Name the cost of the current workflow.
“Reduce time spent copying incoming PDFs into our records” names work you can inspect and a result you can measure.
Follow a record from arrival to completion. Measure hands-on effort and elapsed time. Delays may come from missing information or authority, rather than a task AI can improve.
Name the intended change: more team capacity, faster decisions, fewer corrections, or clearer status. Each needs its own measure.
Choose a workflow you can evaluate.
Compare candidates using the records and exceptions your team actually encounters.
| Criterion | What to establish |
|---|---|
| Business value | What changes if the work takes less effort or reaches a decision sooner? |
| Frequency | Does the work recur often enough to justify implementation and provide a useful sample? |
| Source records | Can you access representative documents, including incomplete and unusual cases? |
| Review effort | Can a reviewer check the output efficiently against its source? |
| System access | Can the result reach the system where the next step happens? |
| Ownership | Who owns the workflow, its exceptions, and the decision to change it? |
A large theoretical benefit is insufficient if access is unresolved or reviewers cannot check the output.
Scope the first implementation around a complete task.
Define the start and completion point, sources, systems, permissions, reviewer, and exception path. Test through the step where the team uses the result.
Set a baseline. Compare total effort, including review and correction, and observe actual use. Keep expected benefits separate from measured results.
Can the team complete the task with less effort and acceptable quality? Use the evidence to expand, revise, or choose another technique.
Further reading
For voluntary guidance on AI risk management, see the NIST AI Risk Management Framework. It does not replace assigning ownership of your workflow.