The pilot always works. The rollout is the hard part.
Field notes from Claude deployments that outgrew the pilot team. This hub collects what that stage actually requires: the Model Context Protocol as an integration layer, a decision about where inference may run when the data is European, cost measured per token rather than per seat, and review of code an agent wrote.
From one enthusiastic team to a company-wide system.
This area covers what happens after a team has proved Claude is useful and someone asks for two thousand seats: how identity and entitlements are wired, how internal systems get connected through the Model Context Protocol, how usage is measured in tasks completed rather than licences issued, and what review looks like when an agent is writing code that ships.
Our position: the seat count is the least interesting number in the whole programme. A company where four hundred people paste text into a chat window has bought a faster typewriter. The value appears when the assistant can reach the ticket system, the document store and the repository under the asking user's own permissions — and that is an integration and identity project wearing an AI badge.
The second thing that gets underestimated is review. Coding agents shift the bottleneck from writing to reviewing, and a team that does not change its review practice at the same time simply moves its queue one step to the right. Several posts here are about that shift specifically, because it is the part nobody budgets for.
Common questions
What usually blocks a Claude rollout past the first team?
Access, not enthusiasm. The pilot team works from public context and their own files, and the moment the tool has to reach a document management system or a ticket queue, someone has to decide whose permissions apply. That decision is architecture, and it is where most rollouts stall.
What is the Model Context Protocol and why does it matter here?
MCP is the open standard Anthropic published for connecting models to tools and data sources through a common interface. It matters because it turns every internal integration into one server other assistants can also use, instead of a bespoke plugin tied to a single vendor.
Not a sales call. An architecture call.
Thirty minutes with the architect who would actually run the engagement.