MRH Trowe ran 400 employees on self-service agents for about $14 a seat in month one

Four hundred seats at about $14 is roughly $5,600 for the month ^[15]. That $5,600 is the first-month production bill at a company where individual teams had already started experimenting with AI on their own ^[11]. The post does not break out how much of the $14 is infrastructure and how much is model usage ^[17].

The savings lever it does name is capacity: right-sizing and scheduled scaling, for about 40 percent of infrastructure cost ^[4]. Both act on hours you are billed for while nothing is running. Right-sizing pays back where the first sizing was too big. A broker with staff in Germany, Switzerland and Austria ^[6] has a working day and a weekend. An operation with users in every timezone works through both.

The regulated part of the design is in the runtime. AWS says AgentCore was the deciding factor because it natively supports open source frameworks such as Strands Agents, isolates each agent session at the compute and filesystem level, and bills by consumption ^[8]. Per-session filesystem isolation is the specific property a broker needs. One employee's agent writing a client document into scratch space cannot be read by the next session that lands on the same compute. Strands is an open source SDK that handles orchestration and reasoning and scales from a builder's laptop into production ^[9], so the agent code moves between hosts.

LibreChat handles authentication and authorization across the employee base, plus a built-in token budget system ^[10]. A per-seat figure means something only when a per-seat limit is enforced somewhere. Here that is the chat layer.

MRH Trowe's stated vision, as the post gives it, is that "every question should be first answered by AI before any human intervention, and repetitive processes should be automated by those who did them in the past" ^[12]. The four gaps the broker found in basic chat were contextual responses grounded in internal data, multi-step workflows, integration with internal systems and document repositories, and central governance instead of tools spread across teams ^[18].

This is AWS's account of its own customer on its own services, published on the AWS machine learning blog ^[14]. For $14 a seat to transfer, a reader needs employees who use agents about as hard as this cohort did in its first month, a comparable always-on footprint, and the same requirement set the broker was buying against: security, data residency and the compliance controls of the German financial sector ^[13]. MRH Trowe was among the first German insurance brokers to run an exclusively cloud-based IT estate, with AWS as its preferred partner ^[7].