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

# Test on past tickets

> See what the Support AI employee would have proposed for tickets your team already solved, judge each reply, and choose a confidence threshold.

Before you let the Support AI employee prepare replies, run it on tickets your team already solved
and compare what it would have proposed with what your team actually replied.

> [!WARNING]
>
> AI employees are in **beta**, so you will not see this unless your deployment has turned them on.
> **It has not yet run on a real Zendesk, Freshdesk or Intercom account.** See [AI
> employees](https://docs.mithunai.com/team/overview.md).

## What a test does, and does not do

A test reads your most recently solved tickets, up to 10, from the helpdesk connected to the Support
inbox -- solved and closed in Zendesk, *Resolved* and *Closed* in the last 30 days in Freshdesk, closed
conversations in Intercom -- and asks your chosen assistant once for each. **It only reads.** Nothing
is sent, filed or tagged in your helpdesk, and nothing appears in your
approval queue.

Each ticket uses one answer from your monthly allowance, and appears in Conversations as asked by an
AI employee. You can run one test at a time, and up to five a day. Results are kept for 14 days.

## What you see for each ticket

* What the customer wrote, and what your team replied.
* What Support would have proposed: a drafted reply, or that it would go to a person, and why.
* The sources the reply cites, and its **confidence**.
* Three buttons to judge the reply: **Good**, **Needs edit** or **Wrong**. Judging needs a person who
  may approve actions.

## What confidence means

Confidence is how closely the strongest source a reply cites matches the question, as your knowledge
search measured it, from 0 to 1. It is **0** when your knowledge does not support an answer. It is
**not** the model grading itself, and it is **not** a guarantee that the reply is right: a close source
can still be answered wrongly. That is why you calibrate a threshold against your own verdicts.

## Choose a threshold

Move the slider to see what a threshold would have done on the latest test: how many tickets would
have got a drafted reply, how many would have gone to a person, and how your reviewers judged the
replies that would have been drafted. **Save threshold** applies it to live work.

With a threshold saved, a reply below it is not drafted. The ticket is handed to a person instead,
and that hand-off adds the tag `needs_a_person`, so a view, trigger or rule in your helpdesk can route
it.
**Either way, nothing is sent without a person's approval.** A threshold can only turn a drafted reply
into a hand-off; it can never send anything.

The threshold is a setting of your helpdesk connection. If you detach the helpdesk, attach it again or
switch to another, set it again. Only an owner or admin, signed in, can set it.

## Related

* [Work that arrives by itself](https://docs.mithunai.com/team/work-that-arrives-by-itself.md)
* [Support](https://docs.mithunai.com/team/support.md)
* [Approval policy](https://docs.mithunai.com/team/approval-policy.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.