Debug a Workflow by Finding the First Wrong Step

The customer asked for next month. The final summary says tomorrow. Rewriting the prompt might help—or the system may have sent the wrong message to the model in the first place. Find where the result first became wrong before changing the whole workflow.

Follow one record Choose a fictional or appropriately authorized example with a stable ID. Follow the stages in the workflow map: source, validation, transformation, optional AI step, destination, and notification. At each stage, record expected input, actual input, expected output, and actual output. Preserve the link to the original tracker record.

Use the mismatch to narrow the cause If the source says next month but the AI input contains an old example saying tomorrow, investigate the mapping or prompt assembly. If the AI input is correct but its summary invents tomorrow, investigate the AI instruction and output checks. If the summary is correct but appears in the wrong customer record, inspect destination IDs and routing. That distinction saves you from adjusting a prompt to compensate for a problem elsewhere.

Control side effects before retrying Pause customer sends and avoid rerunning actions that create records until you understand what already succeeded. Use the approval boundaries during recovery. Zapier’s replay documentation explains recovery options; its duplicate-data guide helps with repeat records. Neither replaces checking your own run history. If reminders were involved, inspect the stop conditions. If the notification failed, inspect the handoff state rather than creating a fresh inquiry.

Make one change, then retest Save a minimized example, expected result, configuration version, fix, and retest outcome. Use the test log so the failure becomes part of your regression checks—the tests you rerun after changes. For working documentation, consult Google’s version-history guidance. Keep actual tool configuration changes identified separately.

Leave the next person a useful note Write “The destination received the previous status because field X was mapped incorrectly,” not “AI was broken.” Add the recovery instructions to the maintenance card. Keep private customer details out of public troubleshooting posts. Your useful result is one identified cause, one checked fix, and a test that catches the same failure next time. If you cannot reproduce the issue safely, stop replaying and investigate the evidence you already have.