MRH Trowe, a German owner-managed commercial and industrial insurance broker, gave roughly 400 employees secure self-service access to AI agents in the first month of production at an infrastructure-and-token cost of about $14 per seat per month, with a clear path to cut infrastructure spend ~40% through right-sizing and scheduled scaling.
“If you and your colleague are doing things twice, consider creating a LibreChat agent!” said Leonid Karlinsky, Board member at MRH Trowe, in AWS’s published case study.
Operating primarily in Germany, Switzerland and Austria, MRH Trowe had to balance employee demand for generative AI with the data-protection and governance obligations of the German financial sector — teams were starting to experiment with AI on their own, fragmenting tools and risking exposure of sensitive client and insurance data.
The deployment reached roughly 400 employees in the first month of production. The first agent in production turns a Microsoft Teams meeting into structured meeting minutes — pulling the calendar entry, fetching the transcript, and drafting a summary with date, participants, agenda, topics and action items. To drive adoption, the team pairs the rollout with workshops to identify and promote power users across teams.
Under the hood the stack layers three technologies: Strands Agents as the open-source SDK for agent patterns, Amazon Bedrock AgentCore as the production runtime with per-session compute and filesystem isolation, and LibreChat as the open-source branded chat front-end with token budgets, multi-model support and Microsoft Entra ID authentication. The deployment runs in a single AWS account inside a VPC in the AWS Europe (Frankfurt) region (eu-central-1) over a transit gateway plus zero-trust tunnel, keeping client and meeting data inside Germany. The application tier runs LibreChat on Amazon ECS with AWS Fargate; the data layer combines Amazon DocumentDB, Amazon ElastiCache, Amazon RDS for PostgreSQL (also the RAG vector store for uploaded documents), Amazon OpenSearch Service (vector store for content ingested from Confluence), Amazon EFS and Amazon S3; the agent layer exposes Strands agents on AgentCore runtime through AWS Lambda and Amazon API Gateway.
In production across the German-headquartered broker, the platform now serves around 400 employees with the meeting-minutes agent in live use; the next use case is “talk to your data,” supporting cross-sell and upsell reviews that combine CRM data with publicly available information, with new agents loaded into LibreChat independently without downtime to the chat application.
MRH Trowe is targeting 10–15 production agents maintained by subject-matter experts by the end of 2026 through a Data and AI Community of Practice, with adoption data used to identify power users and promote their workflows across the organization.