MRH Trowe, one of Germany's leading commercial and industrial insurance brokers, has handed self-service AI agents to about 400 employees, and the first month of production cost roughly $14 per seat. The number turns heads. The quietly impressive part is that a firm bound by German financial-sector rules let staff build their own agents without client data going walkabout.
AWS AI Watch analysis
What happened
The case study, published on AWS's machine learning blog on Thursday, lays out the whole deployment. MRH Trowe, which operates across Germany, Switzerland and Austria, faced the modern regulated-firm dilemma: employees wanted generative AI in their daily work, teams were already experimenting on their own, and unofficial tools touching client data were a compliance problem waiting to happen.
The answer is a three-part stack. Strands Agents, an open source SDK, lets staff build agents in a few lines of code. Amazon Bedrock AgentCore runs them in production and isolates each agent session at the compute and filesystem level. LibreChat, an open source chat interface, provides a branded front door with proper logins, permissions and token budgets, so spending cannot quietly run away from the finance team.
The first production agent is pleasingly mundane: meeting minutes. An employee asks in German for a recent Teams meeting with a named participant, the agent finds it on the calendar, pulls the transcript and drafts a structured summary with agenda, topics and actions.
Two design choices keep the regulator calm. Every request runs as the signed-in employee, with the identity passed server-side through Microsoft Entra ID, so an agent can reach only that person's own calendar and files and cannot be re-pointed from the chat box. And the agents, models and data all run in the AWS Europe (Frankfurt) Region, keeping client information in Germany.
Why it matters
The pattern is the portable part. Regulated firms across Europe are stuck between banning AI and losing their people to unofficial tools. This is a worked example of the third road: make the governed option the easy option, with identity, residency and session isolation designed in rather than bolted on afterwards.
The number matters too. Roughly $14 per seat for a first production month is a budget line people can actually argue about, and cost opacity is one of the quieter ways enterprise AI pilots die. AWS projects a path to about 40 per cent off infrastructure costs through right-sizing and scheduled scaling, which is the vendor's arithmetic, but at least it states its terms.
Our read
Read it as a pattern worth stealing and a bill worth checking. Every figure here comes from the company publishing a report card for its own stack, so treat $14 a seat as a friendly first month rather than a law of nature.
What survives scepticism is the architecture: identity set server-side so the chat box cannot impersonate anyone, in-region processing, per-session isolation and hard token budgets. Those ideas are portable whatever you run. The adoption play is worth copying as well: workshops, power users and promoted workflows instead of mandates. Board member Leonid Karlinsky sums up the culture in one line: "If you and your colleague are doing things twice, consider creating a LibreChat agent!"
The stated vision, that every question should first be answered by AI before any human intervenes, is a bold sentence for a regulated broker. The compliance wrapper above it is doing real work, and it is worth watching how much weight that wrapper has to carry.
What to watch
- Steady-state economics
Whether costs stay near $14 a seat as usage grows, and whether the projected 40 per cent infrastructure saving survives contact with production. - Talk to your data
The newest agent blends CRM data with public information for cross-sell analysis; how far that pattern travels with client data is the real test. - The copycats
Whether other regulated European firms adopt the same Frankfurt-residency, AgentCore pattern for staff-built agents. - The reference stack
Whether Strands Agents plus Bedrock AgentCore becomes the default template pitched to regulated industries.
Discussion spark: MRH Trowe wants every question answered by AI before a human weighs in. For a regulated insurance broker, is that the sensible end state, or a compliance incident waiting for a quiet Friday?
Sources and evidence
- How MRH Trowe enabled secure self-service AI agents in financial services (17 September 2026, 15:36 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)