Cedar Desk human handoff kit
Article: https://quickchat.ai/post/chatbot-human-handoff-workflow

Cedar Desk is fictional. These are editable setup files for a learning Agent, not a one-click import. No real orders, refund API, customer records or email destination are included.

1. Create a separate Agent named Cedar Support.
2. Add each knowledge/*.txt file as a separate Knowledge Base article. Use its topic as the Title and the entire file as Content. Save and wait for Ready.
3. Paste main-prompt.txt into Identity > Main Prompt, allow autosave, close and reopen to verify it.
4. Apply handoff-settings.txt in Actions & MCPs > Human Handoff > Edit Action and Save.
5. Use the Chat Page and Inbox walkthrough in the article. human-reply.txt is the manual operator's reply.
6. Create the Cedar Desk handoff checks dataset in Testing using Paste messages and testing-messages.txt. Use evaluation-criteria.txt in Evaluation criteria. Check the cost before running.
7. Record your manual checks in acceptance-tests.csv. Observation columns are deliberately blank. Linked steps stay in one conversation; independent cases start new ones.
8. Follow after-hours-test.txt on this learning Agent and restore the normal schedule afterward.

Known local failures: handback answered one AI turn then reverted to Unassigned; the unavailable reply promised future human help without a transfer. These are failed acceptance checks, not passing demonstrations. Repeat them on your target deployment and do not rely on the unverified outcomes.

The final prompt contains the tuning change for policy/warranty/guarantee questions absent from the knowledge. The article explains the observed baseline failure. Small controlled runs do not establish a production accuracy rate. Batch reply scores do not prove ownership, AI silence, human delivery or handback.

Credits and entitlements: AI replies and Testing consume credits. The Free plan supports personal learning; repeated tests may need additional credits. Business deployment needs a commercial plan and a monitored human queue.

Replace fictional facts before use. Do not use a typed name or email as account verification. Test optional routing and email notifications separately if you enable them.

observed-runs.csv contains our baseline and final controlled batch observations, dated in UTC. It is separate from your blank acceptance worksheet. Only the missing-policy sentence changed between the compared prompts. Human-delivery evidence is in the article, not inferred from this CSV.
