In this session, I will present the development of our email automation for customer support at Trusted Shops using Amazon Bedrock, from simple LLM API calls to a multi-agent architecture.

We started small: a single Bedrock API call for intent classification, predefined response templates, and a strong focus on data privacy compliance from the very beginning.
Over time, the solution evolved through several stages: from basic prompt engineering to Bedrock Agents, from single-agent to multi-agent orchestration, and finally to Amazon Bedrock AgentCore. Each phase brought new challenges and new learnings about what it takes to run AI solutions in production.
I will also share practical lessons learned along the way, including where simple LLM approaches reach their limits and when agent-based architectures start to add real value.
Today, we can handle over 50% of 1st level support tickets entirely with AI, reduce resolution time by 60%, and continue to achieve high customer satisfaction.

The audience will gain insights into iterative AI implementations, quality control mechanisms and data privacy considerations while keeping customer satisfaction and strengthening the operational service teams.