German insurance broker gives 400 employees governed access to AI agents

German insurance broker MRH Trowe gave about 400 employees access to self-service AI agents during the first month of a production rollout. The system is designed to let staff build useful workflows without sending sensitive insurance and client data into unmanaged consumer tools.

The deployment combines LibreChat for the user interface, the open-source Strands Agents SDK for building workflows and Amazon Bedrock AgentCore for isolated production sessions. Central controls cover authentication, permissions, token budgets, model selection and conversation management. MRH Trowe reported an initial cost of about $14 per employee for the first month and expects infrastructure tuning and scheduled scaling could reduce that figure by roughly 40%.

Its first production agent turns Microsoft Teams transcripts into structured meeting minutes. An employee identifies a meeting, and the agent retrieves the transcript and drafts the date, participants, agenda, discussion and action items. This connects internal data and multi-step tools rather than offering only a general chatbot.

The numbers come from an AWS customer case study and do not establish the same cost or compliance outcome for another organization. The deployment shows a practical response to shadow AI: provide a convenient sanctioned option with session isolation and visible budgets. Human review remains important before generated minutes or agent actions become official business records.