How to Choose an AI Receptionist: A Buyer's Checklist

Every vendor demos well, so the demo tells you nothing. Six things that actually differ — failure behavior, handoff, records, billing, outbound safety, setup — and the two-week test that settles it.

Published 2026-09-06

Every vendor in this category demos well. The voice is good across the board now, which means the demo tells you almost nothing about which one to buy. The differences that matter show up in four places: what happens when it is wrong, how a call reaches a person, what you can read afterward, and how the bill is calculated.

Here is a checklist you can take into a sales call.

1. Failure behavior

Ask directly: what does the system do when it does not know the answer?

There are two possible answers and only one of them is acceptable. Either it guesses, or it stops and says it will get you an answer. Vendors rarely volunteer which, so ask, then test it — call in and ask something deliberately outside the configured knowledge.

Also ask what happens when the speech or language provider is unavailable. "The call rings through to your number" is a good answer. Silence is not, and a dropped call is worse than a busy signal.

2. Handoff to a person

• Can a caller reach a human mid-conversation, and how do they trigger it?
• Does the person receiving the transfer get the conversation context, or start cold?
• Where do transfers go — a verified number you control, or a queue?
• What happens if nobody answers the transfer?

Test this yourself from an outside line before you sign anything. If it is awkward for you, it is worse for a customer.

3. The record

• Are transcripts available for every call, or only completed ones?
• Who controls the retention period — you or the vendor?
• Can you delete a specific call?
• Is recording separate from transcription, and can you enable one without the other?
• Does the greeting disclose recording, in wording you control?

Recording law varies by state and several require all-party consent. You need the controls to match your jurisdiction, not the vendor's default.

4. Billing

• Per second, per minute, or per call — and what is the rounding rule?
• Are ring time, hold time, and wrap-up time billable?
• What is the overage rate past the included bundle?
• Is provider usage passed through at cost, or marked up inside a bundle?
• Does the invoice itemize each meter separately?

Per-call pricing is the model to be most careful with: it charges the same for a 20-second wrong number as for a six-minute booking, and it gives the vendor an incentive to keep calls short rather than useful.

5. Outbound safety, if it can place calls

A system that can only answer calls has a small blast radius. One that can place them does not.

• Is there a destination allowlist you control?
• Are inbound call webhooks cryptographically signed and verified before being trusted?
• What stops a malformed instruction from producing a call you did not authorize?

Terri constrains outbound calls to destinations allowed by account policy and verifies Twilio's signature on every inbound call webhook before acting on it. Ask any vendor for their equivalent answers.

6. Setup reality

• How long until the first real call is answered — minutes, days, or a scheduled onboarding?
• Are instructions written in plain language, or configured as rigid fields and keyword tables?
• Can you change behavior yourself, or does it require the vendor?
• Do you get a dedicated number, or does it attach to your existing one?

The two-week test that settles it

Do not evaluate on a demo. Point after-hours calls at the system for two weeks and read every transcript. You are looking for:

• Answers that were wrong
• Answers that were vague because your instructions were vague
• Calls that should have transferred and did not
• Calls that transferred and should not have

That last list is the actual product review. It costs two weeks and tells you more than any comparison table, including this one.

Common questions

What is the most important question to ask an AI receptionist vendor?

What it does when it does not know the answer. There are two possibilities — it guesses, or it stops and says it will get an answer — and only the second is acceptable. Test it by calling in and asking something outside its configured knowledge.

Which pricing model should I avoid?

Per-call pricing. It charges the same for a 20-second wrong number as for a six-minute booking, and it gives the vendor an incentive to keep conversations short rather than useful. Per-second or metered usage aligns better with the caller being helped.

How should I test an AI receptionist before committing?

Point after-hours calls at it for two weeks and read every transcript, looking for wrong answers, vague answers caused by vague instructions, calls that should have transferred and did not, and calls that transferred unnecessarily.

What safety controls matter if the system can make outbound calls?

Require a destination allowlist you control, cryptographically signed and verified inbound call webhooks, and a clear answer on what prevents a malformed instruction from producing an unauthorized call. A system that can place calls has a much larger blast radius than one that only answers.