Most firms that answered the latest regulator survey of UK finance now run AI agents. Very few let those agents act on their own. The Bank of England and FCA survey of AI in UK financial services, published on 26 October 2026, found that 80% of full-survey respondents had AI agents deployed and a further 14% planned to deploy them within twelve months. Only 2% reported agents deployed with broad autonomy. The same report puts it plainly: "AI agent deployment is widespread, but autonomy is limited."
That headline hides the more useful finding. The survey splits agent autonomy into four levels, and the middle two are not empty. 46% of respondents already run agents that execute actions inside predefined limits, and 22% run agents that act independently but need human approval for certain actions. So agents are acting in UK finance today. The line that almost nobody has crossed is narrower: even counting plans for the next year, no more than 9% expect agents to approve or decline customer outcomes, place trades, move or block payments, or change limits and controls. This guide is for the COO or Chief Data and AI Officer at a bank, insurer or lender who has to decide what changes in controls, ownership and evidence before an agent moves one rung further up that ladder, and why most firms should not rush the last rung.
The scope is deliberately narrow. Live monitoring and the evidence file a supervisor expects are covered in our guide to FCA AI live testing evidence; senior manager accountability sits in the SM&CR phase 2 AI accountability guide; the regulatory constraints the same survey measured are in our 2026 survey guide to AI regulatory constraints; third-party and cloud dependency belongs to the self-hosted private LLM guide; and the mechanics of an approval queue are in our human approval gate design guide. This post holds the autonomy ladder itself.
This guide is shipped by the team behind DATS, the five-stage AI consulting system from DILR.AI, delivered by senior practitioners who ship code, not decks. Or see the AI operating model, which covers governance, RACI and lifecycle, audit-ready by design.
What did the 2026 Bank of England and FCA survey find about AI agents?
The 2026 Bank of England and FCA survey of AI in UK financial services found that AI agents are already common among responding firms and are most often used to recommend rather than to act. 80% of full-survey respondents reported agents deployed and 14% more planned them within a year, while only 2% reported agents acting with broad autonomy and limited real-time human involvement.
The base matters before any of these figures is repeated in a board paper. The survey received 196 responses, 101 from larger firms and 95 from smaller firms, and 105 respondents completed the full questionnaire. Unless the report says otherwise, the agent figures come from that full survey, calculated on valid answers to each question. The headline that 86% of respondents use AI, up from 75% in 2024, is a different base again, because it combines the full and short-form surveys. The report itself warns that its findings are evidence from participating firms, not a complete assessment of AI adoption across the whole UK financial sector. A firm benchmarking itself should say "respondents", not "UK banks".
Within that base, the shape of agent use is clear. Enterprise and coding agents were the most widely deployed types, and client-facing agents were less common. Generative AI accounted for 56% of all reported AI systems. That matters for the ladder. It suggests much agent use sits inside the firm, where staff can check a recommendation before anything reaches a customer. If you are still deciding where agents belong at all, the DATS pillar guide to enterprise AI consulting covers placement before autonomy.
What are the four levels of AI agent autonomy in the survey?
The Bank of England and FCA survey defines four levels of AI agent autonomy. Decision support means agents recommend and humans decide and execute. Constrained execution means agents act within predefined limits and rules. Conditional autonomy means agents act independently but need human approval for certain actions. Broad autonomy means agents initiate and execute actions with limited real-time human involvement.
The wording of each level is worth keeping exactly, because it is the vocabulary a supervisor will use back to you. In the survey's terms, decision support is where "agents recommend actions, humans decide and execute". Constrained execution is where "agents execute actions within predefined limits and rules". Conditional autonomy is where "agents act independently but require human approval for certain actions". Broad autonomy is where "agents can initiate and execute actions with limited real-time human involvement".
The four shares are not a pie. A firm can run a decision-support agent in compliance and a constrained-execution agent in IT operations at the same time, so it counts at both levels, and the four figures add up to well over 100%.
Read the chart as a staircase, not a ranking. Nearly half of respondents have already taken the first step from recommending to acting, inside written limits. About one in five has taken the second, letting an agent act on its own except where a named action triggers an approval. Almost nobody has taken the third. The interesting management question is therefore not "should we let agents act", which most respondents have already answered, but which actions sit on which rung, and what has to be true before an action is promoted.
Where are banks and insurers giving agents more autonomy?
In the Bank of England and FCA survey, the business areas most likely to run AI agents at higher autonomy levels were Operations and IT, Risk and Compliance, and Sales and Marketing. Rows tied to customer accounts and money movement, such as retail banking, payments and general insurance, show low single-digit shares above decision support, so few respondents reported agents acting there.
Table B of the report gives the share of all full-survey respondents reporting agents deployed in each business area at each level. Seven of its rows are shown here:
| Business area | Decision support | Constrained execution | Conditional autonomy | Broad autonomy |
|---|---|---|---|---|
| Operations and IT | 53% | 34% | 14% | 1% |
| Risk and compliance | 30% | 10% | 3% | 1% |
| Sales and marketing | 22% | 9% | 5% | 0% |
| Commercial banking | 11% | 6% | 1% | 0% |
| Retail banking | 8% | 3% | 1% | 1% |
| Payments | 7% | 5% | 2% | 0% |
| General insurance | 3% | 3% | 1% | 0% |
Two cautions apply. First, each cell is a share of all respondents, not of firms in that line of business, so "retail banking 8%" does not mean 8% of retail banks; it means 8% of respondents reported a decision-support agent deployed in a retail banking function. Second, small cells are small numbers of firms, so a 1% in a row is a handful of respondents and should not be read as a trend.
Read with those cautions, the rows suggest a tendency rather than a rule. Sales and marketing is customer-facing yet carries more conditional autonomy than risk and compliance, so the dividing line is not simply internal versus external. What the low rows share is that an agent action there would touch a customer's account or move money. An IT operations agent that restarts a service or reroutes a ticket takes an action whose effect stays inside the firm. A retail banking agent that changes what a customer receives takes a different class of decision, which is why the banking hub on where AI pays in UK banks starts with internal workloads rather than customer outcomes.
Which actions are firms still keeping away from AI agents?
UK financial firms are still keeping AI agents away from decisions that change a customer's position or move money. The 2026 Bank of England and FCA survey found that, even including plans for the next twelve months, no more than 9% of respondents expected agents to approve or decline customer outcomes, execute trades, handle payments, or change limits and controls.
The contrast with the actions firms are willing to delegate is sharp. 16% of respondents already use agents for automated customer communications, rising to 34% once planned deployment is included. For creating or modifying records in core systems, the figures are 8% today and 25% including plans. Those are actions with consequences, but they are actions a firm can see and correct, even where a sent message cannot be recalled. Approving a loan, declining a claim, blocking a payment or moving a credit limit changes the customer's position immediately, and some of those changes cannot be cleanly undone.
Three things make that last group different, and they explain the 9% ceiling better than caution alone:
- The action is the decision. When an agent drafts a decline letter for a human to send, the human made the decision. When an agent declines the application, there is no later point where a person decided anything.
- The harm lands on the customer first. A wrong ticket closure costs the firm an hour. A wrong payment block can cost a customer their rent, and the FCA Consumer Duty asks the firm to evidence good outcomes for that customer, not good intentions about the model.
- The law starts to apply. A decision with a legal or similarly significant effect that is taken without meaningful human involvement is a solely automated significant decision under UK data protection law, with its own safeguards. That is covered below and in more depth in our Article 22A automated decision-making guide.
What has to change before an AI agent moves from recommending to acting?
Before an AI agent in a bank or insurer moves from recommending to acting, three things have to change: the limit the agent works inside must be written as a rule a system can enforce, a named person must own that limit, and every action must leave a record that shows what the agent did and how to reverse it. Each further rung adds to those three.
From decision support to constrained execution. At the bottom rung the human is the control, so weak agent controls are survivable. The moment the agent executes, the control has to move into the system. In practice that means the predefined limits the survey describes are machine-checked: an allow-list of actions, value caps, the systems the agent may write to, and the conditions under which it must stop. It means one person owns each limit and signs off changes to it, in the same way a credit limit or a payment threshold has an owner. And it means every action is logged against the agent's identity with enough context to reverse it. The AI operating model is where those owners and limits are written down, not in the agent's prompt.
From constrained execution to conditional autonomy. Here the agent stops working inside a fixed box and starts choosing its own actions, with a human approval required only for named classes. The design work moves to the classification: which actions are pre-approved, which need sign-off, and what the approver sees. The approver must be able to change the outcome, not just watch it, because an approval that cannot say no is not a control. The mechanics of the queue, the approver's view and how to stop approvals turning into a rubber stamp are covered in the approval gate design guide.
From conditional autonomy to broad autonomy. The last rung removes the routine human check, so the evidence has to come from monitoring instead. This is where the reported controls thin out. Across AI generally, human approval, pre-deployment review and vendor due diligence were each reported by more than 90% of respondents as data risk controls, while ongoing monitoring, leakage testing and periodic review or recertification were each reported by only around half. Those figures cover AI data risk controls generally, not agents specifically, but they point the same way: firms report the controls that run before deployment far more often than the ones that would replace a human reviewer afterwards. Our guide to live AI monitoring evidence covers what that monitoring file contains.
The same diagnostic logic underpins the six named solutions on our enterprise AI solutions page, which include evaluation and observability for production agents.
How do UK rules apply to an AI agent that acts on its own?
No single UK rule governs AI agent autonomy in financial services. Instead, existing duties bind different parties: the FCA Consumer Duty binds regulated firms on customer outcomes, UK GDPR Article 22C binds the data controller where a significant decision is solely automated as defined in Article 22A, and PRA SS1/23 binds banks on model risk management. An agent moving up the ladder engages more of them.
The data protection test is the sharpest, because it turns on exactly what changes at the top of the ladder. Article 22A, inserted into the UK GDPR by the Data (Use and Access) Act 2025 and in force since 5 February 2026, says that a decision is solely automated where "a decision is based solely on automated processing if there is no meaningful human involvement in the taking of the decision". Where such a decision is significant for the person, Article 22C requires the controller to put safeguards in place, including letting the person make representations, obtain human intervention and contest the decision. The duty binds the controller, which is the bank or insurer, not the agent vendor.
The test applies to each decision, not to the rung. Decision support keeps a human in the decision. But a constrained-execution agent that approves or declines within its limits, or a conditionally autonomous agent acting outside its approval class, can take a significant decision with no meaningful human involvement on a lower rung too. What climbing the ladder changes is how often that happens: the higher the rung, the more significant decisions are likely to be taken without meaningful human involvement, and the more weight the Article 22C safeguards have to carry.
Model risk is the second lens. The PRA's SS1/23 model risk management principles for banks, effective since 17 May 2024 with a revised version effective from 23 April 2026, sets expectations for model inventory, validation and ongoing monitoring at PRA-regulated banks. Whether a given agent falls inside a bank's model definition is a judgement each bank makes against its own framework, but an agent that executes actions is a strong candidate for the inventory, and the survey found that applying model risk management principles to third-party or vendor-supplied models was the most commonly reported area of governance uncertainty, selected by 54% of respondents.
Accountability is the third lens and is deliberately ceded here: who signs for an agent that acts is the subject of our SM&CR phase 2 guide.
Why should most firms not rush to broad AI agent autonomy?
Most UK financial firms should not rush AI agents to broad autonomy because much of the preparatory work an agent does is available on the lower rungs, and the post-deployment controls the top rung depends on are the ones firms report least often. Broad autonomy over customer outcomes adds legal safeguards, monitoring demands and reputational exposure.
Consider where the work sits in a typical case. If an agent drafts the suspicious activity narrative, assembles the KYC refresh pack or classifies the complaint, it has done the preparatory labour. Whether a person or the agent then presses the final button changes the risk profile more than it changes the preparation time. A firm that moves its best agents from decision support to constrained execution, with enforced limits and an action log, can capture much of that labour without handing customer-outcome decisions to the agent.
The second reason is visibility. For high-materiality systems developed in-house, 77% of respondents relied on deep knowledge of how the system works; for third-party systems that figure was 24%, with 81% relying instead on vendor assurance and contractual arrangements. The survey does not split agents by provider, but third-party implementations made up 39% of all reported AI systems. Where an agent runs on a third-party model, giving it broad autonomy means relying on a system the firm mainly understands through a contract.
The third reason is that broad autonomy is easy to slide into rather than choose. An agent promoted to conditional autonomy with an approval class that is too narrow is effectively autonomous for everything outside the class. That is why the promotion decision belongs in a governed operating model with a named owner rather than in a configuration change.
What is the best way to govern AI agent autonomy in a UK bank in 2026?
The best way to govern AI agent autonomy in a UK bank in 2026 is to assign every agent action, not every agent, a rung on the four-level ladder from the Bank of England and FCA survey, with a named owner, an enforced limit and a reversal path for each. Promote one action class at a time, and only with monitoring evidence for the rung above.
Four criteria separate a workable approach from a policy document:
- Action-level classification. Autonomy is a property of what the agent does, not of the agent. The same agent may draft freely, update records inside limits and need approval to touch a customer outcome.
- Enforced, owned limits. Every limit is checked by a system, not stated in a prompt, and has one owner who approves changes to it.
- Reversal before promotion. No action class moves up a rung until the firm has shown it can detect a wrong action and reverse it within a stated time.
- Monitoring that replaces the human. At the top rungs, ongoing monitoring has to do the job a reviewer used to do, so it must exist and be tested before the reviewer is removed.
There is more than one route to that. Large consultancies such as Deloitte, PwC, EY and Accenture bring established model risk and controls practices, and for a bank that needs a firm-wide framework rolled out across many business units at once, or an assurance opinion its audit committee already recognises, that scale is a genuine advantage. Specialist firms such as Faculty bring deep data science benches. Where a firm needs a small number of agent placements governed and shipped into production on its own systems, a senior-led delivery team that writes the controls and the code together can be the shorter path. The criteria above apply either way.
Where DATS fits
DATS, the AI consulting system from DILR.AI, works on exactly this decision. The DATS five-stage system runs through Discover and Diagnose, Prioritise and Place, Operating Model, Pilot to Production, and Scale and Run. The diagnostic produces a ranked roadmap of where AI belongs and where it does not, which is the right place to decide which actions an agent should take at all. The operating model covers governance, RACI and lifecycle, audit-ready by design, which is where autonomy levels, limit owners and promotion rules live. An embedded execution office then ships the placements, with production placements the client owns.
Where a firm wants the work record for an agent team as well as the controls, Cognibl, from DILR.AI, attaches the definition of done, the evidence behind it and every tool call to the work, and its two AI flows summarise and flag but never decide. See how Cognibl records agent work.
Does the 2% figure mean only 2% of UK banks use autonomous AI agents?
No. The 2% figure in the Bank of England and FCA 2026 survey is the share of full-survey respondents reporting any agent deployed with broad autonomy, across all financial sectors. It is not a share of banks. Separately, 46% reported agents with constrained execution and 22% with conditional autonomy, and because a firm can report several levels these overlap, so agents that act within limits are far more common than 2% suggests.
The respondent base was 105 firms for the full survey, so treat the figure as a signal about the market rather than a precise sector rate. Our AI regulatory constraints survey guide explains how the same base affects other headline figures.
Is a human approval step enough to keep an AI agent decision out of automated decision-making rules?
A human approval step keeps an AI agent decision outside the UK GDPR solely automated decision rules only if the human involvement is meaningful. Article 22A treats a decision as solely automated where there is no meaningful human involvement, so an approver who cannot change the outcome, or who approves without the information to judge it, may not take the decision out of scope.
That is why approval design matters as much as approval existence, and why the survey's finding that human approval is reported almost universally as a data risk control does not, on its own, settle the question for an agent that acts. If you are testing an approval design today, talk to the DATS team about the evidence a reviewer would expect to see.
Want to work through this on your own agent estate? Start with DATS AI consulting, read our approach to placing AI inside regulated systems, see the named enterprise AI solutions, or browse more AI guides by industry, including the UK enterprise AI consulting guide.
Decide which agent actions move up a rung.
30-min scoping call · No deck · Confidential. We will tell you whether DATS fits, and which actions should stay with a person.
Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.