Pick one task with one owner, and leave the ordering to people. One possible first test prepares supplier confirmations for review: proposed updates, unresolved fields and a draft question for the supplier. It sends nothing and changes nothing.
What if the team is uneasy?
Don’t treat concern as resistance. Ask which work they want improved and what support they need, and agree what the test will and won’t change.
1. An example scope: supplier confirmations.
- Input
- Approved supplier replies and a read-only snapshot of selected open purchase orders.
- Output
- A proposed update with its source, the matched line, unresolved fields and a draft question for the supplier.
- Boundary
- No messages sent to suppliers, no new supplier selected and no order changed during the first test.
Practical check: If fixed rules are enough, use them.
2. Include the cases you hope not to see.
Normal. A clear confirmation with the right order, line, unit and date meaning.
Exceptions. Partial quantities, a revised price, an unclear date and conflicting versions.
Failures. A duplicate message, a missing identifier, an out-of-date record and an unavailable connection.
Practical check: Test that unsupported actions stay blocked.
3. Define success before the run.
Agree field accuracy, correct matching, useful handling of exceptions and total review effort, and compare with the current process. Use known outcomes first. Repeat important cases, and look at the failures, not only an average score.
Practical check: Don’t choose the pass criteria after seeing the results.
4. Name the owners and end with a decision.
Agree who approves results, who controls access and how to stop it, then finish with one of four outcomes: keep, change, expand or stop. Adding live access or actions needs its own review and a new test plan. How to Run a Small AI Agent Pilot covers the limits and the decision record.
Read more: Can an AI Agent Change a Purchase Order? Approve Every Change First · Does Your Supply Chain Need AI, or a Better Handoff?
The scope is an example. A small test gives evidence about that scope only. This is general guidance, not advice for your situation.