Every one of these products demos well. The demo runs the happy path, the voice sounds good, and the voice sounds good on all of them now.
What separates them is a shorter list than the feature comparison suggests, and most of it sits in areas that don’t demo well at all.
The four that actually matter
It writes to your calendar. Not “integrates with scheduling”. Creates a real appointment, at a real available time, with the job details attached, while the caller is still on the phone.
This is the single biggest fork in the market. A system that books is a receptionist. A system that takes a message is an answering service with better diction, and the two are frequently sold at similar prices. Ask to see a booking appear in a live calendar rather than a screenshot.
Its answers come from your material. When it says you service a suburb or charge a certain rate, that should be retrieved from something you supplied, not composed on the spot.
Ask directly: if it tells a caller something wrong, can you trace which document produced it? A vendor who can answer that has built retrieval. One who talks about how advanced the model is has not, and the failure mode is a confidently invented price that nobody hears about until the job is booked.
It hands over cleanly. Every system meets a caller it can’t help. What happens next is the thing you’ll live with.
The good version transfers to a person with a short summary of what the caller wants, or takes a genuinely detailed message and creates a task. The bad version loops, which is worse than voicemail because the caller has already spent two minutes getting nowhere.
You can see what it said. Transcripts of every call, searchable, retained somewhere you can reach. This is how you discover it’s been quoting last year’s rates for a month, which is the most common way these setups quietly decay.
The rules you should be able to set
Less discussed and worth insisting on. You need to be able to state what the system must never do without a person.
Common ones: never quote a price above a threshold, never commit to a same-day callout, never handle a complaint, never promise a date the calendar can’t support. If those can’t be configured, you’re relying on the model’s judgement for decisions that carry real cost.
This is a business decision rather than a technical one, and it belongs with the rest of your AI guardrails. Write the list before you shop, then check each product against it.
Things that sound important and aren’t
Voice quality. Uniformly good. Choosing on this is choosing on the one axis where everything is equivalent.
Number of supported languages. Impressive in a deck, irrelevant unless your customers actually call in those languages.
A human name and personality. You gain a little warmth and lose the ability to say plainly that this is an automated system that will pass you to a person. Not a good trade.
Sentiment analysis dashboards. Attractive reporting on a metric almost nobody acts on. Transcripts are more useful and less impressive.
Questions that reveal the most
“Can I hear a call that went badly?” Everyone has a highlight reel. A vendor who’ll play you a failure has been doing this long enough to have them and is honest enough to admit it. That’s a strong signal in both directions.
“What does a typical call cost me?” Then multiply by your real volume, including nuisance calls, which get answered and billed like any other. Separate that from the one-off cost of building it, which is a different question with a different answer.
“Who updates it when my prices change?” If the answer is vague, it’s you, and it will be forgotten. Settle it now.
“What happens at 2am?” Out-of-hours behaviour is usually configured separately and tested by nobody.
“Whose account is this built in?” Underasked, occasionally expensive, and the same question worth putting to any automation provider. If everything lives in the vendor’s account with credentials you’ve never seen, leaving is harder than arriving.
Privacy, briefly but properly
Call recordings are personal information, and phone calls have their own rules on top of that.
Three things to establish before you sign: whether callers are told the call is recorded and how, where the recordings are stored, and whether anything you send is retained for training. All three have defensible answers, but only if somebody picks them deliberately rather than accepting a default.
Worth twenty minutes now rather than an awkward conversation later.
How to actually test it
Ring it yourself, and don’t use the script.
Ask the ten things your customers really ask, in the words they really use. Call from a car with the window down. Give it something ambiguous. Give it something it can’t possibly handle and watch the handover. Book an appointment end to end and check what lands in your diary, with what detail, and whether the customer gets a confirmation.
Half an hour of that will tell you more than every comparison table on the internet.
For what these systems can and can’t do generally, the honest overview is here, and if you haven’t yet decided whether one is justified at all, work the numbers first.
If you’d like a second opinion on a specific product or quote, send it over and I’ll tell you which of the four capabilities above it actually has. It’s usually fewer than the page suggests.
Frequently asked questions
What features should an AI receptionist have?
Real calendar booking rather than message-taking, answers grounded in your own documents, a clean handover to a person, configurable rules about what it must never handle, and a transcript of every call. Natural-sounding speech is no longer a differentiator because almost everything on the market has it.
What's the difference between an AI receptionist and an answering service?
An answering service takes a message and passes it on. An AI receptionist should complete the task while the caller is still on the line, which usually means booking into your calendar. If a product can't write to your diary, it's an answering service with better speech, and it should be priced like one.
Should an AI receptionist integrate with my CRM?
It's worth it if you want callers recognised rather than treated as strangers every time, or if enquiries need to land in a pipeline automatically. If you only need bookings, calendar access may be enough. Decide which you need before shopping, because integration depth is where quotes vary most.
What questions should I ask an AI receptionist vendor?
Where do its answers come from, what happens when it can't handle a call, what exactly is it connected to, what does a typical call cost, who updates it when my prices change, and can I hear a recording of a call that went badly. That last one tells you more than any demo.
Do I get recordings and transcripts of calls?
You should, and check this explicitly. Transcripts are how you find out what the system is telling people, which is the only way to catch it answering confidently from out-of-date information. Also check how long they're retained and where, since call recordings carry privacy obligations.
Wondering what this would look like in your business? A short chat is usually enough to tell.
Let’s chat