European neobank bunq has redefined customer service by integrating a sophisticated multi-agent AI system named Finn, built on the foundations of Amazon Bedrock. By transitioning from a traditional support model to an autonomous agentic framework, bunq now successfully handles 97% of user interactions, with 70% of these cases being fully automated without human intervention.

The technical backbone of this success lies in an "orchestrator" pattern that manages specialized AI units as tools. Instead of a single model attempting to master every banking nuance, a central orchestrator directs queries to specific primary agents. These AI Agent entities can then dynamically call upon other specialized tools to perform tasks like analyzing transaction logs or interpreting complex regulatory documents. This decentralized decision-making allows the system to remain agile as bunq scales its services to 20 million users across the continent.

The Multimodal system provides real-time speech-to-speech translation in 38 languages and handles visual tasks like receipt processing through image recognition. This transformation, completed in just three months, illustrates how Retrieval-augmented generation (RAG)—a method of providing AI with specific, up-to-date data—can turn a general Foundation Model into a highly reliable financial expert. By minimizing manual escalations, bunq has slashed average response times to a mere 47 seconds.

European neobank bunq has redefined customer service by integrating a sophisticated multi-agent AI system named Finn, built on the foundations of Amazon Bedrock. By transitioning from a traditional support model to an autonomous agentic framework, bunq now successfully handles 97% of user interactions, with 70% of these cases being fully automated without human intervention.

The technical backbone of this success lies in an "orchestrator" pattern that manages specialized AI units as tools. Instead of a single model attempting to master every banking nuance, a central orchestrator directs queries to specific primary agents. These AI Agent entities can then dynamically call upon other specialized tools to perform tasks like analyzing transaction logs or interpreting complex regulatory documents. This decentralized decision-making allows the system to remain agile as bunq scales its services to 20 million users across the continent.

The Multimodal system provides real-time speech-to-speech translation in 38 languages and handles visual tasks like receipt processing through image recognition. This transformation, completed in just three months, illustrates how Retrieval-augmented generation (RAG)—a method of providing AI with specific, up-to-date data—can turn a general Foundation Model into a highly reliable financial expert. By minimizing manual escalations, bunq has slashed average response times to a mere 47 seconds.