AI Receptionist vs Answering Service, the Numbers Nobody Publishes
We sell an AI receptionist. You should read everything below with that in mind, and I have written it so you can check every claim without trusting me. Every price is a published price you can go and look at yourself. Every study is named, dated, and labelled with who paid for it. Where the evidence goes against us, I have said so, and there is a section further down listing seven situations where you should buy a human answering service instead of anything we sell.
I read fifteen of the pages currently ranking for this comparison before writing. Every one of them had a commercial stake. Three disclosed it. None of the fifteen published complete pricing mechanics for both sides. They compared a headline monthly fee against a headline monthly fee, which is the one comparison that tells you nothing, because the fee is not where the money is. The money is in the rounding increment, the billing cycle, the patch-through charge, and the notice period. That is what this page is for.
Key takeaways
- The headline price is not the price. Per-minute answering services commonly round every call up to the whole minute. On short message-taking calls that alone can add 70% to your bill with no change in your usage.
- Ask how many invoices you get a year. Some services bill on 28-day cycles, which is 13 invoices, not 12. That is an 8.3% uplift that never appears on the pricing page.
- One vendor prices both products on the same page. The UK service answer.co.uk charges £2.20 + VAT per call for human answering and £10/month + £0.85 per call for its own AI. Same company, same calls, roughly 2.5x apart.
- Check who owns the phone number before you sign. alldayPA's terms state that any number they allocate to you "will remain our property at all times", and require 90 days' written notice to cancel. Keep your own number and divert to them.
- Disclosure reduces hang-ups, it does not cause them. In the largest published analysis of AI receptionist calls, agents that said they were AI got about 20% fewer hang-ups than ones that did not.
- The five statistics on every competing page are laundered. "62% of calls go unanswered", "$1,200 per missed call" and three others trace back to nothing citable. They are listed below with what they actually trace to.
- Human answering services are genuinely good. Trustpilot ratings across the big UK providers sit at 4.8 to 4.9 across thousands of reviews. If someone tells you the human option is bad, they are selling you something.
The short answer
If you want the decision without the reasoning:
| Your situation | What to buy |
|---|---|
| High volume of similar calls (hours, prices, availability, "do you cover my postcode") | AI |
| Out of hours and overflow only, currently going to voicemail | AI, and this is the least arguable case for it |
| The outcome you want is a booking in a calendar | AI |
| Bursty demand, storm weeks, campaign spikes, outage days | AI, because it answers every line at once and a human rota queues |
| Bereavement, funeral first calls, crisis and safeguarding lines | Human, without qualification |
| Callers who are predominantly elderly, non-native speakers, or on bad lines | Human, and this is the best-evidenced case against AI |
| Fewer than roughly 20 calls a month | Probably neither, buy voicemail-to-text and a disciplined callback window |
| Every first call is consultative discovery, high ticket, low volume | Human |
| A large routine majority plus a small high-stakes minority you can identify in the first 30 seconds | Hybrid: AI fronts, human takes the escalations |
Everything after this is how to check that answer against your own numbers rather than taking mine.
Five statistics you will see everywhere, and where they actually come from
Before the comparison, the ground clearing. These five numbers appear on most pages in this category, usually with no citation or with a citation that goes to another blog post that also has no citation. I tried to trace all five to a primary source.
| The claim | Where it actually leads |
|---|---|
| "62% of business calls go unanswered" | A 2016 study by an SEO firm called 411 Locals, sample size 85 businesses. Not representative of anything, and a decade old. |
| "85% of people who reach voicemail won't call back" | No traceable primary source at all. Cited blog to blog in a loop. |
| "$75 billion lost annually to poor customer service" | A 2018 NewVoiceMedia figure about customer service generally, repeatedly re-attributed to missed phone calls specifically. It was not about missed calls. |
| "80% of callers don't leave a voicemail" | Dead-ends at a 2009 New York Times lifestyle piece. The traceable estimates in this space actually range from about 67% to over 80% depending on source and business type. |
| "$1,200 in lost revenue per missed call" | Widely attributed to Invoca. It does not appear in Invoca's published reports. |
I am not using any of them in this article, and I would treat a page that does as unresearched. If you want a figure for what missed calls cost you, the only honest one is the one you calculate yourself: your average job value multiplied by your close rate, multiplied by the number of calls your phone system shows as unanswered last month. Your phone bill has that number in it.
Two more to handle carefully. "AI receptionists cost 85-95% less" and "saves £1,600-£5,700 a year" are both vendor marketing with no methodology attached. And containment benchmarks in the 40-70% range are vendor-published; they are usable as targets to measure yourself against, not as promises.
What each one actually is
Short section, because everyone covers it and nobody covers it wrong.
A human answering service (also sold as a virtual receptionist or telephone answering service) is a contact centre. Trained agents answer in your business name from a script you supply, take a message, and either email or text it to you or patch the call through to your mobile. Some can book into your calendar as a paid tier. They are staffed 24/7 by rota, which means your calls sit in a shared queue with everyone else's.
An AI receptionist is a voice agent on your line. It answers, understands what the caller wants, works from the information you have given it, can look things up and write into your calendar or CRM, and hands off to a person when it should. It answers every call the moment it arrives regardless of how many arrive at once, because there is no rota.
The three differences that matter operationally, before price:
- Concurrency. A human service queues once its staffed agents are busy. If you run a roofing firm and a storm hits, everyone in your area calls in the same two hours, and hold time is exactly what you cannot afford that morning. AI answers each line on its own thread.
- Consistency of capture. An AI asks the same fields in the same order on every call and produces a transcript every time. Human agents vary, and diagnosing a pattern across a hundred calls is much harder without transcripts.
- Judgement. A human can improvise. That is worth more than every other line in this comparison put together on the calls where it matters, and worth nothing on the calls where it doesn't. Which calls you have is the whole decision.
The cost mechanics, which is the actual comparison
This is the section the ranking pages are missing. Here are the two cost curves.
HUMAN ANSWERING SERVICE (variable, roughly linear in billable minutes)
Monthly cost = plan fee
+ (billable minutes − included minutes) × overage rate
+ patch-through fee × number of transfers
+ add-ons (scheduling, CRM sync, bilingual, out-of-hours)
+ surcharges (bank holidays, weekends)
AI RECEPTIONIST (fixed, with a variable tail)
Monthly cost = flat plan fee
+ (used minutes − included minutes) × overage rate
+ integration or seat tier
The shapes are different, and that is the point. Per-minute billing makes your busiest month your most expensive month. Flat billing makes your quiet month your most wasteful. Which risk you would rather carry is a real business question and neither answer is stupid.
Published prices, checked August 2026
Half the UK market will not publish a price at all. Moneypenny, alldayPA, Answer4u and JAM all gate pricing behind a form. That is itself worth knowing before you start comparing. Here is what is actually published.
| Provider | Model | Published price |
|---|---|---|
| answer.co.uk (human) | Per call | £2.20 + VAT per call |
| answer.co.uk (their own AI) | Monthly + per call | £10/month + £0.85 per call |
| AnswerConnect UK | Plan + per minute | From £45/month, £1.45/min overage |
| Face for Business | Plan | From £49 / £99 / £115 per month, billed per second |
| ReceptionHQ | Plan | From £15/month |
| PATLive (US) | Plan + per minute | $49 + $2.99/min, up to $6,899 for 5,000 minutes |
| Ruby (US) | Minute bundles | 50 min $250, 500 min $1,725 |
| Smith.ai (US) | Per call bundles | $300 for 30 calls ($11.50 overage), $2,100 for 300 calls |
| Specialty Answering Service (US) | Plan + per minute | $44 + $1.54/min, billed by the second |
Prices move. Check them before you sign anything, and note the date you checked.
The one comparison that settles the economics
answer.co.uk sells both products. Human answering at £2.20 + VAT per call. Their own AI at £10/month plus £0.85 per call. Same company, same calls, published on the same site. Run it at three volumes:
| Calls/month | answer.co.uk human (£2.20 + VAT) | answer.co.uk AI (£10 + £0.85/call) | Ratio |
|---|---|---|---|
| 40 | £105.60 | £44.00 | 2.4x |
| 150 | £396.00 | £137.50 | 2.9x |
| 400 | £1,056.00 | £350.00 | 3.0x |
Smith.ai in the US shows the same shape more starkly: their live-agent bundles work out at £7 to £11.50 per call equivalent, and their AI-handled tiers price the same call at a small fraction of that. Both columns are on their pricing page; go and look.
That gap is the honest economic case for AI, and notice what it is not. It is not "85-95% cheaper". It is roughly two and a half to three times cheaper at UK published rates, from a vendor with no interest in flattering AI. Anyone quoting a bigger multiple is comparing against a salaried receptionist, which brings us to the next trick.
The wrong-baseline trick
Most pages in this category compare an AI receptionist against hiring a full-time receptionist at £25,000-£30,000 plus employer's NI, pension, holiday cover and a desk. Naturally AI wins by 95%. But almost nobody choosing between AI and an answering service was ever going to hire a full-time receptionist. That was never the alternative.
The three honest baselines are:
- AI vs an answering service. Roughly 2.5-3x on published UK rates, per the table above.
- AI vs voicemail. This is the real comparison for out-of-hours, and it is not close, because voicemail converts approximately nothing.
- AI vs a part-time person who already works for you answering the phone between other tasks. This is the most common real alternative in a small business and almost no article models it.
There is also an upper crossover that vendors never mention. Past a certain call volume, an actual employed receptionist becomes competitive again, gives you total control, and covers far more minutes than a per-minute service. If you are spending £900 a month on answering, price a part-time hire before you renew.
The rounding increment, the most under-reported cost in this category
Industry-standard practice on per-minute plans is to round every call up to the whole minute. AnswerConnect's own help centre states it plainly: calls over thirty seconds are billed by the whole minute, rounded up. alldayPA rounds up to the whole minute too. Some providers bill in 30-second blocks. A few, like Face for Business and Specialty, bill by the second, and they say so because it is a selling point.
Here is why this is not a detail. Take 40 calls a month on a £1.45/minute rate:
| Average call length | Actual minutes | Billed minutes (rounded up) | Actual cost | Billed cost | Uplift |
|---|---|---|---|---|---|
| 1 min 10 sec | 46.7 | 80 | £67.67 | £116.00 | +71% |
| 2 min 10 sec | 86.7 | 120 | £125.67 | £174.00 | +38% |
| 3 min 10 sec | 126.7 | 160 | £183.67 | £232.00 | +26% |
Rounding hurts most when your calls are short, which is precisely the profile of the businesses that look at per-minute pricing and conclude it is cheap. A message-taking book of short calls is the worst possible case for whole-minute rounding.
Two more things go into "billable minutes" that are not talk time:
- Hold and transfer time is commonly billable. The clock does not stop while they are dialling your mobile.
- After-call work is sometimes billable. AnswerConnect's help documentation states that the time spent writing up and dispatching your message is part of the minutes you pay for.
AnswerConnect does not charge for the first 30 interactions under 30 seconds, which is a genuinely fair concession on wrong numbers, and worth asking every provider whether they match it.
Per-call billing swaps one problem for another
A flat per-call fee removes rounding risk entirely. It also makes wrong numbers, robocalls, silent auto-diallers and eight-second hang-ups fully billable at the same rate as a real enquiry. If a meaningful slice of your inbound is junk, per-call pricing is worse for you than per-minute, not better. Ask what happens to a call that lasts four seconds.
The billing cycle question nobody asks
Some providers bill on 28-day cycles. Thirteen cycles of 28 days is 364 days, so you get 13 invoices a year, not 12. That is an 8.3% annual uplift that appears nowhere on the pricing page. AMBS Call Center is one that documents this model openly.
Ask every provider: "How many invoices will I receive in a twelve-month period?" It is a one-line question and the answer is occasionally surprising.
The full list of things to price before you compare
COST MECHANICS CHECKLIST, ask every provider in writing:
1. Base fee, and exactly what minutes or calls it includes
2. Overage rate, and how it compares to the in-plan effective rate
3. Billing increment: per second, 30 seconds, or rounded up to the minute
4. Is hold time billable?
5. Is transfer / patch-through time billable, and is there a separate
per-patch fee?
6. Is after-call work (writing up and sending the message) billable?
7. What happens to a call under 30 seconds? Wrong numbers? Robocalls?
8. Out-of-hours, weekend and bank holiday surcharges, as a percentage
9. Is appointment booking included or a paid add-on, and at what rate?
10. Is CRM or calendar write-back included, add-on, or not offered?
11. Setup fee, one-off
12. Billing cycle: calendar month or 28 days? How many invoices per year?
13. Minimum contract term
14. Notice period to cancel, in writing
15. Who owns the phone number if one is allocated to me?
Total the twelve-month cost, not the monthly one. Items 11 to 15 only
show up in the annual figure.
The switching costs nobody prices
Getting into an answering service takes between five minutes and three days. Getting out can take three months. That asymmetry is deliberate and it is not disclosed on pricing pages.
Notice periods. alldayPA's terms require 90 days' written notice to terminate. If you sign in January and decide in March that it is not working, you are paying until June.
Number ownership. This is the one that catches people. alldayPA's terms state: "Any telephone number allocated by us to you will remain our property at all times." If the service gave you a number and you put it on your van, your website, your Google Business Profile and 3,000 leaflets, that number is not yours and it does not leave with you.
The fix is one sentence, and it works with any provider, including us. Keep your own number. Divert it to whichever service you are trying. Never advertise a number the provider gave you. Then switching is a five-minute change to a divert rule instead of a rebranding exercise.
Minimum terms. Twelve-month minimums are common and rarely on the pricing page.
Timing note. The UK PSTN switch-off completes in January 2027, so a very large number of businesses are already changing their phone setup for unrelated reasons. If you are going to test something, the migration you are doing anyway is the cheapest moment to do it.
Will callers hang up when they realise it is AI?
This is the fear that actually drives the decision, and the evidence on it is better than you would expect.
Disclosure reduces hang-ups. The largest published analysis of this is a study of 450,702 calls across 503 AI receptionists through June 2026, run by the vendor Upfirst. It is vendor-run, so weigh it accordingly, but the methodology is disclosed and several findings cut directly against the commissioner's interest. The headline: only about 13% of hang-up variance came from controllable settings at all. The other 87% tracked call quality itself, meaning junk and spam calls rather than customers recoiling from a robot. Among the controllables:
| Setting | Effect on hang-ups |
|---|---|
| Disclosing that it is AI | ~20% fewer |
| Mentioning that the call is recorded | ~30% fewer |
| Stating capabilities up front (can text, transfer, book) | ~30-40% fewer |
| Ending the greeting with a question | ~15% fewer |
| Opening by mentioning the owner's absence | ~50% more |
| Voice gender, greeting length, question format | No measurable effect |
The practical reading: pretending to be human is the failure mode, not being AI. And "sorry, the owner isn't available" is the single worst opening line you can use.
UK consumer comfort, from a human answering service. Moneypenny commissioned Censuswide in January 2026 to survey 5,001 nationally representative UK adults. Moneypenny sells human answering, so again, evidence against interest:
- 59% were comfortable with an AI answering their call promptly.
- 54% were comfortable with an AI that clearly states up front what it is.
- 68% were comfortable once they knew they could reach a person at any point.
That last number is the important one. The escape hatch is worth about nine points of acceptance on its own. Hiding the route to a human is the most expensive design decision available to you.
The counterweight, stated fairly. From the same body of research: only 25% of consumers said AI receptionists met their expectations well, against 51% of business decision-makers who thought they did. 83% would still rather speak to a person for high-stakes moments. Separately, Gartner surveyed 5,728 customers in July 2024 and found 64% would prefer companies did not use AI for customer service at all, with 53% saying they would consider switching to a competitor over it.
Both things are true. People are more comfortable than the doom-posting suggests, and they still prefer a person. Neither fact settles the decision on its own; what settles it is which calls you have.
One field observation worth more than any of the surveys. A small business owner running around 200 calls a month through a human answering service reported that only three customers ever noticed they were not speaking to someone in the office. The thing callers detect is that the phone has been outsourced at all, not that it was outsourced to software.
The hard cases, specified rather than asserted
Every competing page has a line like "AI may struggle with strong accents". None of them say what happens next, and none of them cover the group with the most at stake.
Accents and speech differences. In July 2026, Healthwatch Rotherham reported that patients using an AI GP receptionist found it did not respond well to variations in speech, including regional accents and speech impediments. One resident's account is worth quoting exactly, because it is the actual failure mode and not the one people expect:
"I could never get it to understand me, I ended up just hanging up and not bothering to try and book an appointment."
The failure was not a bad call. It was a lost patient, and some were frustrated into travelling to the surgery in person. The vendor's response was that the system is trained on a wide range of accents and dialects, supports 17 languages, transfers when it cannot understand, and never requires a caller to continue with it. Both of those things can be true.
This is not random flakiness, it is systematic. Published evaluations of commercial speech recognition have found roughly 19 errors per 100 words for white speakers against roughly 35 for Black speakers, with the widest gap for Black men. Accuracy degrades with distance from standard American or British English, and non-native accents fare worse than native ones. Peer-reviewed work in JAMIA Open in December 2024 found the same pattern in transcribed patient communication.
The combination is what actually breaks it. A strong accent on a clear landline is usually fine. A strong accent on a speakerphone in a van with the engine running is where recognition collapses. It fails the callers furthest from its training data, who are frequently the ones who most need to reach a person.
Elderly callers. This group gets essentially zero coverage in the comparison pages and it is a large fraction of inbound for GP surgeries, funeral homes, care providers, plumbers and locksmiths. The specific difficulty reported by people who work on call-captioning is that some older callers do not model the interaction as machine-mediated at all: they converse with it, they wait for it to respond conversationally, they raise their voice when it does not. This is not a recognition problem you can fix with a better model. It is an interaction-design problem, and the mitigation is a short greeting, an early and explicit offer of a person, and a low patience threshold before escalating.
What a well-built agent should actually do when it does not understand:
FALLBACK BEHAVIOUR, in order:
1. Clarify, do not guess. "I didn't catch that, are you calling about
a repair, a quote, or something else?" A closed choice recovers
far better than an open re-ask.
2. Read critical data back. Address, callback number, spelling of the
name. Confirm before acting, every time.
3. On a bad line, stop retrying. Offer a callback on a clearer line, or
transfer. Repeated "sorry, could you say that again" is the loop
that makes people hang up.
4. Escalate on the second failure, not the fifth.
5. Never end the call as the resolution to a misunderstanding. Take a
message and send it.
If you want the full version of that logic, including who to route to and what to say while their phone rings, we wrote it up separately in when and how a voice agent should transfer to a human.
Escalation to whom, at 2am?
Every vendor on both sides sells "escalation" as though it were a solved primitive. It is a checkbox on a comparison table. Nobody asks the follow-up question, which is the only one that matters: escalate to whom, and what happens when that person does not answer?
An answering service has the same problem. "We'll patch the call through to you" means your mobile rings at 2am, and if you are asleep the caller gets your voicemail, which is exactly the outcome you paid to avoid.
Write your escalation rules as trigger, action, owner, fallback. Not "escalate if urgent". Like this:
ESCALATION RULE FORMAT
TRIGGER: What in the call fires this rule
ACTION: Exactly what the agent does
OWNER: Which named person or rota receives it
FALLBACK: What happens if the owner does not respond in N minutes
WORKED EXAMPLE, out-of-hours emergency
TRIGGER: Caller states an emergency keyword (leak, flood, no heat,
gas, lockout) between 18:00 and 08:00
ACTION: Collect name, callback number, postcode, one-line problem.
Do not qualify further. Send SMS to on-call rota. Tell the
caller the on-call engineer has been notified and will
call back.
OWNER: On-call rota, current week
FALLBACK: No acknowledgement within 10 minutes, SMS the second name
on the rota and email the owner. Tell the caller, if still
on the line, that a second engineer has been alerted.
Escalation is a spectrum, not a binary. The options, cheapest to most expensive:
- Tagged summary for morning review
- Email to a shared inbox
- SMS alert to a named person
- Callback task created with a deadline
- Warm transfer with a spoken briefing
Choosing the cheapest sufficient escalation for each trigger is what keeps the whole thing affordable, on either side of this comparison. A business that warm-transfers everything has bought a very expensive switchboard.
And always offer the human exit unprompted. People have developed folk techniques for escaping phone systems, including saying nonsense words until the system gives up and routes them to a person. If your callers are doing that, you have designed the system badly. Given that the escape hatch is worth nine points of comfort on its own, hiding it is indefensible.
The hybrid cost trap
"AI first, human escalation" is the answer a lot of people land on, and for the right call mix it is genuinely correct. Moneypenny, the largest human provider in the UK market, launched its own AI Voice Agent in October 2025, and its chief executive has said 2026 is when hybrid becomes standard. When the biggest incumbent in the human category builds an AI front end, that is a signal worth taking seriously.
But there is a cost trap in hybrid and nobody writes it down.
You pay for both, and the human minimum does not shrink in proportion to your escalation rate. If you buy a 50-minute human bundle and an AI plan, and the AI successfully handles 80% of calls, you do not pay 20% of the human bundle. You pay the whole bundle, because it is a minimum. Your effective cost per escalated call goes up as your AI gets better, right up until you can drop to a smaller human tier, which is a step change and not a slope.
Model it this way before you buy hybrid:
Hybrid monthly = AI plan
+ human plan MINIMUM (not your expected usage)
+ human overage if escalations exceed the minimum
Effective cost per escalated call
= (human plan minimum + overage) ÷ escalated calls
Run this at your best-case escalation rate AND your worst-case one.
If the best case makes the human tier look wasteful, buy the smaller
tier and accept overage on bad months. Overage on a small tier usually
beats waste on a large one.
When hybrid is genuinely correct: when your call mix is bimodal, a large routine majority plus a small high-stakes minority, and the minority is identifiable within the first thirty seconds. That last clause is the real test. Hybrid works when the trigger can be written down. It fails when the hard calls only reveal themselves at minute four, because by then you have paid for the AI leg and the human leg and annoyed the caller.
Two hybrid shapes, and most people should start with the first:
- Time-based. Humans in hours, AI out of hours and on overflow. Simplest, cheapest, lowest risk, and it competes against voicemail rather than against a person.
- Intent-based. AI fronts every call and escalates by topic or sentiment. Higher ceiling, needs real configuration discipline and weekly transcript review to stay honest.
Reported splits for well-tuned setups land around 15-25% live transfers with 60-75% captured and called back, but those figures are vendor-published. Treat them as targets to measure against, not as promises.
Test it yourself, a scripted call battery
This is the part you can actually reuse. Every vendor on both sides will give you a demo number or a trial. Demos are run on the calls the vendor knows it handles. Run these instead. Place each call, score it, and keep the recording.
THE 12-CALL TEST BATTERY
Run identically against every provider you are shortlisting.
Score each 0-2: 0 = failed, 1 = handled but clumsy, 2 = handled well.
--- BASELINE ---
1. "What are your opening hours?"
Pass: correct answer in one turn, no transfer.
2. "Do you cover [a postcode just outside your area]?"
Pass: says no clearly, offers something useful. Fail: says yes.
3. Book a routine appointment for a specific day and time.
Pass: it lands in your calendar with the right details.
Then check: did it WRITE BACK, or just email you a message?
--- THE HARD CASES ---
4. Call from a mobile, outdoors, with traffic or a fan running.
Pass: recovers or escalates. Fail: loops "sorry, say that again".
5. Have someone with a strong regional or non-native accent place
call 1 and call 3. Score separately. This is the single most
predictive test in the battery.
6. Speak slowly with long pauses, the way an older caller often does.
Pass: waits. Fail: interrupts or times out.
7. Give a name with an unusual spelling and a postcode.
Pass: reads both back before acting.
--- THE FAILURE BRANCHES ---
8. Ask for something outside its knowledge entirely.
Pass: says it does not know and takes a message.
Fail: invents an answer. This is disqualifying.
9. Ask for a human on the first turn, plainly.
Pass: routes immediately, no friction, no third question.
10. Trigger an out-of-hours emergency at an actual out-of-hours time.
Then DO NOT ANSWER the escalation. Wait 15 minutes.
Pass: the fallback fires. This is the test almost everything fails.
11. Hang up in silence after 4 seconds. Then check the invoice or
usage log at the end of the trial. Were you billed for it?
12. Place three calls simultaneously from three phones.
Pass: all three answered. Human services will queue at least one.
--- AFTER THE TRIAL ---
- Request the transcripts or call recordings for all 12.
A provider that cannot produce them cannot help you improve.
- Reconcile every call against the invoice. Check the rounding.
Calls 5, 10 and 12 are the ones that separate providers. Almost everyone passes calls 1 to 3, which is why demos are useless.
Who should honestly buy a human answering service
We sell AI. These are the situations where you should not buy it from us or from anyone.
1. Bereavement, funeral first calls, and death notifications. The content of these calls is procedurally simple and the tone is absolute. Leading funeral answering services train representatives for months before they take calls unsupervised. A standard greeting is actively wrong here: "thank you for calling, how may I direct your call?" is a failure when the person on the other end has just lost their spouse. If your business takes first calls, a person takes them.
2. Mental health, crisis-adjacent, and safeguarding-exposed lines. The clinical literature is direct that the ability to share distress with another person is part of what makes crisis services work, and that AI errors in this context risk making acute distress worse rather than better. AI's defensible role here is supporting human workers, not answering. Note also that even the human service is a supplement to a documented crisis pathway, never a substitute for one.
3. Fewer than about 20 calls a month. Below this, any subscription is mostly waste, and per-call human billing or good voicemail-to-text plus a published callback window will beat both. The honest test for a solo practice is how many prospects would have hung up rather than leave a message. If that number is three, the prize is three.
4. Every first call is consultative discovery. High ticket, low volume, bespoke, where the first conversation is the sales process and improvisation is the whole job. The maths that matters is conversion rate, not cost per minute, and a 2% conversion difference dwarfs the entire price gap.
5. A caller base that is predominantly elderly, non-native speaking, or speech-diverse. This is the best-evidenced case in this article and it is the one AI vendors are quietest about. The failure is not annoyance, it is abandonment. Read the Rotherham quote again. If a meaningful share of your callers sit in those groups, run test call 5 above before anything else, and treat the result as the decision.
6. Where the call itself constitutes regulated advice, or the provider cannot evidence the compliance stack. For UK healthcare and legal work, ask for the data processing agreement and check it covers the entire chain: telephony, speech-to-text, the language model, and text-to-speech. Not just the front end. A vendor that can only produce a DPA for its own layer has not answered the question. In the US the equivalent question is whether the vendor will sign a BAA covering the same full stack. Compliance is a gate, not a criterion. Fail it and price is irrelevant.
7. Where the first impression is the product. Some businesses are bought on the feel of the first phone call. If yours is one, you already know it.
And a general point of fairness: the human services in this market are good. Trustpilot shows Answer4u at 4.9 across 494 reviews, AnswerConnect at 4.9 across 1,517, Moneypenny at 4.8 across 1,281, alldayPA at 4.8 across 433, Ruby at 4.6 across 848. AnswerConnect markets itself explicitly on "people, not bots", which is a legitimate position honestly held. The lowest-rated of the group at 4.2 across 348 reviews is Smith.ai, the most AI-heavy of them, and the recurring AI-specific complaint in reviews across the category is worth naming: loop behaviour, the agent re-asking a question the caller already answered. That is the concrete failure to test for, not "robots lack empathy".
Who should buy AI
1. High volume of similar calls. Opening hours, do you cover my area, what do you charge, are you open Saturday, is my order ready. If most of your inbound could be answered from a one-page script by a competent temp, that is the AI-fit signature. Note it is homogeneity, not volume that predicts fit. A moderate number of similar calls is a much better fit than a high volume of varied ones.
2. Out of hours and overflow. The strongest and least arguable case, because the competition is voicemail, not a person. And the urgency is real: research in this space consistently finds only about a quarter of people will try a second time after a failed call attempt.
3. When the outcome is a booking. Where success is a slot in a calendar rather than a nuanced conversation, same-call booking beats message-taking by a wide margin, and human services usually put calendar booking behind a paid add-on at a higher per-minute rate anyway.
4. Qualification and routing. Capture name, number, location, issue type, urgency, route accordingly. This is a script, not a judgement call.
5. When you need CRM or calendar write-back. A human service can do this too, usually as a premium tier. AI's structural edge is consistency: the same fields, every call, with a transcript. Ask one question here that catches most vendors out: is the integration native or is it Zapier? Google Calendar, Outlook, Calendly and HubSpot are commonly native. The systems trades actually run, ServiceTitan, Jobber, Housecall Pro, Commusoft, are frequently reached through webhooks or Zapier instead. Then test write-back, not just read.
6. When predictability matters more than the absolute lowest price. Seasonal and weather-driven businesses especially, where per-minute billing spikes in exactly the months cash is tightest.
7. Simultaneous-call spikes. Storm weeks for roofers and plumbers, outage days for IT support firms, campaign launches for retail. Concurrency is the one dimension where the AI advantage is architectural rather than incremental.
8. Solo operators who physically cannot answer. Up a ladder, under a sink, mid-treatment, driving. Electricians, driving instructors, massage therapists, single-handed dental practices. Here the comparison is not AI against a receptionist. It is AI against a phone that rings out.
The questions to send every vendor, both kinds
Paste this into an email. The answers, in writing, are worth more than any comparison table including this one.
PRICING
1. Your full fee schedule, including overage rates.
2. What billing increment do you use? Per second, 30 seconds, or
rounded up to the whole minute?
3. Is hold time billable? Transfer time? After-call write-up time?
4. What do I pay for a call that lasts under 30 seconds?
5. Out-of-hours, weekend and bank holiday surcharges, as percentages.
6. Is appointment booking included or an add-on? At what rate?
7. Calendar month billing or 28-day cycles? How many invoices a year?
8. Setup fee?
CONTRACT
9. Minimum term?
10. Notice period to cancel, and does it have to be in writing?
11. If you allocate me a phone number, who owns it if I leave?
12. Can I export my call data, transcripts and contact records, and
in what format?
CAPABILITY
13. Which of my systems do you integrate with NATIVELY, as opposed to
via Zapier or a webhook? Name them.
14. Does the integration write back, or only read?
15. Can I see a transcript or recording of every call?
ESCALATION
16. What exactly happens when you try to reach my on-call contact and
they do not answer? Walk me through the next 15 minutes.
17. Can I set different escalation rules by time of day and by
call type?
FOR AI PROVIDERS SPECIFICALLY
18. What does the agent say when it does not know the answer?
19. How many failed understanding attempts before it escalates?
20. Show me your data processing agreement, and confirm it covers
telephony, speech-to-text, the model, and text-to-speech.
21. Can I have a trial line to run my own test calls, including calls
from accented and elderly callers?
FOR HUMAN PROVIDERS SPECIFICALLY
22. How many of your agents will be trained on my account?
23. What is my hold time at peak? How is that measured?
24. Do you use AI on any part of my calls, and will you tell my
callers if so?
Question 24 is not rhetorical. Several human providers now run AI on some portion of calls, and the boundary is not always disclosed.
What this comes down to
The honest framing is not AI against humans. For most small businesses the real comparison is AI against the call going unanswered, because the answering service was never actually going to get bought and the receptionist was never going to get hired. If your phone currently rings out after six, that is the baseline, and almost anything beats it.
Where a person is genuinely better, buy the person. The seven cases above are real and they are specific. Where the calls are routine, the volume is bursty, or the alternative is voicemail, buy the software. And whichever you buy: keep your own phone number, ask how many invoices a year you will get, and run the twelve calls before you sign.
If you want to see how an AI receptionist handles the hard branches rather than the demo ones, we have written up the two that break most setups: when and how it should transfer to a human, and the guardrails that stop it inventing answers.
Frequently asked questions
How much does an answering service cost per month in the UK? Published entry prices range from about £15 a month (ReceptionHQ) to £45-£115 a month (AnswerConnect, Face for Business), plus per-minute or per-call charges on top. Per-call pricing at answer.co.uk is £2.20 + VAT. Half the UK market, including Moneypenny, alldayPA, Answer4u and JAM, does not publish prices at all. Your actual bill depends far more on the billing increment and your average call length than on the headline plan fee.
Is an AI receptionist cheaper than an answering service? On published UK rates, roughly two and a half to three times cheaper. The cleanest comparison is answer.co.uk, which sells both: £2.20 + VAT per call for human answering against £10 a month plus £0.85 per call for their own AI. Claims of "85-95% cheaper" are comparing against a full-time salaried receptionist, which was probably never your alternative.
Will my customers know they are talking to AI? Some will. The evidence says telling them is better than hiding it: in an analysis of 450,702 AI receptionist calls, agents that disclosed they were AI saw roughly 20% fewer hang-ups than those that did not. UK survey data from Censuswide in January 2026 found 68% of consumers were comfortable with AI answering once they knew they could reach a person at any point.
What happens if the AI cannot understand my caller? A well-configured agent should clarify with a closed choice rather than re-asking openly, read critical details back, and escalate on the second failure rather than looping. If a caller is on a bad line it should offer a callback or transfer rather than retrying. Test this before you buy: place a call from a moving car and one from someone with a strong accent.
Can an AI receptionist book into my calendar and update my CRM? Usually yes for Google Calendar, Outlook, Calendly and HubSpot, which are commonly native integrations. Trade systems like ServiceTitan, Jobber, Housecall Pro and Commusoft are frequently reached via webhooks or Zapier instead. Ask specifically whether the integration is native, and test that it writes back rather than only reading.
Is a hybrid setup, AI first with human escalation, the best of both? For the right call mix, yes: a large routine majority plus a small high-stakes minority you can identify within the first thirty seconds. Be aware of the cost trap. You pay the human plan minimum regardless of how few calls escalate, so a very effective AI can make your cost per escalated call rise until you can drop a tier.
Do I lose my phone number if I switch providers? You can. alldayPA's terms state that any number they allocate to you remains their property. The fix works with any provider: keep your own number, divert it to the service, and never advertise a number they gave you.
How long does it take to cancel an answering service? Check before you sign. alldayPA requires 90 days' written notice. Setup typically takes between five minutes and three days, so the asymmetry between getting in and getting out is substantial and it is not on the pricing page.