The builder is on leave. A source document changed, cases are failing and nobody knows whether to retry or stop. The workflow depends on one person's availability.
Photo by My Profit Tutor on Unsplash.
An AI workflow runbook should help a named owner operate, investigate and safely pause the process. Keep it usable during interruptions, with references to controlled technical documentation where necessary.
Start with the boundary
Write what the workflow does, what starts it and where responsibility ends. Name the business owner, support route and backup.
State accepted inputs and actions requiring approval. Reference current instructions, acceptance checks and authoritative sources. Avoid duplicating documents that could drift independently. Include the location of approved access requirements without embedding credentials.
Add a normal-case walkthrough
Use an illustrative case to show expected stages and completion evidence. Identify what staff should see at intake, review and final record.
List routine checks:
Is incoming work reaching the right queue?
Are sources current and accessible?
Are failures visible to an owner?
Are approvals distinct from completed actions?
Are waiting cases receiving attention?
Set frequency according to volume and consequence. A checklist nobody can perform at the chosen rhythm is not an effective operating instruction.
Make failure instructions specific
For common failures, describe recognition, investigation owner and pause conditions. Include a check for partial completion before retrying.
'Check whether the destination record exists before restarting' resolves a practical question. 'Try again later' does not.
Document the manual fallback and who can activate it. Keep specialist configuration with appropriate people rather than offering unsafe shortcuts. Explain how work processed manually returns to the normal route without duplication.
Include change and handover rules
Describe how instruction, tool, source or connected-system changes trigger review. Reference evaluation cases and where release decisions are recorded.
Ask the backup owner to walk through a normal case and failure using the runbook. Update anything needing verbal explanation. That exercise checks whether knowledge actually transferred.
Keep an owner and last-reviewed date for the runbook itself.
Related guidance: guide to mapping the work before changing it.
