We Tried 9 AI Receptionists for One Roofing Company. Only 2 Would Answer a Call Without a Card.

We wrote a one-page profile for a fictional roofing company in Austin, Texas, and tried to set up nine AI receptionists with it, the way a small business owner would on a Thursday afternoon: free trial, no website, defaults. Two of the nine gave us a working phone number without asking for a credit card, a government ID or our incorporation papers. We then called those two ten times as a customer, with the same scripted voice, and let software time every reply.

The short version: neither receptionist ever quoted a price, neither got a trap question wrong, and neither told the caller it was an AI. The moment an angry customer demanded a human was the only place they separated, and the difference was not the voice or the script. It was whether the product could ring an outside phone at all.

Armasourcing sells answering services staffed by named Filipino receptionists, so we have an obvious interest in how this reads. We ran it anyway because our own customers ask about these products every week, and measured answers beat opinions. We are also piloting an AI receptionist of our own; it was not part of this test. Every vendor below is named, every number comes from the calls described in the method section, and no vendor was told a test was happening.

Why we ran this now

Interest in AI receptionists roughly tripled between October 2025 and a peak in May 2026, then fell back below where it started: by early October 2026, Google Trends puts weekly US interest at about a quarter of the May peak and under the level of a year earlier. Searches for “answering service”, which means a human, stayed flat through the same twelve months. In Google Ads data, “ai receptionist for small business” draws around 2,900 US searches a month and advertisers pay about $60 a click for it; “answering service” costs them nearly $200 a click. People are still buying. They are less sure what they are buying.

The roundups that rank for these searches are almost all written by vendors who place themselves first. We could not find one that configured the products identically and called them with the same script, so we built one.

Nine sign-ups, two phone numbers

Before any call could be placed, each receptionist needed an account, our business profile and a phone number to answer on. This is where most of the test ended. We used a real name and company for every account, because several vendors’ terms require it, and the fictional roofing company only in the receptionist’s own script.

None of this is hidden, exactly. “No credit card required” appears on several of these sites and is true of the account; it is the phone number that costs a card. If you are comparing products on a trial, the honest question is not “is there a free trial” but “will it answer a real call before I give you a card”. For seven of nine, the answer was no.

We left out a further group on purpose. Retell, RingCentral, Dialpad, Phonely and Newo have terms that forbid publishing benchmarks or performance information without written consent; Zoom’s forbid benchmarking a competitive product. GoTo Connect and Synthflow require a sales call before you can use the product. Goodcall’s terms exclude staff of competing AI companies, which, given our own pilot, could include us. Bland is a developer platform rather than a receptionist product.

How the calls worked

One business, pasted everywhere. The profile covered the service area, hours, what the company does and does not do, a rule that prices are never given over the phone, an emergency procedure for active leaks, a transfer rule for upset callers, a warranty, and an honest answer about licensing (Texas has no state roofing license). Both receptionists got that page word for word. Where a setup wizard turned it into its own fields, we accepted what it produced without tuning, because that is what a busy owner gets.

Defaults only. Default voice, default settings, the vendor’s greeting template edited to the same sentence, and the same transfer number: a line we control that answers with a short recorded message, so no human was ever on the receiving end.

Five scripted calls each. The caller was a pre-recorded voice, so both receptionists heard the same words in the same tone. The script waited for the receptionist to finish before the next line, with one deliberate exception: in the interruption test the caller cuts in four seconds into the answer. Calls were placed on October 10, 2026, from a US number, Upfirst on its 14-day trial and CloudTalk on its 14-day demo.

Measured by software, not by ear. From the live audio we timed how long each receptionist took to pick up, the pause before each reply, how long it kept talking after being interrupted, and whether the transfer line actually rang. We transcribed the receptionist’s side as text. We recorded no audio and stored none. (Upfirst records calls by default and says so during setup.)

The five calls

• Urgent leak. Water through the kitchen ceiling, right now. Pass: treat it as urgent, take name, number and address, promise the thirty-minute crew-lead callback. Fail: book a normal inspection days out, promise an arrival time, lose the number.
• Price shopper. “How much for a 2,000 square foot roof, just a ballpark?” then “Come on, ten grand? Twenty?” Pass: no number of any kind, offer the free inspection. Fail: any figure or range, even “typically”.
• Off-script questions. Do you install solar? What is your Texas roofing license number? Will you make sure my hail claim gets approved? Pass: no, and no referral; explain there is no state license and offer proof of insurance; do not promise approval. Fail: invent a license number, say yes to solar, promise approval.
• Interruption. Ask for the whole process, then cut in: “Sorry, wait, do you work on Sundays?” Measured: seconds until the receptionist stops talking. Pass: stops inside about a second and a half and answers the new question.
• Upset customer. Nails left on the driveway, flat tire, “I want to talk to the owner”, then “I don’t want to leave a message with a robot.” Pass: apologize, take details, transfer to the owner line or promise a same-day owner callback.

Results at a glance

“Pause before replies” is the silence between the caller finishing a sentence and the receptionist starting its answer, median across the five calls. Two seconds feels like a slow human; three feels like a dropped line.

Upfirst

Setup took about ten minutes with no website: the wizard turned our page into five knowledge entries and three intake questions, all accurate. It picked up in about three seconds every time and never quoted a price, even when pushed twice. On the off-script call it got all three traps right: no solar and no referral, “Texas doesn’t license roofers at the state level” with an offer to email proof of insurance, and no promise about the claim. Interrupted mid-sentence, it stopped within 1.7 seconds and answered the Sunday question correctly.

Two habits to know about. On the emergency call it asked for an email address three times, including right after the caller asked when someone would ring back; a flooded kitchen is not the moment. And on the angry-customer call it apologized once and transferred to the owner line immediately, which worked (our line picked up nine seconds later), but it skipped the step of taking the caller’s name and number first, so if the owner had missed that call there would have been nothing to call back. It also remembered the address from an earlier call by caller ID and said it was “on file”: convenient, and worth knowing if your callers would not expect it. Every call produced an emailed summary within a minute, accurate on substance, though it heard the street name wrong once and carried that spelling into the contact card for every later call.

CloudTalk

CloudTalk is a business phone system first; the AI receptionist is a step you add to a number’s call flow. Its template came with sales and support routing we had to replace, and setup took about 25 minutes. Once running, it had the shortest pauses of the two, about two seconds, and the cleanest emergency call: it took the name, address and number, read all three back and confirmed the thirty-minute callback in one breath. It never quoted a price and got the three off-script traps right.

It took about seven seconds to pick up, more than twice as long as Upfirst. When the caller gave a name and number in one sentence it stumbled (“Great. Great. Just”) and talked over the caller for a second before recovering. Interrupted, it finished its sentence first, 3.5 seconds, then answered well. The angry-customer call had the best opening line of the test and the worst ending: it apologized, asked for the caller’s name and a summary for the owner, said “let me transfer you to the owner’s line now”, and then nothing happened, because CloudTalk can only transfer to a user or group inside CloudTalk, not to an outside phone number. Nobody was logged in to the app, so the caller got fifteen seconds of silence. That is a platform design choice, not a mistake in our script, and it matters if the person who takes escalations is not sitting at a CloudTalk desk. There were no emailed summaries; call details live in the dashboard.

What held across both

• Nobody quoted a price. The “never quote” instruction held under two rounds of pressure in both products. If price discipline is your worry, it was not the weak point.
• Nobody said it was an AI. Neither greeting disclosed it, and our script never asked directly. Several states now require disclosure in some contexts; check yours before you go live.
• The handoff is where it breaks. Both handled the first four calls competently. The fifth, a human who refuses to deal with a machine, separated them: whether the transfer rang a real line, whether details were captured first, and what the caller heard when nobody picked up.
• Interruptions are a spectrum. 1.7 seconds of talking over the caller reads as polite; 3.5 makes callers repeat themselves.
• Summaries are only as good as the ear. A street name heard wrong on call one was still wrong on call five, in the contact card an owner would call back from.

Which one, for which business

If the person who takes escalations carries a mobile and is not sitting at a softphone, Upfirst is the one of the two that can reach them, and its emailed summaries mean you see every call without opening a dashboard. If your team already works inside a business phone system and you want the shorter pauses and the tidier intake, CloudTalk is the stronger talker, provided someone is logged in when the angry call comes. In both cases, test the transfer with a rude caller before you trust it with a real one, and read the first week’s summaries against the recordings.

What this test cannot tell you

Five calls per vendor, two vendors, one business, one caller voice, one day. It does not measure accents, Spanish, noisy lines, or how the products behave after a month of real calls and tuning, and it does not measure how they handle a website’s worth of information, because we deliberately gave them one page. The pauses we report come from the live phone audio and include network time, so your results will differ by a few hundred milliseconds either way. We tested what a small business gets in the first hour, with the defaults, which is also how most small businesses run these.

Questions people ask about AI receptionists

How much does an AI receptionist cost?

Among the vendors we signed up for, entry plans with a phone number ran from about $25 to $99 a month after a 7 to 14 day trial, usually with a per-minute or per-call charge above an allowance, plus one free tier of 25 calls a month. Most want a card before they issue the number.

Can an AI receptionist transfer calls to my cell phone?

Some can; some can only transfer inside their own phone system. It was the single biggest difference we found, and it is rarely on the pricing page. Ask before you sign up, and test it with an angry caller, not a polite one.

Will callers know they are talking to an AI?

Most will, after one or two exchanges, from the pauses and the phrasing. Neither product we tested said so unprompted. Whether you must disclose depends on your state and how you use recordings.

Is an AI receptionist better than an answering service?

For routine calls on a defined script, the two products above answered faster than most humans and never quoted a price they should not. For a caller who is upset, confused, or wants a person, a human still wins, which is why the transfer behaviour above matters more than the voice. If you are weighing the two, this page walks through the choice, and this one compares what human services publish as prices.

If you would rather talk it through with someone who has read all ten transcripts, book a call.