What to ask before you buy an AI receptionist

Most AI receptionist demos are recorded under perfect conditions: a quiet room, a cooperative caller, a question the script was built for. Your callers are in a truck with the window down, they interrupt, and they ask things nobody anticipated. These are the questions that separate a system that holds up from one that only demos well.

How fast does it actually answer and respond?

Two different numbers, and vendors like to blur them. Pickup speed is how long the phone rings. Response latency is the pause after the caller stops talking before the AI starts. Anything past about a second and a half reads as a dropped call, and people hang up on dropped calls.

Ask for a live number you can call yourself, from your own phone, at a time you choose. Not a recording. If a vendor will not hand you a number, you are being sold a prototype.

What happens when the caller goes off-script?

This is the real test. Every system handles “what are your hours.” The question is what happens when someone says their basement is flooding and they need somebody tonight, or asks a pricing question with six conditions attached.

• Does it say something confidently wrong, or does it admit the limit and route the call?
• Who does it route to, and how fast: a human, a text to the owner, a callback promise with a real time attached?
• Does the caller ever hear the phrase “I didn't understand that” more than once in a row?

A system that says “let me get someone who can answer that properly” beats a system that guesses. Confidently wrong is worse than honestly limited.

Can it write to your calendar, or only read it?

Read-only means the AI tells the caller you are free Thursday, then a human still has to key in the appointment. That is a message-taking service with better manners. Write access means the booking lands on the calendar during the call, the caller gets a confirmation, and nobody touches it again.

Ask specifically: which calendar system, does it handle double-booking, and what happens if two callers try for the same slot in the same minute.

Where does the call data go?

Every call is a lead record: name, number, what they wanted, whether they booked. If that data lives only inside the vendor's dashboard, you are renting your own customer list. It should flow into a CRM you own, so you can follow up, retarget, and see the pipeline without asking permission.

Who trains it, and who fixes it at 9pm on a Saturday?

The cheapest tools hand you a prompt box and wish you luck. That is fine if you enjoy prompt engineering. If you do not, you want the build, the training, and the ongoing tuning done for you, including the part where you listen back to a bad call and someone corrects the behavior before Monday.

The pricing question underneath the pricing question

Per-minute pricing looks cheap until a busy month. Flat monthly pricing looks expensive until you compare it to one missed job. Ask which model you are on, what counts as a billable minute, and what the bill looked like for a comparable business in a busy season.

The short version

• Call the demo line yourself, from your own phone, unannounced.
• Ask it something deliberately awkward and listen to how it fails.
• Confirm calendar write access, not just calendar visibility.
• Confirm the lead data lands somewhere you own.
• Confirm who tunes it after go-live, and how fast.

Then run your numbers through the calculator to see whether the math actually works for your business.