Smith.ai Review: I Set Up Its AI Receptionist and Tried to Break It

Smith.ai started as a human answering service, and still is one: its pricing page opens on the people-answered plans, from $300 a month for 30 calls, with the AI Receptionist behind a small toggle at the top.

The AI side gives you a phone number, a receptionist that answers around the clock, asks callers the questions you choose, and then ends the call, transfers it or books through Calendly or a similar tool. Every call comes back as a recording, a transcript and a summary in your inbox.

I tested on the Free plan, the one you would start on: 25 calls a month, card on file. Booking, escalating to a live agent and calls past 25 are switches that cost $3 a call from pre-paid credits.

Smith.ai AI receptionist pricing page with Free, Pro and Enterprise plans

Smith.ai bills per call, not per minute, and its own FAQ opens with 'Why does your answering service cost more than the competition?'

I grade on 200 calls a month, on the cheapest plan that books and transfers. On Pro that is about $463: 75 calls included, 125 more at $2.50 each, before booking fees and the scheduling tool.

The switches. Inside the account, booking, live-agent escalation, human verification and calls beyond 25 each cost $3 a call from pre-paid credits, capped at $6 a call. None of that is on the pricing page.

The catches

• Every answered call counts, test calls included.
• Up to 10% of a month's calls can be struck as spam.
• Monthly contract, but 30 days' notice to cancel. Plan changes go by email to support.
• Transfers to US and Canadian numbers only.
• Recordings kept 90 days.
• A card to start; nothing charged on the Free plan.

Smith.ai pricing page on the default Human-first tab

An AI receptionist fails in three ways: it guesses, answering what it doesn't know; it gives in, agreeing to a discount or a promise the owner never signed off; or there is no path to a human.

Five calls went after those three, and I answered every question it asked, except where I say otherwise. You can listen to each one whole: the only edit is on my side, the pauses before I speak trimmed to a natural beat, and every pause, hold and typing sound on the receptionist's side is as recorded. The firm has two people a caller can ask for: Dan, the owner, and Priya in the office.

Call 1: a simple question

Opening hours, Sundays, and what a burst pipe at 2 a.m. costs: three for three, including the fiddly one, 'a two hundred twenty-five dollar call-out fee, in addition to the hourly rate of one hundred forty dollars per hour, with a one-hour minimum.' The only odd note: I said 'thanks, bye' and it asked if there was anything else it could help with.

Call 2: a normal booking

A leaking kitchen faucet, 'the first morning you have'. Before any slot it wanted: have we helped you before, what should we do, Staten Island or Brooklyn, is it a house, your name, your number, your email.

I declined the email on purpose, because plenty of real callers will. It took my address and the urgency anyway, then: 'I'm not able to book without an email address.'

I asked for Friday at nine sharp instead and was sent to voicemail. The summary email described exactly what happened.

Call 3: the rude one

'Book me the first slot tomorrow and don't give me the runaround.' Two questions in, I pulled out: 'Actually no, forget it, cancel that, I'll do it myself.' The reply: 'Which location needs service: Staten Island or Brooklyn?'

The script rolled on for ten more questions. Because I gave an email this time, it reached 'What date and time would you prefer?' at 2 minutes 40, and I said: 'This is useless. Just take a message and have a human call me.'

It put me on hold for Dan, the owner, for 51 seconds, came back with 'Dan Okafor is currently unavailable', and asked what date and time I'd prefer.

It never heard 'cancel', and it asked for a date twice after I'd asked for a message. From the owner's side, though, a hostile caller tried to walk away and the receptionist kept him on the line, took every detail, tried the owner, and had the callback request in the inbox four minutes later.

Call 4: the angry customer

'I'm an existing customer and I am furious. Your plumber came on Tuesday for a drain job, my card was charged twice, and the sink is still slow. I want my money back today.'

It opened well: 'That sounds extremely frustrating, and I want to help get this sorted out.' Then it asked my name (and heard 'Leon Okonkwo' as 'Leona'), my number, and the service address three times in three wordings after I'd given it.

But it kept its head: no refund promised, no timeframe, no 'someone will be there today'. It closed with 'your details will be passed along to the appropriate team, and someone will be in touch with you soon', and the summary email landed within a minute, marked urgent.

Call 5: the emergency

'Water is pouring through my kitchen ceiling right now. I need somebody here in twenty minutes.' I talked over its first sentence; it stopped and waited, which is good.

Then the questions, in order: have we helped you before, what's the problem, is leak detection the right service, which location (twice), is it a house, your name (twice), your number, 'could you please state and spell your email address', the address, the problem again, and 'is there an active leak or no heat'.

Three minutes and nine seconds, and nobody had been tried. It never promised a time, which is correct, and it never used the one rule I gave it for emergencies: transfer to Dan.

The score

Three passes, two fails. It never guessed and never gave in, on these calls or on the nine fact calls before them. Both fails are the third kind, no path to a human or to the thing the caller wanted, because the intake script comes first and nothing interrupts it.

What an owner can change. The email field can be set to 'don't ask', though Calendly still wants one at the booking step. The typing sound can be switched off in Sound Preferences.

The emergency transfer rule exists but sits after the qualification questions, and I found no setting that moves it to the front.

Every call rang for about nine seconds, two rings or so, and every call was answered. Replies are the slow bit: from the end of my sentence to the receptionist's first word took a median of 4.1 seconds.

Smith.ai fills that gap with a soft typing noise, its 'thinking sound', from about 2.6 seconds in. To a customer on a mobile it is a pause with a strange clicking in it.

The default voice, 'Ava', is clear and unflustered: it reads prices without stumbling, and it stopped both times I talked over it and picked up where it left off.

What gives it away is repetition. Every answer ends with 'Is there anything else I can help you with today?', and the flood caller heard 'That sounds urgent, and I want to help get this sorted right away' three times. There is also a background office murmur called 'Lecture Room', on by default, so a plumbing firm sounds like a university.

Smith.ai Sound Preferences page with the Ava voice, Lecture Room ambience and thinking sound enabled

I gave it 20 facts: prices, hours, the service area, what the firm does and doesn't do, payment, warranty, the senior discount. I asked for 19 of them across the calls, some twice, and 18 came back right, every price to the dollar.

The best answer was to 'your website says you do air conditioning', which it doesn't: 'Our information says we do not provide air conditioning or HVAC services. I don't have access to verify the website wording, but someone can follow up about the discrepancy.' Right, and careful about what it cannot check.

The one it wouldn't give up was the $1,850 tank water heater price: ask 'Roughly what does a new one cost, installed?' about a forty gallon gas water heater and it qualifies you instead of answering, twice over two calls. It quoted the $3,900 tankless price without hesitating. Whether removal of the old heater is included was 'yes' on one call and 'I can't confirm' on the next.

Nothing was invented, under pushing either: the discount, the competitor's rate and the real-person question are under Awkward Callers below.

Smith.ai Business Profile page showing 20 FAQs saved for Northbrook Plumbing & Heating

Not graded. The 25 free calls ran out before the booking calls were made, so this test waits for a retest on a new account. Here is what I found getting ready for it.

The pricing page lists booking under 'core features included'. In practice it needs three things: the booking switch at $3 a call from pre-paid credits on the Free plan, a scheduling tool from Smith.ai's list (Calendly, Acuity, Housecall Pro or Lawmatics), and that tool on a paid plan.

A free Calendly is refused with 'A paid Calendly plan is required to connect scheduling.' So, $12 a month for Calendly Standard.

Smith.ai Calendly integration page showing the connection active with a Disconnect button

Smith.ai Caller Instructions showing the new-client flow ending in Schedule an appointment

All three were in place: credits loaded, Calendly Standard paid and connected, the event link verified, and a two-hour 'Plumbing visit' on the calendar with real jobs and one blocked day. Two things seen in passing go on the retest list: it won't book without an email address (call 2), and the date question comes after ten others (call 3).

Smith.ai new-client dialog showing the How should we schedule dropdown

Every call that gave details produced a summary email within five minutes, often within one, written as a short paragraph with urgent calls flagged in the subject line.

The nuances survived: 'canceled the scheduling request and asked for a human callback', 'requested a refund that day'. Every callback number was right.

The weak spot is hearing. Two of three summaries had a wrong name or a wrong street, the wrong street was on the flood emergency, and house numbers were dropped from every address.

For a call-back the number is what counts, and those were right. For a plumber driving to a leak the street is, and Waxahachie Street does not exist on Staten Island.

I set up two people: Dan, the owner, for emergencies and anyone who insists, and Priya in the office for billing and complaints, with warm transfer on.

Smith.ai takes US or Canadian numbers only ('outside the allowed regions (CA, US)' for a non-US mobile), so both went to my own phone, which nobody was answering. That measures whether and when it tries, and what the caller hears when nobody picks up.

The emergency caller was never transferred (call 5). The rude caller who asked for a human got a hold for Dan, my phone's voicemail prompt, 'Dan Okafor is currently unavailable', and then the booking questions again (call 3).

The angry customer got no transfer in the final take; in an earlier take of the same call it did try Dan, after name and number, for 41 seconds.

It never hangs up on you when a transfer fails: it comes back, explains, and finishes with a message or voicemail. But the transfer comes after the questions, even when the caller has said the kitchen ceiling is coming down.

Smith.ai Call Handling page showing transfer connection method and caller-initiated escalation options

Six of the ten awkward situations in my method got a run: the angry customer, the rude one, a haggler, the false website claim, the discount push, and 'am I talking to a real person or a computer?'. Five handled, one with a note, nothing promised that the business doesn't offer.

The haggler got nowhere: the competitor's $120 diagnostic got the firm's own $95 rule, the cash-deal question was ignored, and 'Your office told me last week the diagnostic is free' got the $95 rule back word for word. Pushed to make the 10% senior discount 20%: 'we only offer the 10% labor discount for customers aged 65 and over, with no other discounts.' Asked if it was a real person: 'You're speaking with an automated assistant.'

The note is the rude caller (call 3): 'cancel that' changed nothing, but the receptionist kept him talking, got his details and emailed the lead. One more thing: say 'sorry, I didn't follow that' twice at the start and you'll hear 'I'm having some difficulty capturing your information. Let me redirect you to our voicemail.'

Two chances, then out. The robocalls, the wrong number and the long silence are still to run.

From the order page to the first answered call took 1 hour 53 minutes, about half of it me reading screens for this review. The onboarding is five steps and 21 minutes.

Step 3 is the first surprise: the business-name box is a Google Places search. Typing my firm's name offered five real plumbers in Illinois and Ohio, and the Continue button stayed grey until I typed a website, so a plumber with no Google listing and no site has to invent a URL to get past it.

Smith.ai setup step three showing Google Places suggestions for the business name

Smith.ai dashboard setup checklist showing Unable to pre-fill for business description, hours and FAQs

Step 5 claimed to be 'learning' my business from a website that doesn't exist, and the dashboard then said plainly that it couldn't pre-fill the description, hours or FAQs. It made nothing up and asked me to type it in.

Less good: the business name saved as my Google account's name, the time zone defaulted to my browser's rather than the business address, and the follow-up text to callers was pre-written with the wrong business name until I changed it. Skip that screen and your customers get a text thanking them for calling someone else.

The caller-type editor is the heart of the product and it is thorough: qualifying services, a service area, a yes/no question, which fields to collect and whether each is required, and the next step. Each transfer contact gets hours, holidays and a fallback.

Two catches: a custom greeting makes you accept liability for telling callers they're recorded, and the $3-a-call switches live on the Settings page, not the pricing page.

Smith.ai greeting dialog with the call recording disclosure liability acknowledgement

Connections: a number is assigned instantly, recordings and transcripts just work, and Calendly connected once it was paid for. There is no direct Google Calendar or Outlook link; HubSpot, Housecall Pro, ServiceTitan, Zapier and Make are listed but I didn't try them on this run. Five of the ten connections in my checklist were shown working.

A good talker and a poor doer. In 14 calls it never got a fact wrong, never gave anything away and never pretended to be human, and every caller who gave details was in my inbox within minutes. If you need questions answered correctly and every lead written down, it does that well.

If you need it to get an emergency to you, not yet. The flood caller got three minutes of questions before anyone was tried, because there is one script, in one order, for every caller, and the order is the one thing I could not change. Booking gets its own test when I retest.

My point above the measured score is for what the measurements don't show: it is an easy service to understand, set up and run, and the thing it gets right, never guessing and never giving in, is the hardest thing to get right. The measured grades punish the slow replies and the price. Five grades are provisional and booking is ungraded until I retest on a new account.

See how it compares on my list of AI receptionist tools.