bunq: Multi-agent generative AI assistant on Amazon Bedrock to handle 97% of support
bunq, Europe’s second-largest neobank, upgraded its in-house generative AI assistant Finn to improve multilingual customer support and automate banking operations while maintaining security and compliance requirements. The solution uses Amazon Bedrock with Anthropic Claude models, Amazon ECS for orchestrator and agent services, Amazon DynamoDB for memory and conversation history, Amazon OpenSearch Serverless for vector search in RAG, and Amazon S3 for document storage. bunq redesigned the assistant around an orchestrator agent and an agent-as-tool pattern so primary agents can dynamically invoke specialized tools for tasks such as transaction analysis, document retrieval, failed payment handling, and image/document recognition.
- Organization
- bunq
- Industry
- Finance
- Location
- Netherlands
- Published
- January 2026
Reported outcomes
97%
quantified impactOther quantified impact
Strategic outcomes
Primary read
Use case focus
Showing 3 of 5
- 1Customer Service Automation
- 2Intelligent Virtual Assistant
- 3Workflow Automation
- Deliver consistent, real-time, multilingual customer support across banking workflows.
- Automate support and operational tasks such as failed payments and receipt/document processing.
- Reduce manual escalations and routing bottlenecks while meeting strict security and compliance needs.
- bunq built Finn as an in-house generative AI assistant on Amazon Bedrock using Anthropic Claude models.
- The assistant runs a scalable multi-agent architecture with an orchestrator agent on Amazon ECS.
- Primary agents can invoke specialized tool agents dynamically through an agent-as-tool pattern.
- Amazon DynamoDB stores agent memory, conversation history, and session data.
- Amazon OpenSearch Serverless provides vector search for RAG over bunq's knowledge base.
- Amazon S3 is used for document storage and the broader architecture includes security and observability services.
- Finn handles 97% of bunq’s user support activity.
- More than 82% of support work is fully automated.
- Average response time was reduced to 47 seconds.
- bunq went from concept to production in about 3 months.
- The app was expanded to 38 languages and supports real-time speech-to-speech translation.
Architecture
The article describes a multi-agent orchestrator architecture where an orchestrator agent on Amazon ECS routes requests to a small set of primary agents, and those agents dynamically invoke specialized tool agents. Supporting services include Amazon Bedrock for Claude models, DynamoDB for memory/session state, OpenSearch Serverless for RAG vector search, and S3 for document storage.
Sources & evidence1
AI-generated summary. Verify important details with the linked sources before relying on this case.
Explore related AI use cases
Was this useful?