Most of a bank's AI agents carry light duties under the EU AI Act, and a few carry heavy ones. Agents that assess individuals' creditworthiness are high-risk, with oversight, logging and impact assessment duties. Customer-facing service agents must tell people they are AI. Every agent requires trained staff, and no agent may use prohibited practices.
What does the EU AI Act require of a bank?
The Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and phases in. Prohibited practices applied from 2 February 2025. Dates for high-risk duties have been subject to proposed changes, so build to the substance now.
It is risk-based, and for a bank that means four questions per agent:
- Is any use prohibited? The banned practices include manipulating people through techniques that bypass their judgement and exploiting vulnerabilities such as age or financial hardship to distort their behaviour. A collections or sales agent must be designed well clear of both.
- Is it high-risk? Annex III lists creditworthiness assessment and credit scoring of natural persons, among other uses. Some uses outside banking operations, such as decisions about staff, can also be listed.
- Does it interact with people or generate content? Article 50 requires that people are told when they interact with an AI system.
- In every case, Article 4 requires staff dealing with AI systems to have sufficient AI literacy.
Banks are usually deployers, which brings Article 26 duties: use the system as instructed, assign competent human oversight, monitor it, keep its logs and inform affected people. Deployers of credit scoring systems must also run a fundamental rights impact assessment under Article 27. A bank that builds its own agents, or substantially modifies bought ones, takes on provider duties as well.
Which banks does it apply to?
Any bank that places AI systems on the EU market or uses them in the EU, and banks outside the EU whose AI output is used inside it. A third-country bank serving EU customers through an agent is in scope for that agent.
Banks have a particular position. The Act lets some deployer and provider duties, such as monitoring and quality management, be met through the internal governance arrangements banks already run under financial services law. Supervision of AI used by financial institutions is also pointed towards the existing financial supervisors. That saves duplication but raises the bar: the AI duties will be examined by people who already know your risk framework. Our companion piece for fintechs deploying agents covers firms without that governance in place.
What does it mean when the system is an agent that acts?
The Act was drafted around systems that produce outputs. Agents also take actions: they retrieve data, call tools, draft messages and sometimes move money. We discuss the wider point in the EU AI Act and agents that act.
For a bank, the practical consequence is that classification has to happen per capability, not per agent. One agent might answer balance questions (transparency duty), gather documents for a loan (part of a high-risk process) and set up a payment (outside Annex III but squarely inside your conduct and operational resilience rules). Classify the agent by its highest-risk capability, or split the capabilities so each can be governed and certified on its own.
The three main areas look like this:
- Credit. Agents that read statements, compute affordability or recommend decisions on individuals are high-risk. Keep them at depth 2 so a trained person decides.
- Payments. Payment agents are not generally listed in Annex III, and fraud detection is carved out of the credit listing. The risk is irreversibility, so the controls come from payment law and operational resilience rules.
- Service. Service agents need clear AI disclosure, and they must not steer customers in ways that cross into manipulation. Hand-over to a person must be easy.
How does a bank meet the duties?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Know which agents are high-risk | Inventory of every agent and each capability, classified under the Act | Classification register signed by an accountable owner |
| Human oversight (Articles 14 and 26) | Named, trained overseers for each high-risk capability; actions held for approval at the right depth | Overseer assignments, training records, approval logs |
| Logs (Articles 12 and 26) | Every agent action and verdict written to an append-only record | Retained logs that reconstruct any case |
| Monitoring | Outcome and drift monitoring for each high-risk agent | Monitoring reports with escalations |
| Fundamental rights impact assessment (Article 27) | Assessment before first use of any credit scoring capability on individuals | Signed assessment, updated on material change |
| Transparency (Article 50) | Customers told when they deal with an AI system | Disclosure wording and channel evidence |
| AI literacy (Article 4) | Role-based training for staff who build, oversee or rely on agents | Training completion by role |
Two of these deserve care. Oversight must be real: a person who approves every recommendation without reading it is not oversight, and a supervisor will look at override rates and review times. And validation should follow your existing model risk process, which our piece on how banks validate an AI agent describes.
What will the supervisor ask?
- Show us your inventory of agents and how each was classified.
- For the credit agent, are you the provider or the deployer, and why?
- Who oversees it, how were they trained, and how often do they override?
- Show us the fundamental rights impact assessment.
- How are customers told they are dealing with an AI system?
- How does this fit with your model risk and ICT risk frameworks?
What does the Act not cover for a bank?
It does not replace prudential rules, conduct rules, AML law or payment law. DORA still governs ICT risk, incident reporting and third-party risk for the systems agents run on. GDPR still governs personal data, including Article 22 on solely automated decisions. National supervisors still expect model risk management. The Act adds a layer; it does not consolidate the others. For the operating picture, see how banks use AI agents safely and our guide to what AI agents in banking may and may not do.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Does every bank chatbot become high-risk under the Act?
No. A service agent answering questions carries a transparency duty and the literacy duty. It becomes high-risk only if it takes part in a listed use, such as assessing an individual's creditworthiness.
Can a bank rely on its vendor's compliance?
Partly. The provider's documentation and logs help, but the bank's deployer duties, including oversight, monitoring and impact assessment, stay with the bank and cannot be outsourced.
Should a bank wait for the high-risk dates to settle?
No. Supervisors already expect model risk, conduct and resilience controls on these agents, and the Act's controls largely overlap with them. Building now avoids a rush later.