Use this worksheet when an AI workflow stops after changing something, or when its final message leaves the result uncertain. Check the affected system before deciding what to repeat.
Duplicate this page for each review. Fill in the blank review and keep the worked examples collapsed. When testing a workflow, use disposable data. Keep passwords, access tokens and personal customer information out of shared copies.
| Field | Your notes |
|---|---|
| Requested work | Describe the task and the intended final state. |
| Run reference | Add the run reference and when it happened. |
| Systems affected | Name the systems that may have changed. |
| Reviewer | Name the person checking the result. |
| Permission for the next action | Record who may authorise a retry or a change. |
Duplicate this section for each operation that could change an external system. Check each step separately, including those before the reported failure.
| Field | Your notes |
|---|---|
| Operation | Describe the change this step should make. |
| Before and expected after | Record enough detail to recognise the intended result. |
| Observed result | Choose Confirmed happened, Confirmed did not happen, or Unknown. |
| Evidence | Record what the affected system shows, its record or message reference, and when you checked it. Link to evidence only where the reader has appropriate access. |
| Effect of repeating the step | Could this create another record, send another message or overwrite newer work? Record how you would check. |
Use Confirmed happened when the evidence establishes the intended change. Use Confirmed did not happen when reliable evidence establishes that the change did not take effect. Keep Unknown when the evidence cannot settle the question. A recorded tool call or an empty search result alone may leave the outcome uncertain.
Before another attempt:
Next action: Describe the smallest action that addresses the remaining work, or the check needed before choosing one.
Owner: Name the person who will take that action.
Expected result and check: Write what the affected system should show afterwards and how you will verify it.
Result after the action: Record the evidence and any remaining uncertainty. Keep the review open while the outcome is unresolved.