Humans stay on judgment calls — publish, approve registry changes, resolve ambiguous recruiter fit — not on repetitive glue.
In my daily workflow
- I label which workflow steps are automated vs require approval.
- Draft content stays in admin until I explicitly publish.
- High-impact copilot mutations confirm before commit.
- I review edge cases flagged by users or low-confidence scores.
How it makes me work smarter
HITL prevents silent corruption. The CMS status dropdown (draft / published / archived) is a human gate the RAG index respects. Automation handles embeddings and revalidation; I handle what represents me publicly.
My setup
- Admin auth on all write paths
- Draft-by-default knowledge content
- Confidence thresholds routing to review queues
- n8n or Slack notifications for approval steps
On this portfolio
Knowledge status select in admin controls what the world and the assistant see. Copilot registry updates run through validated tools but still assume an authenticated operator — not anonymous visitors.
