When is a use case worth building?
Most AI ideas don’t die in the demo. They die months later, in delivery, when it turns out no system holds the data the answers must come from, nobody can prove the thing saved a franc, or nobody owns the process it was supposed to improve. All three were knowable on day one, and three questions surface them: grounding (which system holds the truth this needs, and who controls access?), measurement (which number moves, and can you read it today?), and ownership (who owns the process, and who decides?). Pass all three and the idea deserves serious effort. Fail one in principle and you should kill it before it consumes a budget, however good the demo felt.
That is the whole test in one paragraph. The AI use cases track is the working version of it: what each gate actually asks, whether you have an assistant or an agent, how far to let it act, whether to buy or build, and how to measure the value without fooling yourself.
These gates are how we qualify sniper-shot engagements at machtsinn: three questions in the first call, evidence before the blueprint, and a written measurement method before any build starts.