copilot as a company gpt – what actually decides the rollout

TL;DR. We turned Microsoft 365 Copilot from a tool used by a handful of people into a company GPT for everyone at a construction company of around 25 employees — including self-built MCP servers that let Copilot agents talk to external line-of-business systems. The technical part was not the hard one. Training was what mattered. And the real work starts after the rollout. Last updated: 2026-08.

the starting point: licences in place, barely any usage

The client is a mid-sized construction company, around 25 employees, entirely at home in the Microsoft world. Copilot was already there — a handful of people used it. That is the most common state we encounter, and it is not a governance problem. It is a usage problem: the tool sits ready, but nobody has shown people what it is good for in their own working day.

Which is why the brief was not "roll out Copilot" but "roll out Copilot so that it actually gets used".

why copilot — and when we would have advised against it

We pick the platform to fit the client, not our preference. Here the decision was quick: anyone already working entirely inside Microsoft 365 gets the least friction from Copilot. The data stays in the tenant, the surfaces are already familiar, and no additional tool with its own login enters the picture.

With a mixed stack or strict EU compliance requirements we would have advised differently — Langdock or Claude Enterprise are the more honest answers there. And if you are not on Microsoft 365, do not buy it just to add Copilot on top. The full comparison is in our company GPT overview.

mcp: why copilot no longer stops at your m365 data

Support for the Model Context Protocol in declarative agents has been generally available since December 2025. In practice that means a Copilot agent can reach systems that have nothing to do with Microsoft through an MCP server — as if it were a built-in capability.

For this project we built those MCP servers ourselves, so the agents can communicate with the client's external tools. That is the point where a well-integrated chat assistant becomes a company GPT: it no longer just answers questions about documents in the tenant, it works with the systems the actual business runs on.

For the Microsoft world itself, ready-made servers now exist. Everything beyond that you write once — and that effort pays off, because the connection is then available to every further agent instead of ending up as a point-to-point integration. What MCP is technically, and why it became the standard, we explained in AI agents and MCP.

the part that really counts: training

If we take one thing away from this project, it is this: without training, the tool does not get used. Not used badly — not used at all.

That sounds obvious, yet it is underestimated in almost every rollout, because the technology looks so invitingly simple. A chat field looks self-explanatory. It is not. The difference between "I tried it once" and "I use it daily" is not functionality. It is whether someone has shown, using your actual work, which task now goes faster.

So we trained everyone, not just the tech-confident. And next time we would go one step further: people who had little exposure to tools like this need their own format — a smaller group, more time, examples from their own work rather than generic demos.

what keeps us thinking: the day after the rollout

The impact was good. Even so, we do not consider this a closed project but a beginning — and we say so openly, because the honest question is: in six months, will everyone still be using the tool, and using it well?

Three things, we believe, are the answer:

* A monthly exchange. A short, fixed slot where people show what worked for them — and what did not. * A living prompt library. Not a one-off collection handed over at rollout, but a document that grows out of real usage. * Lessons learned that actually flow back. What surfaces in practice has to change the next training session.

A rollout is a date. Adoption is a process. Plan only the date, and in six months you have licences nobody uses — exactly the state we started from.

who this is not for

Copilot is not a universal answer. If you are not working in Microsoft 365, if you deliberately run a model-agnostic stack, or if you need processing strictly inside the EU, another platform serves you better. The platform question is settled after discovery — before that, it would be guesswork.

If you are wondering what a company GPT could look like for you: talk to us. A first conversation costs nothing and is more honest than any off-the-shelf platform recommendation.