AI agents are software systems that read inputs, apply a model or a rule set, and act on the result. European banks are already using them: more than 85% of the banks under European banking supervision use AI in some form, the ECB reported in February 2026. What almost none of them are doing is letting those systems act alone. That gap between how agents are sold and how they are actually run is the useful part of this subject, and it is where the operational gains sit.
What AI agents actually do inside a bank
The term covers everything from a scripted assistant to a system that plans several steps and calls tools. In banking the distinction is not academic, because the regulatory position of each is different.
Banks deploy agents across three layers. The customer-facing layer handles account queries, payment disputes and product questions. The middle office runs know-your-customer checks, fraud scoring and alert triage. The back office reconciles transactions, prepares regulatory returns and watches system health. Most of our AI agent development sits in the second and third of those, where the work is high volume and the output has to be defensible.
Layer | What the agent does | What it produces |
|---|---|---|
Customer-facing | Answers queries, routes disputes | A resolved case or a routed escalation |
Middle office | KYC checks, fraud scoring, alert triage | A score with its reasoning, or an escalation |
Back office | Reconciliation, report preparation | A draft return for human sign-off |
One correction worth making early, because vendor material gets it wrong. Agents in banks do not improve on their own in production. In the workshops the ECB ran with 13 supervised banks across nine countries in 2025, no bank allowed self-learning after deployment. That is not the same as not retraining: models are retrained, and have to be, but as a governed change rather than a side effect of running the thing. Under the ECB's guide to internal models, switching a credit model to a mostly machine-learning approach is itself a material change requiring supervisory approval.
Fraud detection and risk monitoring
Fraud is where the agents have the longest record, and where the case for them is easiest to make in euros. Payment fraud reported across the European Economic Area reached EUR 4.2 billion in 2024, up from EUR 3.5 billion the year before, with credit transfer fraud losses of EUR 2.2 billion and card fraud of EUR 1.3 billion. Payment service users bore around 85% of the credit transfer losses, mostly through scams that persuaded them to make the payment themselves.
That last detail is what changes the engineering problem. A rule that flags unusual geography does not catch a customer who was talked into paying. Behavioural models that score against a customer's own baseline do better, and the agent's job is to combine the signals, assign a score, log the reasoning, and either act or escalate depending on confidence. The escalation path is what makes the decision auditable afterwards, and it is the same pattern we describe in AI orchestration for fintech.
Regulatory reporting and the part agents genuinely reach
Reporting is a volume problem before it is an accuracy problem, and it is expensive in a way that is actually measured. The European Banking Authority put the annual cost of complying with its supervisory reporting requirements across the EEA banking population at EUR 5.5 billion, roughly 1.4% of the total operating costs of all consolidated EU banks, at a median of EUR 790,700 per institution for ongoing costs.
Inside that figure sits the number worth noticing: EUR 0.7 billion a year goes on ad hoc data requests, at about EUR 160,000 per institution. Scheduled returns are already automated in most banks. The unscheduled ones, where somebody assembles an answer from four systems under a deadline, are the unstructured part of the burden and the part an agent plausibly addresses. That is a narrower claim than the vendor pitch, and it survives contact with a chief financial officer.
Choosing the platform the agents will run on?
Our free vendor-neutral scorecard rates any mini-app or super app platform on seven criteria, including SDK openness, runtime portability, FINMA and DORA fit, exit terms and total cost, with the exact question to put to each vendor.
Where AI agents create friction and how banks manage it
Three constraints recur, and none of them argues against adoption. They argue for treating it like any other critical operational change.
Friction point | What it takes to manage it |
|---|---|
Model risk | Validation, monitoring and a governed retraining path, with drift detection that prevents automatic material change |
Data quality | Preparation before deployment; a model fed poor data produces confidently wrong output, not cautious output |
Vendor concentration | A pre-contract assessment under DORA Article 29 and an exit plan under Article 28(8) that has been tested |
The third has become concrete. On 18 November 2025 the European Supervisory Authorities designated the first 19 critical ICT third-party providers under DORA, among them Amazon Web Services, Microsoft and Google Cloud. Every model API a bank calls is a line in the Article 28(3) register of information. An exit plan that exists only on paper is not an exit plan, and the regulation uses the word “tested”.
There is also a liability point that catches banks buying rather than building. Under Article 25(1) of the EU AI Act, a deployer that puts its own name on a high-risk system, or modifies it substantially, becomes its provider and takes on the full provider obligations. For a bank running credit scoring, the penalty tier is EUR 15 million or 3% of worldwide annual turnover, not the headline figures quoted for prohibited practices. The security consequences of giving an agent real access are covered separately in our piece on smart agents and secure banking.
Banking operations work we have shipped
Putting an agent into a banking workflow?
WislaCode builds regulated banking software where the governance layer is part of the design: decision logging, escalation thresholds and the integration underneath them.
What is an AI agent in banking?
A system that reads inputs, applies a model or rule set, and acts, rather than only answering. In a bank it typically scores a case, records why, and either completes the task or escalates it. The difference from older automation is that it handles variation in the input rather than breaking on it.
Do AI agents in banking learn and improve on their own?
Not by themselves. In the ECB's 2025 workshops with 13 supervised banks, none allowed self-learning after deployment. That is not the same as not retraining: retraining is a controlled model change with validation and, for internal models, supervisory approval behind it.
Where do AI agents actually save money in banking operations?
The clearest case is the unscheduled work. The EBA puts EEA-wide supervisory reporting compliance at EUR 5.5 billion a year, of which EUR 0.7 billion goes on ad hoc data requests, around EUR 160,000 per institution. Scheduled returns are largely automated already; assembling an answer from several systems under a deadline is not.
What are the main regulatory constraints on AI agents in EU banking?
DORA governs the resilience and the vendor relationship: the register of ICT arrangements, the concentration assessment before contracting under Article 29, and a tested exit plan under Article 28(8). The AI Act adds logging under Article 12 and human oversight as a design property under Article 14. GDPR Article 22 already governs decisions made solely by machine.
Which European banks are running autonomous AI agents in production?
None that has published it. The supervisory evidence points the other way: in the ECB's sample no bank used generative AI for credit scoring or fraud detection, and every deployment kept a human in the loop. Treat any vendor claim of autonomous banking agents in Europe as unevidenced until the bank itself says so.