Look for friction you already recognize
Your first useful workflow is probably close to your existing work. Think about a task you repeat, understand well, and find slightly tedious: organizing notes, drafting an internal update, comparing document versions, or turning rough ideas into an outline. Familiarity matters because you need to judge whether the output is useful.
Choose a manageable experiment
Pick a task with clear inputs, a visible output, and limited consequences if the draft is wrong. Avoid connecting multiple applications before you understand the basic workflow. A manual copy-and-review process is enough to test the idea. Write a one-sentence goal and define what would make the experiment worth repeating.
Record the baseline
Complete the task normally once and record the time, revisions, and quality concerns. Use a comparable example for the AI-assisted attempt. This is a small personal experiment rather than a controlled study, so treat the result as a useful signal instead of a universal productivity claim. Record anything that makes the two attempts meaningfully different.
Break the task into steps
List what information goes in, what transformation is needed, how you will check it, and where the finished work goes. Decide which parts AI can help with and which need your judgment. For a weekly update, AI might group notes and draft headings while you choose priorities, verify status, and approve the final message.
Try a second case
One successful result can be luck or an unusually easy input. Test a messier example with missing information or conflicting notes. Does the workflow flag uncertainty or hide it? Revise the instructions and document the failure. Repeated problems are useful evidence that a task needs a different approach or more human involvement.
Save a recipe you can understand
Capture the purpose, allowed inputs, instruction, output format, review checklist, and stop conditions. Name the recipe after the result, such as “Weekly project update,” rather than the tool. Add the date and tool used so you know what was tested. Keep it short enough that you or a colleague can follow it next week.