How bunq Handles 97% of Support With Amazon Bedrock

bunq is Europe's second-largest neobank, founded in 2012 by Ali Niknam, with 20 million users across Europe. Its generative AI assistant Finn, launched in 2023 and built fully in-house, answers user questions, can translate the bunq app into 38 languages and translates speech-to-speech calls to the support team in real time. The AWS post describing the system was co-written with Benjamin Kleppe, Machine Learning Engineering Lead at bunq. The first multi-agent support system used a central router agent that sent each query to a specialised sub-agent for areas such as technical support, general inquiries and transaction status. As bunq added more specialised agents, three problems appeared: the router needed ever more complex logic to pick the right destination, several agents needed the same capabilities, and the router became a single point of failure and a development bottleneck. bunq redesigned the system around an orchestrator agent. Instead of routing to every possible agent, the orchestrator sends queries to only three to five primary agents, and those agents can invoke other agents as tools when a request needs them. The agents run on Amazon ECS. Amazon Bedrock provides Anthropic's Claude models for natural language understanding, alongside fine-tuned models hosted on Amazon SageMaker for specialised banking scenarios. Agent memory and conversation history are kept in DynamoDB, OpenSearch provides the vector store for semantic search across bunq's knowledge base, Amazon S3 stores documents and Amazon MemoryDB manages user sessions. AWS WAF and CloudFront protect the front end, and CloudTrail, GuardDuty and CloudWatch provide monitoring. The project started in January 2025 and went from concept to production in three months. A team of 80 people, from AI engineers to support staff, tested the system and deployed updates three times a day during the initial rollout. bunq reports that Finn now handles 97% of support interactions, with 70% fully automated and an average response time of 47 seconds; another passage of the same post puts the fully automated share at over 82%.