> ## Documentation Index
> Fetch the complete documentation index at: https://docs.trellistech.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Launch and hand off your AI Agent

> Test the AI Agent in review mode and define when a person must take over

## What this helps you do

Launch an AI Agent with clear information, safe review, and a dependable human
hand-off. Start with one property and one channel.

## Before you start

* Complete the [Quickstart](/quickstart) through the channel test.
* Decide which teammate owns escalations.
* Prepare one property fact, one company policy, one repeat procedure, and one
  request that must go to a person.
* Keep automatic replies off during testing.

## Prepare the source information

Use one source of truth for each kind of information:

| Information                             | Put it here        |
| --------------------------------------- | ------------------ |
| Property facts such as parking or Wi-Fi | Property           |
| Company policies                        | Document           |
| Voice, limits, and hand-off rules       | Agent instructions |
| Repeat procedures                       | Skill              |

## Test in review mode

<Steps>
  <Step title="Open the agent">
    Go to **AI Hub**, choose the agent, and review its instructions and access.
  </Step>

  <Step title="Test a property fact">
    Ask a question whose answer belongs to one property. Confirm the answer uses
    that property's current details.
  </Step>

  <Step title="Test a policy and procedure">
    Ask one question answered by a company document and one that should follow a
    skill. Confirm the sources do not conflict.
  </Step>

  <Step title="Test a hand-off">
    Use a safe example involving a complaint, refund, safety concern, or another
    case your team requires a person to review. Confirm the agent pauses or
    prepares a draft instead of making the sensitive decision.
  </Step>

  <Step title="Review the conversation">
    Check tone, facts, recipient, property, and requested action. Correct the
    source information, not only the draft.
  </Step>
</Steps>

## Decide when a person takes over

Write short hand-off rules in the agent instructions. Name the situations, the
person or team responsible, and what the agent may do while waiting. Common
hand-offs include safety issues, money, policy exceptions, angry guests, owner
decisions, and uncertain property facts.

## Move beyond review mode

Only consider automatic replies after your team has repeated the tests on real
approved examples and reviewed the results. Expand one channel or property at a
time. Keep sensitive actions behind approval.

## How to check it worked

* Each test uses the intended source.
* The agent follows your tone and limits.
* The hand-off reaches the responsible person.
* No sensitive promise or action happens without the required review.
* A manager knows where to inspect the conversation and improve the source.

## Common problems

<AccordionGroup>
  <Accordion title="The answer is generic or wrong">
    Follow [Fix wrong AI information](/guides/fix-wrong-ai-information) and
    update the source that should own the answer.
  </Accordion>

  <Accordion title="The agent cannot see the property or conversation">
    Review the agent's property, Inbox, and tool access in AI Hub.
  </Accordion>

  <Accordion title="The agent acts when it should ask">
    Tighten the hand-off instructions and require approval for the action. Test
    the same case again in review mode.
  </Accordion>
</AccordionGroup>

## Related articles

* [AI Hub](/platform/ai-hub)
* [AI Agent](/platform/ai-agent)
* [Approve a sensitive message](/guides/approve-sensitive-message)