Start with the workflow, not the model
The most useful starting point is a recurring business process with a clear owner, known inputs, and an outcome people already understand. Map the process as it works today, including the unofficial spreadsheets and manual checks.
AI is usually one part of the solution. Integrations, business rules, permissions, and exception handling often matter just as much.
Look for bounded judgment
Good early opportunities involve classification, summarization, extraction, drafting, or retrieval where a person can review uncertain results.
- The work repeats often enough to learn from
- Inputs are available and permitted for use
- A reviewer can recognize a good result
- Failure is recoverable and visible
Build the smallest complete loop
A narrow workflow that reaches a real outcome teaches more than a broad demo. Connect the inputs, produce the output, route exceptions, and measure whether people actually use it.
Expand only after the first loop is reliable and the team understands its controls.