The problem
“We should use AI” is too broad to guide a useful experiment. Start with a recurring task where the current process is understood.
A method you can use
- List three recurring tasks and their owners.
- Record frequency, typical effort and common errors.
- Choose one task with approved inputs and an output a person can verify.
- Test one step, using a clear quality threshold and a stop condition.
- Compare the outcome with the baseline before expanding.
For business leaders
Choose a bounded experiment with an accountable owner. Ask what happens if the output is wrong and who makes the final decision.
For practitioners
Write the input, instructions, expected format and review checklist. Try varied examples and document failure cases.
Illustrative example
A team drafts internal meeting summaries from approved notes. A person verifies every action, owner and date. This example does not claim measured savings or client results.