AI agents in financial services: what is allowed, what is required and what comes next

How much AI does finance use?

86% of the UK firms surveyed by the Bank of England in 2026 used AI, but only 2% had agents with broad autonomy.

Widely, but mostly with a human still in charge. In its 2026 survey of AI in UK financial services, the Bank of England asked 196 firms about their use of AI. 86% used it, up from 75% in 2024, and third-party implementations made up 39% of all reported AI systems. 80% of the firms had AI agents deployed for decision support, where the agent recommends and a human decides, and a further 14% planned to deploy agents within twelve months. Agents with more room were much rarer: 46% of the firms had agents that carry out actions within predefined limits, 22% had agents that act independently but need human approval for certain actions, and only 2% had agents with broad autonomy. Operations and IT was the largest area, with approximately 31% of all reported AI systems, up from around 22% in 2024. The firms saw the greatest current benefits in operational efficiency, automation and productivity, followed by cybersecurity and operational resilience, and risk management and control effectiveness.

A worked example from the Netherlands shows how early the step to agents still is. The Dutch market regulator AFM surveyed 323 asset managers (report published in March 2026). 53% used AI or planned to within a year, but 78% had no plans to deploy AI agents or multi-agent systems within 12 months. 89% had no formal policy for the use of AI agents, and only 2% had a policy written specifically for agents.

A large share of the work in financial organisations is repetitive and predictable. That is exactly where AI agents deliver major savings and better work, and with that a competitive edge and a better experience for customers.

What sets an AI agent apart?

An AI agent does not just give an answer, it also acts, and that changes supervision.

An ordinary AI model gives an answer or a prediction. The Financial Stability Board (FSB) describes agentic AI as autonomous systems that are capable of planning, reasoning and executing complex high-level goals, made up of multiple tasks, independently. The FSB notes that financial institutions have shown increasing interest in AI agents that can operate autonomously across end-to-end processes. That is what makes the difference for supervision. As soon as a system prepares a payment, makes an offer to a customer or places an order, it is no longer advice to an employee but an action by the institution. In its final report of 15 October 2026, the FSB lists the risks that come with that autonomy:

• unauthorised actions
  an agent pursues a goal and takes illegal, unethical or unauthorised steps without adequate human approval;
• erroneous actions
  an agent can take hundreds of intermediate steps and make errors in any of them, which can be very hard to detect in real time;
• data security and privacy
  agents may expose or manipulate customer details, trade secrets or internal communications;
• disruption to connected systems
  agents work through APIs, databases and other agents, all of which can be disrupted if an agent malfunctions or is compromised.

The FSB adds that these risks “can materialise at great speed”.

Are AI agents allowed?

Yes: none of the supervisors cited here bans AI agents, and existing rules already apply.

Most of the supervisors cited here point to rules that already apply. The BIS Financial Stability Institute concluded in a December 2024 paper that most financial authorities have not issued AI-specific regulations for financial institutions, because existing frameworks already address most of the risks. It does name areas that need further attention, including governance, skills, model risk management, data governance and third-party AI service providers. The individual supervisors say the same in their own words:

• UK
  the FCA does not plan to introduce extra regulations for AI and relies on existing frameworks, such as the Consumer Duty and the Senior Managers and Certification Regime, which puts accountability on senior managers.
• Investment services in the EU
  ESMA expects firms that use AI to comply with MiFID II, in particular the obligation to act in the best interest of the client. It names overreliance on AI, by firms and clients alike, as one of the risks.
• Insurance in the EU
  EIOPA published an opinion in August 2025 that clarifies how existing insurance legislation applies to AI systems, with a risk-based and proportionate approach.
• A Dutch example
  in a joint report from 2024, the Dutch central bank DNB and the AFM write that supervisory standards do not depend on the technology used, and apply in full when AI is used.

What applies in the EU today?

AI literacy, transparency and DORA already apply, also to firms that are only starting with agents.

There are three obligations. AI literacy. Since 2 February 2025, organisations that use AI must take measures for the AI literacy of their people, according to the European Commission. From 2 August 2026, national authorities are responsible for supervising and enforcing the AI Act. Transparency. The transparency rules of the AI Act came into effect in August 2026. A customer who talks to a chatbot must be made aware that they are talking to a machine. Digital operational resilience under DORA. The Digital Operational Resilience Act has applied since 17 January 2025, according to the European Banking Authority (EBA). Financial entities in scope must keep a register of their contracts with ICT third-party service providers. An agent built on a model from an outside provider may rely on such a provider, so check whether that arrangement belongs in the register.

What changes in December 2027?

From 2 December 2027, stricter rules apply to high-risk AI such as credit scoring.

The biggest step is the regime for high-risk AI. According to the European Commission, AI used to give access to essential services, such as credit scoring that can deny someone a loan, counts as high risk. DNB and the AFM point to a second financial use that the AI Act lists as high risk: pricing and risk assessment for life and health insurance. Those rules apply from 2 December 2027, later than first planned. The date moved because of the so-called AI Omnibus, the simplification of the AI Act that entered into force on 27 July 2026. From then on, high-risk systems must meet requirements that include:

• risk assessment and mitigation;
• high-quality datasets, to reduce the risk of discriminatory outcomes;
• logging of activity, so that results can be traced;
• detailed documentation and clear information for the organisation that uses the system;
• appropriate human oversight.

That sounds like a lot of new work, but the EBA is reassuring. In 2025 it mapped the AI Act against EU banking and payments law, with a focus on creditworthiness and credit scoring, and found no significant contradictions. Some effort will be needed to integrate the two frameworks, but the EBA saw no immediate need for new or revised guidelines.

What do supervisors expect?

The FSB asks boards to use its 12 sound practices, including least agency for AI agents.

The clearest answer comes from the FSB. In October 2026 it published 12 sound practices for the responsible adoption of AI, and it strongly encourages the board and senior management of financial institutions to use them when they develop their AI governance, risk management and controls. The practices are not an international standard, and they do not replace local law. The first four are about governance: the board and senior management set and oversee how AI adoption fits the business model, risk appetite and strategy; there are clear roles and responsibilities; AI risks are part of the risk management framework and AI use is documented; and the organisation keeps learning. The other eight follow the life of an AI system, from assessing each use case and choosing the technology to data governance, explainability, performance, human oversight, cyber and ICT risk, and third-party dependencies.

For agents, the FSB adds specific considerations:

• Least agency.
  Give an agent only the permissions it needs to reach its goal. The FSB warns that risks can compound multiplicatively, not additively, as an agent gets access to more tools and systems.
• A risk tolerance per agent.
  Decide at the start of each agentic use case how much autonomy and access the agent gets.
• An identity for every agent.
  Unique identifiers make it possible to see which actions an agent took, and which humans configured and oversaw it.
• Traceability.
  Record what triggered an action, which information and tools the agent used, which steps it took and whether it stayed within its permissions.

The FSB is open about the limits: sound practices specific to agentic AI and multi-agent systems “remain nascent” and may need fuller treatment later.

What such a stability risk could look like was sketched as a scenario by economist Torsten Slok of Apollo in September 2026: customers’ personal agents could automatically move savings to the accounts with the highest rates, 3.3% to 5.0% in the US against an average of 0.1% on checking accounts. In that case “banks could lose a large share of the cheap deposits they rely on to make loans”. It is a US scenario, not a forecast. Concentration is the wider concern. In its 2024 report on the financial stability implications of AI, the FSB named third-party dependencies and service provider concentration as the first of four vulnerabilities that could increase systemic risk. The 2026 UK survey shows why: the top three cloud and compute providers accounted for 79% of all reported relationships, and the top three model providers for 54%.

An AI agent in practice

In an FSB case study, an agent proposes fraud rules and people approve every one of them.

The FSB report includes a case study, based on what the bank reported. A large internationally active bank built an agentic AI system in three months to detect emerging fraud and scam patterns. The agent runs continuously and proposes new detection rules, but the bank’s fraud analytics team reviews and approves every new rule before it goes live. According to the case study, the agent contributed to three quarters of the bank’s card fraud rules, and fraud losses fell by over 20% in the first half of the 2026 financial year compared with the same period in 2025. The pattern is worth noting: the agent does the searching and drafting at speed, and a human still takes the decision that changes how the bank treats its customers.

How do you start responsibly?

Start small, with a low-risk task, limited permissions and a human in charge.

1. Start with the AI-first organisation.
   Make AI a matter for the board, not only for IT. Agree who owns it and who decides where agents are used. Then you know in time who is responsible for high-risk AI when the stricter rules apply in December 2027.
2. Start with a low-risk task.
   First walk through your processes, from preparing case files to assessing credit, and decide per task what an agent can do and what belongs with people. An agent that summarises internal documents or prepares a case file is a better first step than an agent that grants credit on its own.
3. Decide how much autonomy each agent gets.
   Set a risk tolerance per use case and grant the fewest permissions that will do the job.
4. Give every agent an identity and a log.
   If you cannot see what an agent did and why, you cannot supervise it, and neither can your supervisor.
5. Keep a human in charge of decisions about customers.
   The FSB describes several forms of oversight, from human approval of every decision to intervention only when needed. Firms that build in human oversight now will need fewer changes in December 2027.
6. Check whether the model provider belongs in your DORA register.
   In the EU, the contractual arrangements you have directly with an ICT provider belong in your register of ICT third-party arrangements; whether a provider further down the chain must be included depends on the arrangement, so check it.
7. Train your people.
   AI literacy has been an obligation in the EU since February 2025, and the FSB asks that the board, senior management and control functions have the skills for their roles too.

For banks and insurers

Waiting until December 2027 is not a good plan, because several rules already apply.

The high-risk rules have been postponed, but AI literacy, transparency and DORA already apply, and supervisors from London to Basel are now spelling out what they expect from boards. JPMorganChase is taking a two-track approach. The bank uses agents internally, but in its third-quarter 2026 earnings call Jamie Dimon said the bank is “not going to allow just any outside agents to come in inside the company, because it’s just too risky for clients and for JPMorgan”. CFO Jeremy Barnum added that if customers “give up their credentials to the agent and the agent does something that they don’t want, that’s going to be on them”. Who is liable when a customer’s agent gets it wrong is a question every bank and insurer will have to answer. Firms that introduce agents now, on a small scale and well documented, build the knowledge they will need later. We wrote earlier about how generative AI is reshaping the finance sector and about how autonomous AI will reshape everything you do.

An agent that prepares a case file also changes the work of the people around it. I think managers at banks, insurers and pension funds achieve most when they give their teams room to experiment with this new technology, starting with low-risk tasks. Enthusiasm for what it delivers and a critical eye on the risks should go together. That side of change, getting people moving, is also the subject of the keynote on the organisation of the future. If you want to take your board, leadership team or financial services conference through what AI agents mean for your organisation, take a look at the keynote on agentic AI or the keynote on the AI revolution.

Frequently asked questions

Can a bank use AI agents?

Yes. None of the supervisors cited in this article bans AI agents in financial services. The existing rules for governance, risk management and treating customers fairly also apply to AI. If an agent is used to assess the creditworthiness of individuals in the EU, the stricter rules for high-risk AI apply from 2 December 2027.

Who supervises AI in financial services?

In the EU, the starting point of the AI Act is that the existing financial supervisors also supervise the AI rules where AI is used for financial services, according to the Dutch supervisors DNB and AFM. In the UK, the FCA applies its existing rules, such as the Consumer Duty and the Senior Managers and Certification Regime.

When do the high-risk rules apply?

For uses such as credit scoring, from 2 December 2027, according to the European Commission. That is later than first planned, because of the AI Omnibus that entered into force on 27 July 2026.

What does DORA have to do with AI agents?

DORA is about digital operational resilience, including failures and errors at ICT providers. An AI agent that runs on a model from an outside provider can use such an ICT service. If you have contractual arrangements with that provider, they belong in the register of ICT third-party arrangements that DORA has required since January 2025.

What are the FSB sound practices for AI?

Twelve practices, published by the Financial Stability Board on 15 October 2026, to help financial institutions adopt AI responsibly. Four cover organisation-wide governance, eight cover risk management across the AI lifecycle, and the report adds considerations for agentic AI. They are not an international standard.