AI for small and mid sized companies
Describe the work.
Then let agents do it.
Task Architecture is a practical way to bring AI into everyday work. Start from the task, not the tool, and let the people who know the work best describe it precisely enough for an agent to carry it out.
“We all have to become a bit like software architects. Not for software. For our tasks.”
An employee in her late fifties said this in the back office of a mid sized healthcare company, in the middle of a conversation about AI tools. It is the whole idea in two sentences. Everything on this site follows from it.
Four principles
What changes when you start from the task.
The point is not another AI framework. It is a change in sequence: understand the work first, capture the experience inside it, and only then decide what technology belongs there.
Task before tool.
Every vendor solves one small problem. Together they build a zoo nobody can keep track of. Describe the task first, then decide what an agent should do and where a person stays in.
Experience is the raw material.
The people who have done the job for twenty years know every exception. Written down as a task, their knowledge works at many desks at once and stays when they retire.
We need more coal.
In 1865 Jevons saw that efficient engines raised coal demand instead of lowering it. Cheaper work gets done more often. The demand moves to people who can describe and steer it.
Software in batch size one.
A tool for one process in one company used to be economically absurd. Today it can be a prototype by morning. Open the drawer of ideas that were never worth it.
Proof from practice
The routine went to the machine. The knowledge became a sales channel.
When the chatbot Billie took over routine questions, Ingka Group trained call centre staff as remote interior design advisers. The interesting part is not the chatbot. It is what happened to the human knowledge once routine work became cheaper.
Sources: Reuters, June 2023; Fortune, July 2026.
Start here
The one hour exercise.
No strategy paper. No tool demo. Take one hour with your team. You will learn more about your readiness for AI than from any vendor presentation.
Pick one recurring task.
Describe it so precisely that a new colleague could do it without asking a single question.
Ask for the exceptions.
Sit down with your most experienced colleague. Not the rules. The exceptions. Write every one of them down.
Open the drawer.
List the tasks you once gave up automating because the software was too expensive. Look at them again.
Read the series
Four posts. One article.
About
Thomas Eusterholz
I advise small and mid sized companies on AI, digitalisation and the software they already run. Before that I spent years as a managing director, so I know what happens to a good idea between the strategy meeting and Monday morning.
Task Architecture is where I collect what I learn about bringing AI into real work, together with the people who do it.
Find me on LinkedIn