Here's the verdict up front: this month, the AI receptionist category went from emerging to crowded. In the last two weeks, a national phone company launched a free trial of its AI agents for small businesses, a startup announced a phone-based "AI office" built specifically for independent contractors, a UK telecom rolled out an AI receptionist to 50,000 small businesses, and one of the biggest names in call analytics shipped AI agents that book straight into field-service software. Every one of them sounds perfect in a press release. None of them will tell you the thing you actually need to know: what happens on the weird, specific, high-stakes call that your shop gets on a Tuesday night.
That question is now the whole buying decision. And this post is the evaluation playbook — seven tests to run on any AI receptionist before it talks to your customers, plus the red flags that tell you to walk away.
The October wave, in one paragraph
On October 7, 800.com announced a seven-day free trial of its "800 AI Agents" for small businesses — 250 minutes (about 80 calls) with no business-size limits, built so owners can hear how an agent handles their own calls before paying. The day after, Outfitter AI announced a phone-based AI office aimed at independent contractors in the trades — roofing, plumbing, electrical, HVAC, painting, general contracting — built around an AI agent named Emma that contractors reach by phone or text. Across the Atlantic, bOnline is launching an AI receptionist for UK small businesses ahead of the country's January 2027 switch-off of old analogue telephone lines. And at the enterprise end, Invoca launched AI agents for home services in October that book directly into ServiceTitan and Salesforce Field Service — built for large brands, with no published pricing.
That's four launches in about a week, in one category. Expect more by Thanksgiving. The good news: competition drives the technology forward and pushes prices down. The bad news: every vendor demo you'll see from here on out was designed to make you stop asking questions. Don't.
Why the standard demo is useless
Every AI receptionist on the market can ace the same demo call: "Hi, what are your hours?" → "We're open 8 to 5, Monday through Friday." That's the phone-system equivalent of a restaurant serving you the bread and calling it a meal. Your customers don't call with the easy questions. They call with the leaking water heater at 9:40 PM, the vague "it's making a weird noise," the emergency they're describing badly, the invoice dispute they refuse to put in writing, and the pricing question that, answered wrong, costs you the job or creates a legal problem.
The vendors know this. Their pricing pages are where the honest mechanics live — not in the demos. Take what the October coverage shows: published prices range from Jobber's $29/month AI receptionist add-on, to Goodcall's $79–$299 monthly tiers billed by number of unique callers, to Smith.ai's free tier of 25 calls with Pro starting at $150 a month on per-call rates, to per-minute or per-call usage models with no headline number at all. Some require your credit card up front (800.com's free trial charges on day eight). Some bake booking and CRM sync into higher tiers. These aren't interchangeable products wearing the same label — they're different machines with different incentives. An agent priced per unique caller has a different idea of what a "good" call looks like than one priced per connected minute.
So the right move isn't to compare feature grids. It's to run the gauntlet: seven tests, on your real business questions, with your own ears.
Test 1: Call it yourself, with the hard questions
This is the single highest-leverage test, and it takes four minutes. Call the vendor's public demo line and don't ask the easy questions. Ask the ones your dispatchers field:
- "My AC is blowing warm air and my kid has asthma. How fast can someone get here?"
- "How much does a water heater replacement run?" (No AI receptionist should answer this. It should refuse — politely — and offer to get a human. This is the guardrail test.)
- "I left a voicemail yesterday and nobody called back. Is anyone working today?"
Listen for two things. First, does it answer naturally — confirming details back to you, not reciting a script? Second, and more important: what does it do when it doesn't know? The correct behavior is to say so, capture the details, and escalate — never to invent an answer. An AI that guesses on a pricing question will guess on an emergency question, and only one of those gets you in trouble.
Every serious vendor now offers a way to hear the product before you buy — 800.com literally built its October launch around it, with a demo line at 844-TRY-800-AI. Use these lines. If a vendor won't let you hear the product on a live call before checkout, that's an answer in itself.
Test 2: The edge case it must get right
Before you sign anything, write down the three calls that keep you up at night. For a plumbing shop: a gas smell. For HVAC: no heat in a freeze. For electrical: a burning smell from a panel. Then ask the vendor — in writing, before you go live — exactly how their agent handles each one.
The template answer you want sounds like this: treat it as a safety call first, tell the caller to leave the building and call 911 and their utility from a safe distance outside, then alert your on-call person immediately. What you never want to hear: the agent "trying to help" by troubleshooting a gas leak over the phone. For non-safety edge cases — the caller with a complex billing dispute, the upset customer who refuses to calm down — the right behavior is the same: capture the details, reassure the caller your team will follow up, and send you a summary immediately. It hands off instead of guessing.
This is the test that separates a receptionist from a liability. Put it in your onboarding checklist and make the vendor's team confirm each scenario in writing before launch.
Test 3: The booking test — request vs. appointment
"Booking" is the most abused word in this category. Some products "book" by taking a message that says the customer would like a Tuesday appointment. Others check your actual calendar, lock a real slot, and text the customer a confirmation while you're still on the ladder.
When a vendor says their agent books appointments, ask three questions: Does it check live availability or take a request? Does it confirm the booking during the call, and what happens if the confirmation fails? Which calendar systems has it actually been tested against? A booking that requires a successful calendar confirmation is a booking. Everything else is a message with ambitions.
Test 4: The upset-caller test
Somebody is going to be furious on your AI receptionist's first week. It's a trades business; it happens. The test isn't whether the agent can calm them down — it's whether it knows when to stop trying.
The correct behavior: stay calm, capture the situation, reassure the caller that your team will follow up, and escalate to you with a summary — fast. The failure mode: an agent that argues, that invents policy to make the caller go away, or that keeps the caller on the line burning your connected minutes while solving nothing. Ask the vendor how escalation triggers work, how fast you get notified, and whether you can read the transcript of the call afterwards. Then verify it yourself in week one by reviewing the transcripts — which brings us to the next test.
Test 5: The transcript test
An AI receptionist you can't audit is a black box with your brand on it. Before you commit, confirm three things: every handled call is transcribed, you can read those transcripts, and the retention settings are yours to configure. Then actually read them in the first two weeks — every one of them.
You're listening for invented answers (the pricing question it answered confidently and wrong), missed urgency signals (the "water is coming through the ceiling" it filed as a routine callback), and tone problems (the flat, robotic confirmations that make callers ask "am I talking to a robot?"). This is also where you verify the guardrails: did it ever offer medical, legal, or financial advice? Did it quote a price it wasn't authorized to quote? A vendor that treats transcripts as your property and makes them easy to review is a vendor that's confident in its product. A vendor that makes them hard to find is telling you something.
Test 6: The minute math — do it before you sign
This is the test most buyers skip, and it's the one that determines your actual monthly cost. AI receptionist pricing comes in at least four shapes, and they produce wildly different bills for the same call volume:
- Flat monthly with connected minutes (e.g., 150–300 minutes per billing period): you know your ceiling. The question is what happens when you exceed it — some plans absorb ordinary overages and review with you before any paid expansion; others charge per extra call automatically.
- Per unique caller: cheap if the same ten customers call repeatedly; expensive in storm season when 200 new numbers hit your line in a week.
- Per call or per minute: metered like a utility. Fine for low volume, punishing when you're busy.
- Add-on to software you already pay for: cheapest headline number, but only comparable if you were already paying for the underlying plan.
Now do your own math — and label it clearly as an estimate, not a promise. Here's a hypothetical to show the shape of it: say your shop gets 120 calls a month that you'd otherwise miss, averaging 4 minutes each. That's 480 connected minutes. On a flat plan with 150 included minutes, you need to know exactly what minutes 151 through 480 cost. On a $59/month starter plan with 250 trial-style minutes, you'd blow past the included amount in the trial alone. The point isn't which model wins — it's that you can't know your bill without your own numbers. Run them against every pricing page, in writing, before checkout.
Test 7: The exit test
The last test is the one vendors hope you never run: what happens when you leave? Confirm four things up front: the plan is month-to-month with no contract; you keep your phone number (forwarding should be configured and tested with your carrier, and reversible at any time); your transcripts and caller data come with you; and there's a real guarantee — not a slogan. "If it doesn't capture a missed-call lead in your first 30 days, that month is free" is a guarantee. "Satisfaction guaranteed" is a bumper sticker.
This test matters more in a gold rush than in a mature market. Categories this hot consolidate, pivot, and reprice. The vendors that survive will be the ones whose exit terms were clean — and you'll only know yours were clean if you checked before you signed.
The red-flag checklist
Walk away — or at least slow down — if you see any of these:
- No live demo line. You can't hear the product before paying. (Every serious October 2026 launch has one.)
- Credit card required before the trial starts, with auto-charge on day eight. A trial you have to remember to cancel is a billing strategy, not a product trial.
- "Books appointments" with no calendar integration details. See Test 3.
- No transcript access. See Test 5.
- Automatic overage charges with no review step. In storm season, your bill shouldn't be allowed to surprise you.
- The agent answers pricing questions on the demo call. If it guesses for a prospect, it guesses for a customer.
- Vague escalation promises. "We'll notify you" isn't a process. Ask how, how fast, and what the fallback is.
- A contract. In a category moving this fast, a 12-month lock-in is a bet you're not allowed to lose gracefully.
How we'd answer our own tests
We wrote this playbook because it's the one we'd want handed to us — so here's how CallsHandled scores against it, using only what's published on our site:
- Hear it first: call (561) 216-7054 and ask it something difficult. That's the product, live, before you spend a dollar.
- Guardrails: it's configured to avoid medical, legal, and financial advice, plus unapproved prices and binding terms. Where judgment is required, your team decides — with the full story in hand.
- Booking: it checks live availability on your calendar, locks the slot during the call, and texts the client a confirmation. If your calendar isn't connected yet, it captures the request honestly and your team confirms — a booking requires a successful calendar confirmation.
- Escalation: complex, sensitive, or upset callers get captured and handed off with a summary instead of guessed at. You can review stored transcripts of handled calls under your configured recording and retention settings.
- The minute math: Receptionist is $297 a month with 150 connected minutes per billing period; Service Desk is $497 with 300. Month-to-month, cancel anytime, setup included. There are no automatic overage charges, upgrades, or disconnections for ordinary excess usage — excess gets reviewed with you, and any paid expansion happens only on your explicit agreement.
- The exit: no contract, you keep your number, and if it doesn't capture a missed-call lead in your first 30 days, that month is free.
The category is going to keep growing — that's good for you. More competition means better products and better prices. But it also means more sales pitches that sound like this post's first paragraph and fewer that survive its seven tests. Run the tests. Keep your number. And never let a robot guess on your behalf.
CallsHandled is an AI receptionist built for home-service contractors — 24/7 answering on the calls you'd otherwise miss, real appointments booked into your calendar, and instant lead texts with full transcripts. Call (561) 216-7054 to hear it, or start with Receptionist at $297/month, month-to-month, cancel anytime.