Adoption is the hard part: six months of MCP in production at an HVAC company

Most "AI in the enterprise" stories come from tech giants. This one comes from a 300 to 400 person Dutch HVAC company (heating, cooling and ventilation), a hands-on technical engineering business with a small IT team, where AI was brand new. I built a production MCP platform there (nine servers, 50+ tools, the first live for months and the newest shipped just last week), mostly on company time and alongside my other work, with my manager backing it early. Those servers span the whole company: the ERP, our pre-order calculation history, the BIM models, building automation, and the energy use of the buildings we run. Data that used to sit in separate silos, each behind its own login and its own specialist, is now reachable through one agent in plain language. The people using it every day are project managers, engineers and back-office staff, not developers. By the time I'm on stage it will have about six months of real use behind it, so this is a report, not a roadmap.

The reason a small IT team could even build this is leverage, not budget. The agent does the driving and is the interface, so there's no chat UI to build or maintain. MCP is the layer that both unlocks the data and carries the domain guidance, so the agent doesn't just reach the data, it knows what it means. Our existing identity provider handles security, the same roles every other system already enforces. And it all runs on the cloud we already pay for. Smart use of MCP turns minimal investment into enterprise-grade reach, which is what puts this within range of a company our size.

The honest centre of the talk is adoption, because that's the part nobody shows. I'll put the real numbers on screen: how it grew from a handful of early users to around 70 unique people making their own tool calls, and the shape behind that number, because totals hide the truth. Some reach for it daily, some only now and then, and some have only just started. This is still an introduction phase, not a victory lap, and I'll be straight about what pulled people in and what quietly stalled.

Adoption didn't happen by mandate; it happened through people. I'll tell the stories behind the numbers: the manager who became a champion, the early believers who pulled their colleagues in, the demos that landed, and the times I put in extra effort off the clock to solve one person's specific problem and win them over by the next day. These are the concrete strategies that turn a tool you built into a tool people actually open on a Monday morning.

And the thing that made it usable at all is domain knowledge. A clean ERP is useless to an agent that doesn't know what the fields mean. That knowledge isn't hand-written: an AI explores the real data to surface the patterns, and the domain experts who actually understand our systems validate what's true, before it gets encoded directly into the tools. The result is a non-technical colleague can ask the ERP a question in plain language and get an answer a domain expert would recognise as correct. Meaning is half of it; readable presentation is the other half. Instead of a wall of text, results come back as charts, tables and maps the server renders right inside the conversation, which is what turns "I asked the ERP" into something a project manager trusts and opens again the next day.

You'll leave with a realistic picture of what AI adoption looks like in a normal company rather than a tech giant: the champion strategies that drive it, the role domain experts play in making it trustworthy, the honest numbers six months in, and how to start doing the same.