AI only helps once the task has a name.
The first useful AI use case does not start with the model. It starts with a recurring work step you can measure. Optimistic — with a tape measure.
Read the German original
Do not start with the model
The temptation is familiar: new tool, big promises, a use case will appear somehow. We reverse the order. Look for a recurring step where people search, transfer, or prepare information. Who does it? With which inputs? How do you recognise a good result? Only when those three answers stand does the technology question earn its place.
Example: a service team leafs through approved manuals for internal follow-up questions. A possible assist: suggest matching passages. The team checks, rewrites, decides. Small scope, clear ownership — and precisely therefore an honest trial. Whether it holds in your operation shows only practice, not the demo.
Measure the everyday work, not the wow effect
Before anything is connected: how does the step run today? Time for search and rework, typical mistakes, waiting. That is your tape measure. An elegantly phrased answer is not yet proof that work got easier.
For the manual example, what counts are relevant finds, traceable sources, and how costly the final check is. Also decide when the trial ends or gets rebuilt. A good outcome can mean: we stop. That is progress, not failure.
Data and alternatives — before the spark jumps
Before the first transfer, clarify which information the tool sees, who is allowed to, and where processing happens. Early trials often need carefully chosen sample data. A local model still needs permissions and operations; “it runs on our machine” is not a concept.
And check the unromantic alternative: better search, clearer forms, fresher documents. Sometimes the problem is dusty knowledge — then AI would only be varnish on a wobbly board. Name a business owner. When the manual changes or a permission disappears, that must reach the assisted flow. Otherwise you build elegance on sand.
Test the uncomfortable cases
A serious trial includes the normal case, contradictions, and empty hands: no matching answer. Decide how the system shows uncertainty. Check results with the people who actually know the work — their follow-up questions are often the best specification.
Look at the whole path: prepare, use, control, correct. Speed is fine as long as review stays proportional to benefit. And when something goes wrong, people need a clear way back to the previous way of working. No dead-end magic.
End with a decision, not applause
The NIST AI Risk Management Framework 1.0 links context, measurement, accountability, and ongoing risk handling. For a first step we compress that into a working question: do we have enough evidence to keep building this use case responsibly?
Record what holds, where limits sit, and who owns open points. Only then: a larger trial, a different solution — or clean up the data first. AI can be an extraordinary booster. Used with ownership it becomes craft. Left unchecked it stays theatre.
What does this look like in your work?
We can turn the idea into a sensible next step with you.
AI & Linux workplace