Department of 10 · Setup 4 of 5
Quality checks
Where should human judgment stay in an AI-supported routine? At the gates a polished draft can still fail. Quality means accepted work with fewer avoidable errors.
1InputsAre the records current and complete?If not: draft blocked
Old pricing or empty notes block the draft. One deal ID gets one open work item, so a duplicate run does not start a second one.
2Scope and pricingDo the terms match the approved source?If not: corrected first
Sends, pricing and unusual terms all need human approval. In the fictitious sample deal, the rep corrects one price line before approving.
3TemplateIs the correct version being used?If not: back to current version
Managers approve templates. Prompts, context, schemas and scripts are versioned, so the team can tell which version a run used and restore a safe one.
4ExceptionsDoes this case need escalation?If yes: goes to a manager
Any item that fails a check moves to Exception for a manager. A permission failure stops the run and logs it. It never borrows wider access.
5Human approvalHas the authorized person accepted it?If not: nothing goes out
A run log for every job records its source, reviewer and error, so approval can be checked later. The draft sounding good is not the test.
Back to gate 1: learn from reworkFix the source, not only the draft
For example: a proposal draft looks polished but uses an outdated delivery assumption. The reviewer catches it, corrects the source and records the issue so it does not repeat.
When something breaks
DetectMonitor flags it
Failures and exceptions are monitored daily, using the run log.
PauseAgent stops
Stopping first keeps one fault from reaching more work.
RestoreLast safe version
This is why prompts, context, schemas and scripts are versioned.
ManualPeople handle the queue
The manual fallback is documented and trained, so the work continues while the fault is fixed.
- Connected toolsSee where the approval sits in the handoff
- Team trainingSee how reviewers learn these checks
Failure recovery, common failures and all checksThe original recovery path and control list
Quality checks
When something breaks
- DetectMonitor flags it
- PauseAgent stops
- RestoreLast safe version
- ManualQueue handled by people
Common failures
- Stale or missing
- Old pricing or empty notes block the draft.
- Duplicate runs
- One deal ID gets one open work item.
- Permission failure
- The run stops and is logged. It never borrows wider access.
Checks
- Human approval for sends, pricing and unusual terms.
- A run log for every job with source, reviewer and error.
- Daily monitoring of failures and exceptions.
- Prompts, context, schemas and scripts are versioned for restore.
- Manual fallback is documented and trained.