Customer support · Production

bunq handles 97% of user support with Amazon Bedrock multi-agent AI system

The problem

bunq's traditional support systems could not keep up with round-the-clock multilingual demand from 20 million users, creating bottlenecks and straining internal resources. As bunq scaled its initial multi-agent architecture, a router-based design became a single point of failure due to routing complexity, overlapping agent capabilities, and a scalability bottleneck that required comprehensive testing for every new agent added.

First attempt

The initial router-based multi-agent architecture became a single point of failure: routing logic grew too complex, agents required access to the same data sources, and every new specialized agent required comprehensive testing of all routing scenarios.

Workflow diagram · grounded in source
1
User submits natural language request
trigger
“Users can interact with Finn through bunq's application and web interface, using natural language for their requests, such as account information, transaction history, financial advice, and support issues.”
2
Orchestrator routes to primary agents
routing
“Routes queries to only three to five primary agents”
3
Primary agents invoke specialized agents
ai_action
“Primary agents detect when they need specialized help. Tool agents are invoked dynamically by primary agents. Agents can call other agents through a well-defined interface—they become tools in each other's toolkits.”
4
RAG knowledge retrieval
ai_action
“OpenSearch Serverless automatically scales compute resources based on application needs and provides vector search capabilities for Finn's Retrieval Augmented Generation (RAG) implementation, enabling semantic search across bunq's knowle…”
5
Coordinated answer delivered
output
“a user asking about a failed payment can trigger a coordinated response: one agent interprets the question, another checks transaction logs, and a third suggests solutions based on similar cases. They all work together seamlessly to deli…”
Reported outcome

Finn now handles 97% of bunq's user support activity with average response times of 47 seconds, was deployed in 3 months from concept, and supports 38 languages, positioning bunq as Europe's first AI-powered bank.

Reported metrics
user support handled by Finn97%
Support fully automated (operations bullet)over 82%
support fully automated (ML lead quote and conclusion)70%
Average response time47 seconds
Show all 9 reported metrics
user support handled by Finn97%
support fully automated (operations bullet)over 82%
support fully automated (ML lead quote and conclusion)70%
average response time47 seconds
languages supported38
concept to production timeline3 months
team size for deployment80 people
deployment update frequencythree times a day
bunq total users20 million
Reported stack
Amazon BedrockAmazon ECSAmazon OpenSearch ServerlessAmazon SageMakerAmazon S3Amazon MemoryDBAWS WAFAmazon CloudFrontAWS CloudTrailAmazon GuardDutyAmazon CloudWatchAnthropic's Claude models
Source
https://aws.amazon.com/blogs/machine-learning/how-bunq-handles-97-of-support-with-amazon-bedrock?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Finn now handles 97% of bunq's user support activity with average response times of 47 seconds, was deployed in 3 months from concept, and supports 38 languages, positioning bunq as Europe's first AI-powered bank.

What tools did this team use?

Amazon Bedrock, Amazon ECS, Amazon OpenSearch Serverless, Amazon SageMaker, Amazon S3, Amazon MemoryDB, AWS WAF, Amazon CloudFront, AWS CloudTrail, Amazon GuardDuty.

What results were reported?

user support handled by Finn: 97%; Support fully automated (operations bullet): over 82%; support fully automated (ML lead quote and conclusion): 70%; Average response time: 47 seconds (source-reported, not independently verified).

What failed first in this deployment?

The initial router-based multi-agent architecture became a single point of failure: routing logic grew too complex, agents required access to the same data sources, and every new specialized agent required comprehensi…

How is this customer support AI workflow structured?

User submits natural language request → Orchestrator routes to primary agents → Primary agents invoke specialized agents → RAG knowledge retrieval → Coordinated answer delivered.

Related customer support cases