Begin with the task, not the AI tool. Describe the steps people perform, the information they use, and the decision or output the task should produce. If the process is unclear or changes every time, clarify it before trying to automate it.
Look for work that is frequent and bounded, and decide what a useful result would be. Consider what happens when the system is uncertain, receives incomplete input, or produces an incorrect answer. A person should know when to review, correct, or stop the workflow.
Check the information involved. Confirm that the organization is allowed to use it for this purpose, that access is controlled, and that sensitive data is handled appropriately. Also identify who will maintain the workflow and review its behavior over time.
A small pilot can help reveal edge cases before wider use. Treat it as a way to learn: define the limits, involve the people doing the work, and decide in advance what evidence would justify continuing or changing direction.
Bound the task and plan for review
Describe the input, expected output, and limits of the workflow. Decide which cases are routine and which must go to a person. Include a way to correct an output, pause the automation, and record when the result needs review.
Evaluate the workflow on representative examples before expanding its use. Include incomplete, unusual, and ambiguous inputs. Consider the quality of the output, the time needed for review, the consequences of an error, and the handling of the information involved. Revisit those checks when the workflow or model changes.
- Choose a repeated task with a clear start and end.
- Decide what requires human review or escalation.
- Check quality, privacy, and failure handling in a limited pilot.
Add your perspective