RoutineSmithKeep the workflow controlled after the pilot.
Record every material change, the evidence behind it, whether tests must run again, who owns the decision and how the team will recover if the change fails.
Download the free Excel logWorks locally. RoutineSmith does not receive what you enter.An AI workflow maintenance log records what changed, why it changed, which operating boundary is affected, whether earlier tests remain valid and who owns the next decision. Open an entry before approval, not after the new version has already reached live work.
Enough detail to make the next permission decision inspectable.
Version and date
Identify the operating version before and after the change, plus the date it was observed.
Material change
Record changes to the model, prompt, rules, data, permissions, integrations, vendor, cost or incident response.
Evidence
Link the release note, test case, incident record, configuration diff or reviewer observation that supports the entry.
Affected boundary
State whether the change touches inputs, outputs, human review, external actions, data access, cost or reliability.
Retest decision
Choose whether the frozen test set must run again and retain the result before expanding permissions.
Owner and recovery
Name the person responsible for the decision and the exact rollback or follow-up path.
Review the evidence when the operating boundary changes.
- 01
Prompt, model or threshold
Retest the cases whose route, output or human gate could change.
- 02
Permission or integration
Review the action boundary and confirm the workflow cannot reach an unintended system or destination.
- 03
Data or knowledge source
Check provenance, missing context, retention and whether earlier test evidence remains representative.
- 04
Vendor terms, privacy or cost
Recheck the operating assumption and record whether continued use still fits the approved boundary.
- 05
Incident, override or repeated correction
Pause or narrow the workflow until the cause, correction and regression case are documented.
The log preserves a decision trail. It does not grant authority.
- Keep credentials, customer records and confidential source text outside the workbook.
- Do not mark a retest as passed without retaining the tested version and evidence.
- Do not expand permissions while a required retest, owner or recovery path is missing.
- Keep irreversible actions manual when correction or rollback is unclear.
Deployed AI workflows need continuing review.
NIST calls for post-deployment monitoring, incident response, recovery and change management. ISO/IEC 42001 describes a management system for maintaining and continually improving AI use. This template translates those broad principles into a lightweight operating record for a small team; it is not a certification or compliance guarantee.
Questions about the maintenance log.
Is the maintenance log free?
Yes. The editable Excel workbook downloads without an account, payment or email address.
When should we add an entry?
Add one before approving a material change and after an incident, repeated correction or vendor change that could affect behaviour, access, cost or risk.
Does every entry require a full retest?
No. Record the decision either way. Retest when the change could affect outputs, controls, data handling, external actions, reliability or operating cost.
Does a recorded change approve production use?
No. The log preserves evidence and ownership. The named decision owner remains responsible for approval, pause, rollback or retirement.
Should confidential data go in the workbook?
No. Use references, identifiers and approved evidence locations. Do not paste customer records, credentials or confidential source text into the template.
Learn the build, test and recovery path.
RoutineSmith teaches practical workflows with explicit human gates, test evidence and fail-closed behaviour. Review the curriculum only if you need guided implementation.
Review the practical course