I've built six AI colleagues with names, voices and areas of expertise — and now I'm testing how far they can go in real project work. Not as chatbots you ask things, but as colleagues who do the groundwork, produce and soon sit in on meetings too. I test everything in my own company first. The goal is simple: understand what actually works before I recommend anything to anyone.
AI that answers questions is already everyday. What I'm curious about now is the next step: AI that takes on a role — works in several steps, owns an area of responsibility and hands over a finished proposal that a human reviews and decides on.
For the public sector this is both a big opportunity and something that demands the same rigour as any IT project: information security, traceability, GDPR and the new requirements in the EU AI Act. That's why I test it myself, on my own ground, at a calm pace — so no one else has to be the guinea pig.
Why in my own company? Because experiments need the freedom to fail. Advanto is its own lab: here I can give the AI colleagues more and more capabilities, see where it breaks and be honest about it — without any organisation being at stake.
An AI colleague isn't a chatbot with a made-up name. It's built in layers — much like a newly hired consultant grows into the job. Here's what the layers look like, from the ground up:
Layer 1 you can buy. Layers 2–7 are the craft. That's where I spend my time.
No AI colleague gets all its capabilities at once. Each one climbs a staircase, one step at a time — and every step is tested in the lab before the next:
A documented role with a name, background and voice.
Answers questions within its field, in chat and voice.
Produces documents, materials and code to our templates.
Carries out whole tasks on its own and reports back the result.
Joins meetings, takes initiative within its mandate and works with the others.
Today the whole team stands on step 1. Viktor, the developer, is first up to step 2 and on his way to step 3. And one thing is built into the staircase: no AI colleague can raise its own level. That decision is always human.
Can my AI colleagues sit in on a real Teams meeting? Hear who in the room says what, follow the context and chip in when spoken to — with their own voices? I'm building it as a shared meeting presence: the team joins as one participant, and a "conductor" decides in the background whether and who should speak. Most of the time the answer is "keep listening" — it's the ability not to speak that's the hard part. Every meeting starts with the team introducing itself as AI and asking permission.
Can an AI colleague compile status reports and steering-committee materials from project data — so project managers can spend their time on leadership instead of admin? The templates already exist in my project methodology. Now I'm testing letting Elias, the team's project-management assistant, fill them in from real project data — with a decision log where no decision is missing an owner and a date.
Compiling materials, analysing documents and supporting risk reviews in IT projects — areas where I know exactly what the work looks like today, and where a human always reviews and decides. Karin, the team's requirements analyst, now has a written role description with a clear mandate: which questions she should be able to answer, and when she should hand over to me. For government agencies and regions, this needs to work in the long run inside the environment they already live in, often Microsoft 365 — that's part of what I'm evaluating.
Can the AI colleagues work in the same tools as human colleagues — version control, testing tools, operating environments, documents? I test it with the same principle that applies to people: personal permissions, least possible access, and everything logged. Viktor is first out: he gets to build and test freely in the development environment, but anything touching live operations requires my approval. Every colleague's access is gathered in an internal staff register — who may do what, in which system, and where the line runs.
I'm looking for a government agency or region that wants to test from a real, bounded need — with the same structure and documentation as in an ordinary IT project. I promise to be honest about what's mature and what's still an experiment.
A first conversation costs nothing. I'm glad to tell you what I've seen so far — what works and what doesn't.
