How bunq handles 97% of support with Amazon Bedrock | Artificial Intelligence

bunq, Europe's second-largest neobank with 20 million users, built Finn — an in-house generative AI assistant powered by Amazon Bedrock (Anthropic's Claude models) — to handle 97% of its user support interactions. Challenge: Traditional support models couldn't keep up with demand for round-the-clock multilingual assistance across 38 languages and multiple time zones. Solution: A multi-agent architecture with an orchestrator pattern. Instead of a central router directing queries to many specialized agents (which caused bottlenecks and scaling issues), bunq redesigned around an orchestrator that routes to 3-5 primary agents, empowering each to invoke specialized agents dynamically as tools. Architecture: Amazon ECS (orchestrator + primary agents) → DynamoDB (conversation memory) → OpenSearch Serverless (vector store for RAG) → Amazon Bedrock (Claude models) → AWS WAF + CloudFront (security layer). Results: Finn handles 97% of support with 70% fully automated, 47-second average response times. The team moved from concept to production in 3 months, deploying updates 3 times daily. Finn translates the app into 38 languages and offers speech-to-speech real-time translation — a first in global banking. bunq became Europe's first AI-powered bank as a result.