How a 600-Clinic Finance Team Builds Its Own AI Forecast
3 Sept 2026
600 clinics. 4 people in the data team. And finance now builds its own cash flow forecast.
600 clinics. 4 people in the data team. And finance now builds its own cash flow forecast.
They are doing more work than they could before but monthly closing takes six days only. The target is two.
After giving the organisation access to Claude, the finance team started building its own cash flow forecast. M&A tracks its own pipeline. Procurement built a price analysis tool and takes it to meetings.
Far fewer tickets.
That was one of three talks at last night's Claude Community meetup in Zürich by Kilian Dumortier from Colosseum Dental Group.
Third edition, and the room was full again.
Two other things I am keeping:
→ Databricks, Patricia Ferreiro and Vlad Marin: a new model comes out every week and nobody wants to rebuild their setup each time. Their open source tool Omnigent swaps the model or the coding agent underneath without touching the code. Cost limits and permissions stay in place.
→ Caffeine AI, Habibou Sissoko: do not let the builder grade its own homework. In other words, don't ask your AI to revise its own output. One agent builds. A second starts with fresh memory and checks the work against tests it did not write and cannot change. His rule: give the agent room to optimise the implementation, never permission to redefine success.
The questions afterwards ran longer than the talks, I absolutely going down on the techie rabbithole, every time I go to a Claude meetup!
PS: if you have great Claude Code stories from your company, and you'd love to share, drop me a message. Looking for speakers for a next Claude for Leaders event in ZH.
Arnaud Delacour thank you for keeping this going. And it was lovely catching up Andreea Stanescu!
Which team in your company is still doing manually what AI could do in minutes?
Want to think through what AI means for your own leadership? My Private AI Advisory is here.
What separates real AI adoption from a pilot that never scales
The detail that stands out in the Colosseum Dental Group story is not the technology, it is who built the tools. Finance built its own forecast. Procurement built its own pricing tool. Nobody waited for a central AI team to hand them a finished product. That is the actual difference between AI adoption that scales and AI adoption that stays a pilot forever, ownership sits with the team that has the problem.
The second pattern from that evening matters just as much: don't let the builder grade its own homework. I hear a version of this mistake constantly in the organisations I advise, one AI tool produces the output and the same tool checks whether it is any good. A second, independent check, whether it is a person or a separate process, is what keeps quality from quietly drifting, and it is close to what senior leaders actually need from AI in 2026.
If you are trying to move your own team from pilot to real use:
→ Give the team closest to the problem the access and the permission to build, not just a finished tool
→ Separate who builds from who checks, every time
→ Track a number that matters to the business, like closing time, not just adoption or usage
If you want practical starting points for your own team, I have a free AI Tools & Tips Guide with the tools and workflows I use as a coach and business owner. Leave your email here and I will send it over: https://www.thechangerepublic.com/free-resources