400 agent seats at $14:
governed cost goes public

If an agent cost quote has no governance in it, you are not comparing prices — you are comparing different products. MRH Trowe, a German commercial and industrial insurance broker, put roughly 400 seats of self-service AI agents into production in their first month at about $14 per seat — a rare published hard figure for agents in a regulated vertical. Our judgment: $14/seat is the control-included price; the value of the number is not the price tag, it is that someone finally published the governed unit cost. This piece unpacks what the figure includes and gives regulated buyers a transferable minimum set. Source: AWS ML blog, company- and vendor-reported, no third-party audit.

1. A rare hard number: what $14 per seat buys

Roughly 400 employees reached this scale in the first month of production; ~$14/seat/month covers infrastructure and tokens. Board member Leonid Karlinsky states the vision plainly: “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.” Cost transparency comes from consumption-based billing: AgentCore runtime bills on active CPU and memory per second, and I/O wait is not charged — relevant because agentic workloads typically spend 30–70% of their time waiting on model responses, tool calls, or database queries.

Under evidence discipline: this is company-reported data published through a vendor channel (AWS), with no third-party audit. Read it as a floor and an architecture reference for governed agents, not as a market benchmark. What transfers is the minimum control set and the first-agent choice, not the specific figures.

2. The four control levers of a regulated agent program

Add the foundation — private connectivity (transit gateway plus zero trust) keeps employee traffic off the public internet. These five controls are the “governance minimum set” for regulated verticals, and each maps to a question a compliance review will actually ask.

3. Self-service and central governance are not opposites

The core tension is stated in the official narrative: employees should be able to build and use agents themselves, without deep technical skills, while everything stays inside a secure, centrally governed, compliance-ready environment. MRH Trowe’s answer is not policy — it is architecture. One managed platform (LibreChat + Strands + AgentCore) replaces fragmented shadow AI; the LibreChat admin panel governs which endpoints and models each user can reach via ACLs; employees are encouraged to explore, but exploration happens inside governance boundaries.

For digital-transformation leads, this is the answer to why banning shadow AI fails: prohibition does not stop adoption, it pushes it out of sight. Replacing it with a governed self-service floor is how you get both adoption and control.

4. Our judgment: the control-included price is the real price

Any agent cost figure that leaves out residency, isolation, identity, and budgets is a different product priced a different way. The $14/seat is “what it actually costs to let one employee run a governed agent” — comparing governed delivery against a feature is the only apples-to-apples comparison.

Our position: write the control stack into the cost model on day one, not as an afterthought. Cost transparency is a governance property, not a FinOps follow-up — you cannot govern an agent program whose cost you cannot account for, and you cannot fund one that never priced the controls. This is also our own practice: governance is part of the quote, not a discount line.

5. The transferable template

The first agent is deliberately low-risk and high-frequency: Teams meeting minutes, not claims processing. A one-line request becomes a structured protocol. The step proves that “self-service plus governance boundary” works in real employee hands while barely touching high-risk actions.

The adoption mechanics transfer too: cross-team use-case workshops build a power-user group; adoption data identifies power users whose workflows get promoted; and the board member’s line — “if you and your colleague do something twice, create an agent” — is a usable rule of thumb. The scale path: 10–15 domain-expert agents maintained by subject-matter experts by end of 2026, plus ~40% infrastructure cost reduction via right-sizing and scheduled scaling.

6. Action and the question left for buyers

Action: put the control stack in the cost model before the pilot; choose a low-risk, high-frequency first agent; instrument cost per seat and per use case from day one. The question left for buyers: what does one governed agent seat cost in your organization — including residency, isolation, identity, and budgets? If you do not know, you have not priced the product you are actually deploying.

References: AWS ML Blog, How MRH Trowe enabled secure self-service AI agents in financial services (company- and vendor-reported, no third-party audit) https://aws.amazon.com/blogs/machine-learning/how-mrh-trowe-enabled-secure-self-service-ai-agents-in-financial-services/ ｜ MRH Trowe, company site https://www.mrh-trowe.com/en/