In bank operations most cases follow rules, but the exceptions tie up teams: returned payments, expired IDs, incomplete credit documents. Every decision has to be traceable and every change to customer data evidenced. That is where agents help: they prepare cases and close routine work without bypassing the controls.
Industry 01/103 example agentsAutonomy level L3–L4
Example agents
Three agents that take over work here
Examples from typical workflows, not client references. Which agents pay off for you is shown by assessing your own processes.
01
The payment investigations agent
Payments · back office
Handles returned SEPA payments and investigations: reads the return reason and the customer enquiry, finds the original payment, checks account data and mandate, prepares the correction or re-initiation and answers the customer with the actual status.
No new beneficiary accounts without confirmation, amount limits per case type, any sanctions-list hit stops the case immediately.
Human in the loop
The payments team approves corrections above the limit; suspicious cases go to compliance with the agent's reasoning.
What to measure
Turnaround per investigation · share of standard cases closed · customer follow-ups
02
The KYC refresh agent
Customer data · AML
Works through due customer due diligence refreshes: pulls company register extracts, reconciles beneficial owners with master data, requests missing documents from the customer and hands the case handler a complete file with deviations flagged.
Autonomy level
L0
L1
L2
L3
L4
Start L3 · Assisted agents
The agent closes no case on its own: risk rating and sign-off stay with people.
Systems
Company registersCustomer masterSanctions listsDMS
Guardrails
Read-only register access, no change to the risk class, customer communication only through approved templates.
Human in the loop
Case handlers review the prepared file and decide on risk class and closure, with four eyes for high-risk customers.
What to measure
Backlog of due refreshes · handling time per file · document requests per case
03
The credit file agent
Credit processing · corporate clients
Checks incoming credit applications for completeness, extracts figures from annual accounts and collateral documents, fills the assessment template and drafts the file memo with sources. It does not make the credit assessment itself.
Autonomy level
L0
L1
L2
L3
L4
Start L3 · Assisted agents
Preparation at L3; the credit decision stays with people for good.
Every extracted figure with its source, no scoring decision, no documents shared outside the file.
Human in the loop
Credit analysts review the template and decide; the agent flags contradictions between documents instead of resolving them.
What to measure
Time to a decision-ready file · document requests · analyst corrections
Framework
What to factor in for this industry
DORA: Since 17 January 2025, banks, insurers and investment firms must meet the Digital Operational Resilience Act's requirements on ICT risk and ICT third-party providers. That covers agents and their model providers too.
AI Act: AI systems that assess the creditworthiness of natural persons are high-risk; the obligations apply from 2 December 2027. Systems for detecting fraud are excluded.
Automated decisions: Art. 22 GDPR limits decisions made without human involvement. That is why credit decisions stay with people here, with human-in-the-loop as a fixed step.
Note, not legal advice: whether a specific use falls under these rules is for you and your legal counsel to decide. We provide the technical facts for it, from the agent profile to the audit trail.
