Skip to main content
Correct the information behind an AI answer so the next reply uses the right source. Editing one draft fixes that message; updating the property, policy, or procedure addresses the recurring problem.

​
Find the source that should own the answer

Inspect an AI Trace in Trellis
Open the affected reply’s AI Trace or available sources. Read the records it used and compare them with the claim that was wrong. Keep the conversation, property, question, and time available for the retest. Check the source that owns the information:
  • Property: parking, Wi-Fi, access, amenities, and arrival instructions.
  • Policy: how to handle a situation, including exceptions and escalation.
  • Agent instructions: overall role, voice, and stable limits.
  • Document: reference information retrieved for the answer.
  • Skill: steps for carrying out a repeatable procedure.
  • Reservation or PMS: booking dates, guest details, and stay status.
If two sources disagree, correct the outdated copy as well as the primary source.

​
Correct and retest

​
Give a specific correction

Give specific feedback with Teach AI in Trellis
Choose Teach AI on the reply. Explain what was wrong, give the correct fact or behavior, and state its scope. For example: “For this property, rubbish goes in the courtyard bins beside the bicycle storage. Update this property’s disposal instructions.” Trellis can process feedback in the background. Answer any clarification it requests. Some feedback describes an action for a booking, not a reply. Examples are “send the pet rules to every booking with pets” and “bookings with infants get the infant template”. The AI replies only when a guest writes, so Teach AI does not save this kind of instruction as a reply rule. It proposes a draft workflow that starts from the booking instead. Review the draft and confirm it. The workflow does not run until you activate it. If the new rule contradicts an always-on policy or a knowledge document, Teach AI shows the conflict and asks which version to keep. It saves nothing until you decide. If Teach AI cannot finish, it tells you that nothing was saved. Choose Retry to send the same correction again.

​
Check the saved change

Review the resulting change in Trellis
Open the property, policy, or instruction the feedback says it updated. Confirm that the wording is correct and applies to the intended property or audience. Do not assume the source changed just because feedback was submitted.

​
Test the answer again

Repeat the original question with the same property and reservation context while the reply is under review. Inspect the new sources and answer. Then try one related case, including another property if the correction should be limited to one home.
Check sources on a sent message in Trellis
On a reply that cites a source, open the citation to read the actual record. Check that the answer used the corrected fact and did not introduce a new promise or exception.

​
Correct an already-sent reply

Teach AI from an existing reply in Trellis
You can also use Teach AI from an existing sent message. Review the resulting source change in the same way. This does not retract or edit what the guest already received; the conversation owner should send a separate correction if the guest needs one.

​
When the source is a reservation

Compare the reservation in Trellis with the connected PMS and follow the integration’s supported update or refresh process. Confirm the corrected booking detail appears before retesting the agent.

​
Common problems

Check the property and reservation selected for the test, then inspect the new trace. Confirm the agent used the source you changed and that no conflicting policy or document remains.
Review property facts and policy scope. A property-specific exception should not become a workspace-wide instruction.
Keep the example for support with the conversation, time, question, expected answer, and available trace. Remove unnecessary private details from any shared screenshot.