s authors say several findings, notably about “uncertainty about best practice, third-party dependencies and readiness for more autonomous AI”, are consistent with the July 2026 Financial Services AI Adoption Plan from the government’s independent AI Champions for the sector. Firms are already holding back autonomy. The harder test is whether they can see far enough into the systems they buy to know when the same caution is needed there.
Strategic Insight: The agent debate asks how much freedom to give AI. For most of the firms surveyed, the results suggest the more pressing question is how much they know about the AI they already depend on.
Source: Artificial intelligence in UK financial services - 2026 (Bank of England and Financial Conduct Authority, 2026), written by a joint team from both regulators. Additional sources: FSB finalises sound practices for the responsible adoption of AI (Financial Stability Board, 2026); Bank of England Considers Stress Testing AI’s Effect on the Financial Sector (Bloomberg via Claims Journal, 2026), by Tom Rees and Laura Noonan.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching the Bank’s choice of its next system-wide scenario, and the next edition of this survey, for signs that oversight of vendor AI is catching up with reliance on it. Read more analysis at Insights, or get in touch.