Skip to content

AI receptionist, answering service or voicemail: what actually answers your phone

GUIDE · COMPARISON · 10 MIN

Facts on this page verified August 2026.

The short answer: it turns on your call volume and on whether it can book

Voicemail recovers almost nothing, and the other three each win in a different band. A human answering service takes messages well and bills by the minute, which stops making sense as volume climbs. An AI voice agent earns its place only if it can book rather than take a message. Missed-call text-back is the cheapest fallback and almost nobody sets it up.

The leak you are actually trying to close

Before comparing anything, be honest about the size of the hole. Owners are consistently wrong about how many calls their firm answers, and they are wrong in one direction.

97% believed · 66% measured

What home-service firms think their answer rate is, against what it measured at.industry studySource: Service Direct, 2019

The second number is the one that decides which option you need. Answering the phone is not the goal. Booking the job is the goal, and the gap between the two is where most of these products quietly fail.

42% · 24%

Share of inbound trade calls that become booked jobs, and the same share at firms under five technicians.industry studySource: ServiceTitan, 2022

Read together, they set the test for everything below. A solution that raises your answer rate and not your booking rate has moved the problem rather than solved it. That is the single most useful question to ask any supplier on this list, including us: what can it finish without a human?

The five options, side by side

There are five real options and most comparisons only show you three. The two that get left out are the cheapest one and the one the writer is not selling. Both are here.

What each option costs in shape rather than in euros, what it can finish on its own, and the way each one fails.
OptionCost shapeCan finish aloneOut of hoursHow it fails
VoicemailFree, already on your lineNothing. It records and waits.Yes, technicallyCallers no longer leave messages. The ones who do have already called somebody else.
In-house receptionSalary, fixed, whether the phone rings or notEverything, including judgementNoOne person cannot be at the desk, on the phone and in the treatment room. Holiday and sickness are total outages.
Human answering servicePer minute or per call, so cost rises with successTakes a message. Books only where you have given it your diary.Usually yes, often at a premiumIt knows your business shallowly. Callers can tell, and complex questions come back to you anyway.
AI voice agentFlat monthly, largely independent of volumeBooks, reschedules, answers routine questions, qualifies and routesYes, identically at 03:00 and at 15:00Accents, background noise and genuinely unusual requests. Fails badly if it cannot hand over to a person.
Missed-call text-backCents per messageSends a booking link. Recovers the caller, not the conversation.YesUseless for callers who need to speak to someone now, and it is a fallback rather than an answer.

The column that decides it is the third one. Cost shape tells you what happens as you grow. A per-minute service is a variable cost that rises exactly when you are busiest, which means the better it works the more it costs, and the flat-fee options invert that. But neither matters if the thing on the end of the line cannot complete the job the caller rang about.

The decision, by call volume

Volume, not enthusiasm, decides this. Count your inbound calls for one month before you buy anything. Most firms guess low by a wide margin, and the answer changes at two thresholds.

  • Under roughly 100 calls a month. Neither a human service nor an AI agent will pay for itself on call volume alone. Set up missed-call text-back, put a real booking link in your email signature, and spend the money on something else. Anyone who sells you a receptionist at this volume is not doing you a favour.
  • Roughly 100 to 200 calls a month. This is the band where it genuinely depends. A human service works well if most calls are simple message-taking and your margins are healthy. An AI agent wins if a meaningful share of those calls are bookings, because booking is the thing per-minute services do worst.
  • Over roughly 200 calls a month. Per-minute economics have turned against you and the flat-fee options win on cost as well as on capability. This is also the point at which after-hours volume stops being a rounding error, and out-of-hours is where a human service is most expensive and an agent costs the same as at noon.

Put your own numbers through the arithmetic rather than trusting the bands. The missed-call calculator takes your call volume, your answer rate and your average job value and prices the leak per month, with the model shown so you can argue with it. That number is what any of these options has to beat.

Where an AI agent is the wrong answer, from the people who build them

We sell AI phone agents. Here is where we tell people not to use one. Every vendor comparison on this subject is written by somebody who cannot afford to say this, which is precisely why it is worth reading.

  • The caller is already angry. A complaint routed to a machine becomes a worse complaint. Anger detection is the first rule we write into a handover, and the correct behaviour is a fast, unapologetic transfer to a person.
  • The call is clinically or legally sensitive. Symptoms, safeguarding, anything a professional obligation attaches to. An agent may take the name and the callback number. It should not take the substance.
  • Intake is genuinely complex and varies per caller. If the useful version of the call is twenty minutes of a specialist asking follow-up questions, an agent will produce a tidy transcript of the wrong information. Book the specialist instead.
  • Your existing reception is not busy. If the phone is answered in seconds by somebody who knows every caller by name, that is better than anything on this page and it is not close. Automate the overflow and the out-of-hours, not the person.
  • You are not willing to disclose it. A firm that wants the agent to pass for human has already chosen the failure mode. In Germany that is also now a legal question rather than a taste question, and it is a bad one to get wrong.

That last one has moved from etiquette to obligation, and it deserves its own page. Disclosure, call recording and where the data lives are the three things German buyers ask about first, and the answers are more specific than most suppliers admit.

The hybrid almost everyone ends up with

Nobody picks one row of that table and stops. The arrangement small firms settle into is layered, and it is worth designing on purpose rather than arriving at by accident.

  1. Your own people answer first, for as long as they can. Two or three rings. Nothing beats a person who knows the caller, and the layers below exist to protect that person from the phone rather than to replace them.
  2. The agent takes the overflow, out-of-hours and the second line. It picks up in seconds, states plainly what it is, books into your calendar rules, and never queues anybody. The same at 03:00 as at 15:00, which is the part a rota cannot do.
  3. Anything sensitive, angry or unusual goes to a human immediately. The handover rule is written at build time, not discovered in production, and the number it transfers to is a real one that somebody answers.
  4. Missed-call text-back catches whatever still falls through. Nothing is perfect. The last layer costs cents and turns a dropped call into a booking link inside ninety seconds.

None of this is experimental in German healthcare, which is usually the most conservative room in the argument. It is already the norm at a scale that makes the question less about whether it works and more about whether it is configured properly.

16,000 practitioners · ~3M calls a month

German healthcare professionals already running AI phone assistance.public figures (aaron.ai/Doctolib)Source: aaron.ai (a Doctolib company), 2026

What we build is the middle two layers, and we build them to hand over well. The Customer Facing Agent answers in seconds, discloses itself in the first turn, books into your rules, and routes anything it should not handle to a person with the context attached. The first working version runs in ten working days, and the old path stays switched on underneath it until you say otherwise.

Every source on this page

Each claim above is numbered to one of these. Open them and check.

  1. 1. Service Direct: 2019 Home Service Call Performance Report (2019)

    research

    1,000 recorded calls to 94 US home-service firms across 42 categories, scored by hand. The calls came from the publisher's own pay-per-call campaigns, which it states openly, and it is US data from 2019. It is the perception gap that travels, not the exact percentages.

  2. 2. ServiceTitan: Data report: average call booking rates (2022)

    research

    June 2022 data from more than 3,000 trade businesses in the United States and Canada that use the publisher's software. A typical shop booked 42% of calls; shops with fewer than five technicians booked 24% and shops with 25 or more booked 59%. Software users are not a random sample of the trade.

  3. 3. aaron.ai (a Doctolib company): Company figures published on the aaron.ai home page (2026)

    market observation

    Self-reported vendor figures, read on 14 August 2026: 16,000 doctors, therapists and healthcare workers, and 3 million calls a month. Nobody audits them. They are quoted here as evidence that this is ordinary in German healthcare, not as evidence that it works.

Will my customers know they are talking to a machine?

Yes, because it says so in the first sentence. That is our default and since August 2026 it is also what EU law expects when a person is interacting with an AI system. In practice callers mind far less than owners fear, provided the agent is fast, gets things right and hands over to a person the moment it is asked.

What happens on a complicated call?

It hands over. The rules for when are written before the agent goes live rather than discovered afterwards: anger, clinical or legal substance, anything outside its list, and any caller who asks for a person. The handover carries what it already collected, so the caller does not repeat themselves.

Does it work with my existing phone number?

Yes. Your number stays yours and the calls that reach the agent are the ones you route to it: overflow after a set number of rings, out-of-hours, or a second line. Nobody has to learn a new number and nothing changes on your letterhead.

What does it do out of hours?

The same thing it does at midday, which is the actual argument for it. Out-of-hours is where a human service charges the most and where voicemail loses the most, and it is a large share of inbound calls for trades and practices because that is when customers are free to ring.

What if it mishears something important?

It confirms back the things that matter, name, number, date and reason, and anything it cannot confirm becomes a handover rather than a booking. You get the transcript either way. An agent that guesses in order to look competent is a badly configured agent, and that is a build decision rather than a limit of the technology.

Read next

Missed-call cost calculator

Your call volume, your answer rate and your job value, priced per month with the model shown.

Is an AI phone agent legal in Germany?

Disclosure, recording consent and retention, against the primary law rather than vendor blogs.

Trades and home services

The same leak walked end to end for a firm whose phone is in the van.

The Customer Facing Agent

What we actually build for the middle two layers of the hybrid above.

Price the leak before you buy the fix.

The calculator turns your call volume, answer rate and average job value into a monthly number, with every assumption editable.

Run the missed-call numbers

free · no account · the model is shown, not hidden