• 80%

  Of responding UK financial firms have AI agents deployed today.

  Source: BoE and FCA, Oct 2026
• 9% at most

  Expect agents to approve customer outcomes, move payments, trade or change limits, even including the next 12 months.

  Source: BoE and FCA, Oct 2026
• 8% to 25%

  Use or plan to use agents to create or modify records in core systems: 8% today, 25% including the next 12 months.

  Source: BoE and FCA, Oct 2026

Eight in ten firms run agents, and their commonest job is to recommend

The Bank of England and the FCA published their 2026 survey of AI in UK financial services on 26 October. 196 firms responded. 105 of them completed the full survey, and most of the figures in this post are shares of those full-survey responses. 86% of firms use AI, up from 75% in 2024, and 94% have at least one AI system in production.

Agents are already ordinary. 80% of respondents have AI agents deployed, and a further 14% plan to deploy them within 12 months. The report sums up the state of play as widespread deployment with limited autonomy. Decision support is the most commonly reported level, at 80%: the agent recommends, and a person decides and executes.

So if your agents stop at the recommendation, your team is in the industry's largest group. The figures describe UK firms. A risk committee in New York or Singapore asks the same questions, so the request below is written to work there too. What moves a stalled case is a better-shaped request, scoped to one workflow taken to production.

Your committee is being asked for the four hardest verbs

Some actions stay rare. In the Bank of England and FCA survey of October 2026, even including deployment planned over the next 12 months, no more than 9% of respondents expect agents to do any of these:

• approve or decline customer outcomes
• initiate, execute or block payments
• execute trades or place orders
• change limits, thresholds and controls

Those four verbs are the daily work of a lending desk, a payments operation and a servicing team. They are also what a headcount plan built on agents tends to assume. A request framed as “let the agent decide” asks the committee to put the firm among that 9% in one step, with no record inside the firm to point to.

Two other actions are growing, and both are smaller.

Record changes and customer messages are growing. Approvals and payments stay at 9% or less.

Share of responding firms, percent. Dashed: today. Solid: including deployment planned in the next 12 months. Each row is a separate action; firms could report several, so rows do not add up.

• Records in core systemsToday
  8%
• Records in core systemsIncluding planned
  25%
• Automated customer messagesToday
  16%
• Automated customer messagesIncluding planned
  34%
• Approve outcomes, payments, trades, limitsIncluding planned, each
  9% at most
• 010%20%30%40%

What sits between decision support and execution?

The Bank of England and FCA survey sorts agent autonomy into four levels, drawn below from the bottom rung up. The second is constrained execution, where “agents execute actions within predefined limits and rules”. Above it, conditional autonomy lets agents act on their own while certain actions still need a person's approval.

A firm could report more than one level, so the four figures overlap and do not add up: 80% report decision support, 46% constrained execution, 22% conditional autonomy and 2% broad autonomy. A firm with a recommending agent in collections and a constrained one updating records in servicing counts in both of the first two.

The rung to ask for is the second. 46% of respondents already report agents working there, so your committee would be approving a pattern close to half the surveyed firms run somewhere. The committee writes the limits and rules, and every action outside them stays with a person.

Four rungs of agent autonomy, and the one to ask for

Ask for one action inside a limit, with the past cases that prove it, and the committee has something it can approve.

Pick the write before you pick the decision

The actions gaining ground are writes. In the Bank of England and FCA survey of October 2026, creating or modifying records in core systems goes from 8% of respondents today to 25% once planned deployment is counted. Automated customer communications go from 16% to 34%. Including planned deployment, both sit well above the 9% ceiling on approvals and payments.

A record update with a value cap, or a case moved to a review queue, is smaller than an approval, can be reversed, and can be checked after the fact. A templated customer message cannot be recalled once sent, so its limit is the approved template and a daily volume cap.

Choose yours against four tests:

• A wrong write can be reversed within a working day, or, for a customer message, is fixed in advance by an approved template.
• A value cap or a daily volume cap bounds it.
• It happens often enough that you can pull 50 to 100 past cases where an expert made the call.
• One person owns it and can be named in the request.

If two candidates pass, take the one with more past cases. Field 4 below depends on that sample.

What goes in a risk committee request for an AI agent?

Start with the owner. In the Bank of England and FCA survey, 94% of respondents said accountability for AI systems and outputs sits with business area owners or users. For an agent in lending, payments or servicing, that is the operations head, and the committee will read the request as yours. Write it on one page, in seven fields: the action, the limit, the cases, the comparison, the model, monitoring and review, and rollback.

1. The action. One verb, one system, one record type. “Set the document-received status on a mortgage application in the origination system” names all three.
2. The limit. A value cap per action, a daily volume cap, and the cases excluded outright, such as customers flagged as vulnerable or accounts with an open complaint.
3. The cases. 50 to 100 past cases where an expert made the call, chosen by your team, awkward ones included. A sample picked to flatter the agent will be spotted.
4. The comparison. For each case, the write the agent intended beside the one your expert made, with the agent's writes recorded rather than executed. List every disagreement in full, with the expert's note on which answer was right.
5. The model. Which model the agent calls, where case data goes when it does, who wrote the test cases it is scored on, and what happens when the provider changes the model. Expect the first question here: applying model risk management to third-party or vendor-supplied models was the governance challenge respondents selected most often, at 54% (55 of 102 firms).
6. Monitoring and review. Who reads the agent's log, how often, and which result pauses it. Write this field in detail. Among controls for AI data risk in general, ongoing monitoring and periodic review are each reported by only around half of respondents, against more than 90% for pre-deployment review and human approval.
7. Rollback and incidents. How a wrong write is reversed, by whom and how fast, and the playbook if a batch of writes goes wrong. Incident management and response playbooks were the least common governance practice the survey listed, at 49%.

For a wider view of your agents on the same questions, the Governance and Control dimension of the Agent Maturity Index asks whether you can list every agent, what it is allowed to do, and stop it in seconds.

Where SynOS fits

Agents built on SynOS rehearse against your real systems with their intended writes recorded rather than executed, and every release is scored against your team's own test cases. A release that scores worse than the last one does not ship. Every read and write is scoped to the person asking and logged, which is the record field 6 asks for. When your experts correct the agent, the correction is kept as training data for a model you own. If you already know which write you want approved, bring that workflow and we will build the field 4 comparison on your past cases with you.

Questions operations heads ask about agent approval