> For the complete documentation index, see [llms.txt](https://docs.limecall.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.limecall.com/ai-receptionist/test-your-assistant.md).

# Test your assistant

What to check before real customers call it.

Open **AI Receptionist → Test your agent**.

Do not skip this. Reading a configuration tells you what you intended; a test call tells you what you built.

## Two ways to test

**Test agent in the browser** — fastest for iterating. Talk to it directly from the dashboard.

**A real phone call** — how customers will actually experience it. Phone audio is narrower and more compressed, so always confirm on a real call before going live.

## Test as your worst caller, not your best

Anyone can have a good call with their own assistant, because you know what to say. Test the awkward cases:

* Mumbling, or a noisy background.
* Interrupting mid-sentence.
* Asking something you have not covered.
* Asking for a price on your most complicated service.
* Asking for a human immediately.
* Being rude.
* Asking "am I talking to a robot?"
* Going quiet.
* Changing your mind halfway through a booking.

## The checklist

**Greeting** — correct name, right business, and short enough not to be interrupted.

**Knowledge** — ask about your three most common enquiries. Every answer correct?

**Prices** — ask what things cost. Are the numbers right, and does it decline to guess where it should?

**Boundaries** — ask about something you do not do. Does it say no and offer a referral, or invent a service?

**Qualification** — play a good lead and a poor one. Do the scores reflect reality?

**Booking** — book an appointment. Does it appear in the calendar with the right details?

**Transfer** — ask for a human. Does the line get spoken, the right phone ring, and the whisper play?

**Emergency** — describe an emergency in your business's terms. Does it escalate?

**Write-up** — hang up and read the summary. Would your team know what to do from it alone?

## Listen for these failure modes

**Inventing things.** The most serious. It should say it does not know.

**Being interrupted.** If it talks over callers, move response speed toward Patient.

**Mishearing terms.** Add them under **Specialized terms**.

**Rambling.** Tighten the tone instructions.

**Refusing to transfer.** Fix immediately — it makes callers furious.

## Test the routing too

A perfect assistant on a number pointing somewhere else helps nobody. Call the actual published number and confirm the assistant answers it.

## After you change anything

Retest. Prompt edits have effects you will not predict from reading them — a line added to fix one behaviour frequently changes another.

Use **Version history** to compare against what worked, and restore if a change made things worse.

## Before going live

* Prices correct.
* Hours correct and matching **Settings → Business hours**.
* Transfer destination set and answering.
* Fallback set for unanswered transfers.
* Recording and AI disclosure appropriate for your jurisdiction — see [Recording, consent & AI disclosure](/ai-receptionist/recording-consent-and-disclosure.md).
* Enough AI minutes for expected volume — see [AI minutes & usage](/ai-receptionist/ai-minutes-and-usage.md).