Start with the job
Define the workflow before the AI category
Turn a vague request for an AI tool into a bounded workflow with a trigger, inputs, review point, accepted output, and fallback.
What this guide should leave behind
The team can compare an assistant, embedded feature, automation, retrieval system, or custom build against the same observable job.
Work the evaluation in this order
- 01
Choose one frequent job and record the trigger, current inputs, responsible person, current steps, required result, and known exception path.
- 02
Mark which step needs generation, extraction, classification, search, recommendation, routing, or summarization. These are different technology needs.
- 03
Set a human review point and define what the reviewer must be able to see, correct, reject, or reproduce.
- 04
Write the current baseline and the evidence a pilot must improve without creating a new unacceptable risk or operating burden.
Evidence worth seeing
- The workflow has one accountable business owner.
- The requested AI capability is described as a task, not a marketing label.
- Exceptions and low-confidence cases have a safe route.
- The current manual or non-AI fallback remains usable during the pilot.
Know when general guidance stops
Pause general evaluation when the workflow can approve credit, employment, health, insurance, housing, education, legal, safety, or other consequential outcomes without qualified review.