Disclosure, up front
We build AI call handling for trades and service businesses, so we sell this. That is a reason to read the next section sceptically, and also the reason we can tell you what the demos leave out. Where we think you should not buy one of these at all, we have said so, and that section is longer than the one about us.
The demo call always works. It works because the person running it knows exactly which questions the assistant can answer, and asks those. Your customers will not do that. They will ring from a van with the window down, ask something nobody thought to configure, interrupt halfway through, and want a person.
So the useful question is not "how good is the AI". All of these sound convincing now. The useful question is what happens in the ninety seconds after it stops understanding, because that is where the missed job, the annoyed customer and the refund request all live.
The short version
- There are four failure behaviours. Dead end, message, cold transfer, warm transfer. Find out which one you are buying before you sign.
- Warm transfer is frequently a paid upgrade. Rosie publishes it on the $149 plan, not the $49 one, and waterfall transfer on the $299 one.
- The dangerous failure is not silence, it is confidence. An assistant that invents your Saturday hours costs more than one that says it does not know.
- Scope is the safety mechanism, not smarter AI. A small set of facts it genuinely holds, and everything else escalates.
- Make seven test calls before buying. The script is in section 05. It takes twenty minutes and it is the whole evaluation.
- Sometimes the answer is do not. Section 07, and it applies more often than vendors admit.
01The five ways a real call goes off script
In rough order of how often they happen:
1. A question nobody configured
"Do you work on commercial units or just residential?" If that was not written into its knowledge, the assistant has two options: say it does not know and escalate, or guess. Which one it does is configuration, not intelligence.
2. The line is bad
Roadside, a busy workshop, a speakerphone in a stairwell, an accent the model handles poorly. Transcription degrades, and the assistant starts responding to words that were not said. Good ones notice repeated low confidence and hand off. Poor ones confidently answer the wrong question.
3. The caller is already annoyed
Someone calling about a job that went wrong does not want a triage questionnaire. Detecting frustration and getting out of the way is one of the clearest dividing lines between a system built by people who have listened to real calls and one that has not been.
4. The assistant is confidently wrong
The worst one, and the least visible. It does not sound like a failure. It sounds like a smooth, helpful answer that happens to be untrue: the wrong price, hours you do not keep, a service you stopped offering. Nobody reports it, because the caller has no reason to doubt it. You find out when someone turns up expecting something you never agreed to.
5. The caller just wants a person
Some people say "operator" before the greeting finishes. This should be the easiest path in the whole system and is routinely the worst, because it is not what the product demo is designed to showcase.
02The four escalation behaviours, worst to best
Every product does one of these when it gets stuck. This is the single most useful thing to establish before buying, and it is rarely on the pricing page.
| Behaviour | What the caller experiences | What it costs you |
|---|---|---|
| Dead end | It loops, repeats the same prompt, or restates the question until they hang up | The job, and often the customer. They ring a competitor next |
| Take a message | It admits it cannot help and captures a callback number | Acceptable. You still have the lead, but the caller waits |
| Cold transfer | Straight through to a human, who knows nothing about the call | The caller repeats everything. Mildly irritating, rarely fatal |
| Warm transfer | The assistant briefs your team first, then connects | Best outcome. Frequently a paid tier |
A fifth exists at the top end: a waterfall transfer, which tries several numbers in turn until a human picks up. Useful if your team is in vans rather than at desks.
The thing worth noticing is that this ladder maps onto price, not just quality. Rosie, to take a vendor that publishes its tiers clearly, lists warm transfers and live call transfers on its Scale plan at $149 per month, not on the $49 Professional plan, and puts waterfall transfers on Growth at $299. So the behaviour that matters precisely when the assistant is failing can sit two tiers above the number that got your attention.
That is not a criticism of Rosie, which is clearer about it than most. It is a warning about how to read every quote in this market: price the tier that has the escalation you need, not the entry tier.
03The failure that actually costs money
Of the five, number four is the one to design against. An assistant that says "I am not sure, let me get someone" has failed safely. An assistant that invents a price has failed expensively, silently, and in a way that reaches your customer as a promise you now have to honour or refuse.
The instinct is to fix this with better AI. That is the wrong lever. The reliable fix is scope: give the assistant a small set of facts it genuinely holds, and make everything outside that set escalate rather than improvise. A system that knows five things and hands off everything else outperforms one that will attempt anything, because you can predict what it will do.
So when a vendor demonstrates breadth, ask the opposite question. Not "what can it handle" but "what is it configured to refuse". A confident answer to that is the strongest signal you will get in a sales call.
04What these cost, and on what unit
Published prices as of 5 September 2026. The trap is that they are billed on three different units, so the headline numbers are not comparable until you convert them to your own call volume.
| Product | Billed on | Published plans | Escalation notes |
|---|---|---|---|
| Goodcall | Per agent, per month | Starter $79, Growth $129, Scale $249. Annual billing $66, $108, $208 | Transfer to directory contacts; escalation detail not published |
| Rosie | Minutes per month | Professional $49 (250 min), Scale $149 (1,000 min), Growth $299 (2,000 min) | Warm and live transfer on Scale; waterfall on Growth |
| Smith.ai | Calls per month | Starter $300 (30 calls), Basic $810 (90), Pro $2,100 (300) | Offers AI-first and human-first as separate choices |
Read from each vendor's own pricing page on 5 September 2026 and simplified to the fields that matter here; each has more features per tier than shown. Smith.ai's per-call rates work out at roughly $10, $9 and $7 a call as volume rises. Prices change, so verify before budgeting.
Two things fall out of that table. Minute-based billing punishes long calls, which are usually your most valuable ones, so a chatty customer booking a big job costs more than a quick wrong number. And per-call billing punishes junk, since a robocall and a booking cost the same. Neither is wrong. They just reward different call mixes, and you already know which one you have.
05Seven test calls to make before you buy
This is the whole evaluation, and it takes about twenty minutes. Insist on a number you can dial yourself rather than a demo someone drives for you. A vendor who will not give you one has told you something.
- Ask something outside its scope. Something real but unconfigured. You are watching for whether it admits ignorance or invents an answer.
- Call from a moving car with the window down. Genuine background noise, not a quiet room. Does it degrade gracefully or answer a question you did not ask?
- Interrupt it mid-sentence. Real callers talk over the greeting constantly. Does it stop and listen, or finish its script?
- Ask for a human in the first three seconds. Before it has finished introducing itself. Time how long it takes and count how many refusals you get.
- Ask the same factual question three times, in three different ways. Prices, hours, service area. If the three answers disagree, it is improvising, and it will improvise at your customers.
- Call after hours. The whole reason to buy one of these. Confirm it behaves differently at 11pm than at 11am, and that urgent work reaches somebody.
- Make an awkward request. Be a bit rude, or vague, or upset. Not to be unpleasant, but because some of your customers will be, and you need to see it happen before they do it.
Then listen to the recordings
Not the transcripts, the audio. Transcripts hide the pauses, the talking over each other, and the two seconds of silence that make a caller say "hello? hello?". A system can read perfectly on paper and sound broken on the phone.
06Things worth settling in writing
- Does it say it is an AI? It should, in the first sentence. Callers work it out anyway, and the ones who feel deceived are the ones who post about it.
- Who holds the call recordings and transcripts, and for how long? These are recordings of your customers. Some places regulate recording and automated calling, so check what applies where you operate.
- What happens when the service is down? Ask where calls go if their system fails at 8am on a Monday. The correct answer is that they fall back to ringing your phone, not into a void.
- Can you read every call, not just the booked ones? The calls that did not convert are the ones that tell you what to fix.
- Who edits what it knows, and how fast? If changing your holiday hours needs a support ticket, it will be wrong for a week every year.
07When you should not buy one at all
More often than this market admits.
- Your calls are emotional or safety critical. Anything where a worried person needs a competent human immediately. An assistant is a layer between them and help.
- Your call volume is genuinely low. If you can return every voicemail within the hour, a callback habit costs nothing and sounds better. Do that until it stops working.
- Nobody will maintain it. An assistant quoting last season's prices is worse than a missed call, because the caller believes it. If no one owns updating it, it decays into a liability.
- Every job is bespoke. If quoting requires judgement on every call, there is little for it to capture beyond a name and number, which an answering service does more cheaply.
- The real problem is that nobody calls back. An AI receptionist will capture more leads and then those will also sit unactioned. Fix the follow-up first; the missed calls may not be your constraint.
If you have been through that list and the case still holds, you are buying for the right reason: calls arriving when nobody can pick up, and revenue leaking because of it.
08Where we fit
We build AI call handling as part of a wider automation engagement rather than as a subscription you self-serve. In practice that means the assistant answers, captures the job details, books the appointment and texts you a summary, and the parts either side of the call, the intake, the quoting, the follow-up, are wired into the same system rather than left as separate tools.
That model suits you if the phone is genuinely how your work arrives and the surrounding admin is also manual. It is the wrong shape if you want a $49 subscription you configure yourself this afternoon, in which case the products in section 04 will serve you better and cost far less. Our pricing is published in full here: builds start at $1,000 and managed operations are $750 a month, after a free audit.
Whichever way you go, make the seven calls. They are the same seven calls whether you are testing us or anyone else, and we would rather you made them.
09Common questions
What does an AI receptionist do when it cannot answer a question?
That depends on how it was configured, and there are four common behaviours. The worst is a dead end, where it loops or repeats itself until the caller hangs up. Better is taking a message. Better still is a cold transfer, putting the caller straight through to a human. Best is a warm transfer, where the assistant briefs the person before connecting them so the caller does not repeat everything. Ask which of the four you are buying, because the difference is the whole product.
Is warm transfer included in the price?
Often not on the entry plan. Rosie, for example, publishes warm transfers and live call transfers on its Scale plan at $149 per month rather than its $49 Professional plan, and waterfall transfers, which try multiple numbers until someone answers, on its $299 Growth plan. So the behaviour you most need when the assistant is stuck can sit two tiers above the price you were quoted. Check which tier the escalation you want lives on.
How much does an AI receptionist cost in 2026?
Published prices span a wide range because they are billed on different units. Rosie charges $49 per month for 250 minutes, $149 for 1,000 and $299 for 2,000. Goodcall charges $79 per agent per month on Starter, $129 on Growth and $249 on Scale, with roughly 16 percent off for annual billing. Smith.ai bills per call: $300 per month for 30 calls, $810 for 90 and $2,100 for 300, which works out at roughly $10, $9 and $7 per call. Convert every quote to your own expected call volume before comparing.
Can an AI receptionist tell when it is wrong?
Not reliably, and this is the failure mode that costs the most. An assistant that does not know your Saturday hours will not usually say so; it will produce a plausible answer in a confident voice. The protection is not better AI, it is scope: give it a small set of facts it genuinely holds, and make everything outside that set escalate rather than improvise. Ask any vendor what their assistant does with a question outside its knowledge, and treat vague answers as the answer.
How do I test one before buying?
Call it yourself, seven times, and go off script deliberately. Ask something outside its scope. Call from a moving car with background noise. Interrupt it mid-sentence. Ask for a human immediately. Ask the same factual question three times to see whether the answers agree. Call after hours. Give a genuinely awkward request. The demo will always work; these seven calls are what tell you how it behaves on the day it matters.
Will it annoy my customers?
Some of them, yes, and it is worth being honest about that rather than pretending otherwise. What reliably annoys people is not the assistant being artificial, it is being trapped: no route to a person, having to repeat themselves after a transfer, or being confidently told something untrue. Callers tolerate an assistant that is quick, clear about what it is, and hands over cleanly. Compare it against your real alternative, which for most small businesses is a phone that rings out.
Should it say that it is an AI?
Yes. Disclose it in the first sentence. Callers usually work it out anyway, and the ones who feel deceived are the ones who complain publicly. Some jurisdictions also regulate automated calling and voice recording, so check what applies where you operate before recording or transcribing calls. Disclosure costs nothing and removes the worst version of the complaint.
When is an AI receptionist the wrong choice?
When most calls are emotionally loaded, genuinely complex, or safety critical, an assistant adds a layer between a worried caller and the person who can help. It is also wrong when nobody will maintain it: an assistant quoting last season's prices does more damage than a missed call, because the caller believes it. And if your call volume is low enough that you can return every voicemail within the hour, a callback habit costs nothing.
Not sure the phone is your actual bottleneck?
Tell us what happens to a call that comes in while everyone is on a job. If the honest fix is a $49 subscription or just calling people back faster, we will say so.
Ask for a fit checkSources: Goodcall, Rosie and Smith.ai pricing pages, read 5 September 2026. Escalation notes are drawn from what each vendor publishes; where a vendor does not describe its escalation behaviour, the table says so rather than inferring. Outbound mentions are unpaid and carry no affiliate tracking. Prices change without notice, so verify before you budget. Related: Best AI Automation Agencies for Small Business (2026) and Best AI Website Builders for Small Business Owners (2026).