My most useful AI experiment began with a repetitive task: turning a long meeting transcript into a short set of unresolved questions. It did not need a chat interface or a complicated agent. It needed a predictable input and an output I could inspect in a minute.
Make the review part of the tool
I kept a link to the supporting passage beside every proposed question. That simple choice changed how I used the output. Instead of deciding whether to trust a fluent summary, I could check the evidence and edit the wording before sharing it.
A useful tool saves attention without hiding responsibility.
I judge these experiments by the work they remove from an ordinary day. If preparing the input takes longer than doing the task, the design needs another pass. Small scope makes those tradeoffs visible and leaves room to stop when automation is not helping.