AI Receptionist: What It Is, What It Costs, and How to Choose One

An AI receptionist is software that answers your phone, holds a spoken conversation with the caller, and completes the thing the caller rang about. It answers, it understands, and it acts. Anything that only does the first two is a message taker with a better voice.

The term is about three years old and it is already being used to sell at least five different products: true conversational voice agents, human staffed answering services with an AI tier, auto attendants that got a rebrand, add-on features inside phone systems, and developer platforms that are really a toolkit rather than a product. Most of the confusion in this market comes from that one fact.

This page is the plain English version. What one actually is, how the technology works underneath, what it genuinely does well in 2026 and where it still falls over, what it costs, how to pick one, what to never let it near, and how to run a pilot that cannot damage the business. It is written for an owner or an operations manager at a company with 5 to 200 staff who is losing calls and does not want to become a telephony expert to fix it.

What an AI receptionist actually is

An AI receptionist is a voice agent that picks up an inbound call, speaks in natural language, works out what the caller wants, and either resolves it or hands it to the right person. There is no menu, no keypad, and no script the caller has to follow.

Three capabilities separate a real AI receptionist from everything else sold under the label.

• It speaks and listens in free-form conversation. The caller talks the way they would talk to a person, and can interrupt, change their mind, or ramble, without breaking the call.
• It understands intent rather than matching keywords. Someone saying my boiler is making a noise and there is water on the floor gets classified as an urgent maintenance call, not routed by whether they said the word emergency.
• It takes an action in a system you already run. It books the slot, raises the ticket, checks the balance, sends the text, transfers the call. This is the part most products skip and the part that decides whether you bought anything.

A useful test: if the outcome of every call is a message in somebody's inbox, you do not have an AI receptionist. You have an automated notepad. That can still be worth paying for, particularly if the alternative is voicemail, but it should be priced and evaluated as what it is.

Here is what a working call looks like end to end. The phone rings. The agent answers in one ring and says who it is answering for. The caller says they want to book an appointment for next Tuesday. The agent checks a live calendar, offers two real slots, takes a name and a callback number, spells the name back, writes the booking into the calendar, sends a confirmation text, and ends the call. Total elapsed time is around ninety seconds and nobody in your office was involved.

And here is the same call when the product is weaker. The agent answers, sounds convincing, takes the same details, and emails them to your front desk. On Tuesday morning somebody types the appointment in, discovers the slot is gone, and calls the customer back. The call was answered. The work was not done.

What an AI receptionist is not

Four things routinely get sold as AI receptionists and are not.

• An auto attendant. Press one for sales, press two for support. This is a phone tree, it has existed since the 1980s, and putting a synthetic voice on the front of it changes nothing about how it behaves.
• A chatbot with a phone number. Some products are a text-based bot with speech bolted on at each end. They tend to give away the difference by ignoring interruptions and by answering the question you asked two turns ago.
• Voicemail transcription. Useful, cheap, and not a conversation. Nobody is served, nothing is booked, and the caller still hung up on a recording.
• A recorded human voice on a decision tree. Some answering services use pre-recorded prompts assembled dynamically. It sounds better than text to speech and it is less flexible than either a person or a model.

None of these are bad products. They are just not the thing you are shopping for when you search for an AI receptionist, and vendors in all of these categories now use the same three words.

AI receptionist vs virtual receptionist vs answering service vs auto attendant

These are six different products and the only reliable way to tell them apart is to ask who or what is on the line and whether it can write to your systems. Marketing language will not tell you, because all five categories now describe themselves with the same vocabulary.

The single most useful question in the entire buying process is this: when the call ends, what changed in my systems? If the answer is nothing, the category above the answer does not matter.

AI receptionist

Software answers, converses, and acts. The economics are different from every other option on the list because the marginal cost of the fiftieth call in an hour is roughly the same as the first. That is the whole argument for it: capacity that does not degrade at 6pm, on a Saturday, or during the one week a year when the phone will not stop.

The trade is judgment. A human receptionist who does not know something will improvise sensibly, apologize, or find someone. A model that does not know something will often produce a fluent, plausible, wrong answer, and will do it in a confident tone. Every serious deployment is built around constraining that.

Virtual receptionist

A virtual receptionist is, in the overwhelming majority of cases, a remote human being who answers your calls under your business name from somewhere else. The category is roughly twenty years old and predates conversational AI by a long way.

They are good at the things software is bad at. Nuance, tone, an irritated caller, a complicated situation that does not fit a form. They are expensive at volume because you are buying human minutes, and they usually cannot do the work either, because they do not have logins to your systems. Most of them take a very good message.

Traditional answering service

An answering service is a call center that answers overflow and after-hours calls, captures a message, and escalates according to rules you set. It is the oldest product in the category and still the most widely used.

The honest description of its value is deferral. A call that would have hit voicemail at 9pm becomes a structured message in an inbox at 9pm. Your team still does the work in the morning, and the caller still waits. For emergency dispatch that is genuinely valuable, because the escalation is the product. For a leasing inquiry or a new patient call, the caller has usually already called somebody else by the time you ring back.

Auto attendant and IVR

An auto attendant is a menu. Press one for sales, press two for support, press three to hear these options again. IVR, interactive voice response, is the same idea with more branches and sometimes speech recognition for the menu choices.

It is not AI and never was, regardless of what the marketing on your phone platform now says. Zoom's own support documentation still defines its classic auto receptionist as answering calls with a personalized recording and routing them to a phone user, a call queue, a common area phone or voicemail. That is a phone tree, described plainly, inside a product line that also markets a separate AI receptionist.

The distinction matters commercially. If a demo shows you a menu builder with numbered options, you are looking at an auto attendant, whatever the page it was linked from was called.

Phone system AI features

Most business phone platforms now ship some form of AI answering as a feature. This is often the cheapest and fastest route to something working, because the telephony is already in place and the number does not need to move.

The constraint is scope. A phone platform's AI feature integrates well with that phone platform and with the mainstream calendars, and stops there. If the action you need is write a work order into Yardi, look up a patient in Dentrix, or create a matter in Clio, a bundled feature is usually the wrong tool. Ask specifically which third party systems it can write to, and treat a list of Zapier triggers as a partial answer rather than a full one.

Which one you actually need

Why virtual receptionist is the most confusing term in this category

A virtual receptionist is usually a person. That one sentence resolves most of the confusion buyers have when they start shopping, and almost no vendor page says it plainly.

The category was built in the 2000s by companies employing remote staff who answer calls under your business name, using a script you supply. The customer never knows the receptionist is not in your office. That is what virtual meant: not in the building, rather than not human.

You can see the ambiguity in what people search for. In a corpus of 7,478 phrases we pulled for this market, keywords containing virtual receptionist total roughly 28,300 searches a month in the US, and about 21 percent of that is people looking for the job rather than the service. Virtual receptionist positions and virtual receptionist jobs are the fourth and fifth largest phrases in the whole family. When a fifth of the demand for a term is job seekers, the term is describing employment, and employment implies a human.

Meanwhile ai receptionist, at roughly 12,100 searches a month on its own, has essentially no job-seeker contamination. Nobody applies to be an AI receptionist. The vocabulary is telling you which product is which.

Three questions that cut through it

• Is a person on the line, yes or no? Not is it human assisted, not is it human backed. Is there a human being listening while my customer talks.
• What happens between 11pm and 7am, and does the answer change the price? Human services usually charge more for out of hours coverage or route it to a smaller overnight team. Software does not care what time it is.
• What is the billing unit? Human services bill receptionist minutes or calls, because their cost is labor. Software bills flat monthly tiers, per minute of audio, or per customer. The billing unit is a reliable fingerprint of what is actually answering.

Some companies now sell both, which is fine and often sensible, but it means the brand name no longer tells you what you are buying. Read the plan, not the homepage.

The hybrid model, which is often the right answer

The arrangement that works best for most businesses between 5 and 200 staff is not either or. It is AI first with a human fallback.

The AI answers every call on the first ring. It handles the eighty percent that are structured and repetitive: bookings, status questions, intake, directions, hours, simple lookups. Anything it cannot handle, anything where the caller asks for a person, and anything on your never-automate list transfers to a human, either your own team during the day or a paid service overnight.

This is cheaper than staffing for peak, faster than a human queue, and it fails safe. It also gives you the only dataset that matters when you are deciding how far to push automation: a month of transcripts showing exactly which calls the software could not finish.

How an AI receptionist actually works

An AI receptionist is four pieces of software in a loop, plus a way to reach your other systems: telephony, speech to text, a language model, text to speech, and tool calls. There is nothing else in the box. Understanding those five is enough to evaluate any product in the category and to know which part is failing when a call goes wrong.

A single conversational turn runs the whole loop. The caller stops speaking, the system decides they have finished, finalizes the transcript, sends it to the model with the conversation so far, waits for the first tokens of a reply, starts generating audio, and plays it. If a tool call is involved, add a round trip to your calendar or your CRM in the middle.

The knowledge layer, which is where most quality comes from

The model does not know your business. Everything specific it says comes from what you gave it: a prompt describing how to behave, and a knowledge source describing your hours, services, prices, policies and locations.

Products handle this in one of three ways, and the difference shows up in accuracy. Some scrape your website and your Google Business Profile automatically, which is fast to set up and inherits every error and out of date page you have. Some ask you to fill in structured fields, which is slower and much more reliable. Some support retrieval over a document set, which scales to large policy libraries and introduces its own failure mode when the retrieved passage is the wrong one.

Whichever you get, one rule holds: the agent will confidently repeat whatever you gave it. Before launch, audit the source, not the agent.

Why the same call costs different amounts on different products

Because the four layers can be billed separately or bundled into one number, and vendors do both. Retell AI publishes an unbundled stack on its pricing page: voice infrastructure at $0.055 per minute, text to speech from $0.015 to $0.040 per minute depending on the voice provider, the language model billed separately by model, telephony at around $0.015 per minute for a US number, and optional extras such as a knowledge base at $0.005 per minute and PII removal at $0.01 per minute. Its published headline range is $0.07 to $0.31 per minute, which is a floor and a ceiling rather than a price.

Vapi goes further in the same direction, publishing $0.05 per minute for its own hosting layer and describing speech to text, the model and text to speech as at cost, or free if you bring your own API key. Bland publishes the opposite approach: $0.11 to $0.14 per minute depending on tier, with its pricing page stating that one per-minute rate covers the language model, speech to text, text to speech and telephony, with no per-token charges and no separate vendor invoices.

None of these are consumer products. They are platforms that a developer or an integrator builds on. The relevance to a business owner is that if a quote you have been given looks impossibly cheap per minute, check whether the model, the voice and the phone line are inside that number.

Why latency, interruptions and accents are the hard parts

Everything difficult about an AI receptionist happens in the gaps between words. The model deciding what to say is largely a solved problem. Deciding when to say it, over a compressed phone line, to a stranger with an accent who keeps interrupting, is not.

Latency

In ordinary human conversation the gap between one person finishing and the next starting is around 200 milliseconds. That is the number your ear is calibrated to. At half a second the pause is noticeable. At a second it reads as a bad connection and callers start saying hello, are you there. At two seconds they hang up.

Vendors publish response time figures and they cluster in the same place. Bland's site states 400ms against a claimed industry average of 1,240ms. Vapi publishes under 500ms average latency. Synthflow publishes sub-500ms for the platform. Retell AI publishes around 600ms. All four are self-reported marketing numbers with no stated measurement method, so treat them as a rough indication of what the category is aiming at, not as a specification you can hold anyone to.

What you can do is measure it yourself. Call the demo number, ask a question, and count how long the silence lasts before the reply starts. Then ask a question that requires a lookup, such as do you have anything on Thursday, and time that one too, because tool calls add a round trip and this is where products diverge most.

Interruptions and turn taking

The hardest engineering problem in voice AI is knowing when the caller has finished speaking. It is called endpointing, and there is no clean answer to it.

Set the threshold short and the agent talks over people who were only pausing to think, which is what most people do when reciting an address. Set it long and every exchange gains half a second of dead air. Neither setting handles the caller who says uh huh or right in the middle of the agent's sentence, which a person would read as encouragement and a naive system reads as an interruption to stop for.

Barge-in, the ability for a caller to cut the agent off mid-sentence and be heard, is now standard and you should test it directly. Ask the agent a question that produces a long answer, interrupt after three words, and see whether it stops cleanly, whether it actually heard what you said, and whether it picks up from where you took the conversation rather than from where it left off.

The second thing to test is the double request. Say I need to reschedule Thursday, and also can you tell me if you take Delta Dental, in one breath. Weak systems answer one part and forget the other. That single test separates products faster than any feature list.

Accents, names and numbers

Speech recognition accuracy is not uniform. It is very good on clear, unaccented, mid-sentence speech and much worse on exactly the things a receptionist call is made of: proper nouns, street names, spelled out email addresses, policy numbers, and any speaker whose accent is under-represented in the training data.

This is not a complaint about the technology, it is an operational fact you have to design around. A caller named Siobhan or Nguyen or Adeyemi will have their name mangled unless the system is set up to spell back and confirm. A caller reading out a fifteen digit member number will have at least one digit wrong.

Three mitigations actually work. Read back and confirm anything that matters, always. Prefer to send a text with a link rather than reading a long alphanumeric string in either direction. And where you can, look the caller up by the number they are calling from instead of asking them to identify themselves at all.

The test to run before you buy: get the most strongly accented person you know, plus one person with a hard to spell surname, to book an appointment and give a callback number. Read the transcript. That fifteen minute exercise will tell you more than the entire sales process.

The other conditions that break calls

• Background noise. A caller on a job site, in a car with the window down, or in a busy waiting room degrades recognition badly. Some platforms sell denoising as a paid add-on, which tells you it is a real problem.
• Poor cell connections. Dropped packets produce partial words, and a model given partial words will confidently answer the wrong question.
• Other robots. If your number is called by another company's automated system, or if the agent needs to navigate a phone menu to complete a transfer, expect failures. Two machines waiting for each other to finish speaking is a common and very silly failure mode.
• Silence. Callers who go quiet to look something up get prompted, prompted again, then hung up on. Check what the timeout behavior is and how long it gives people.
• Hold music and call waiting. Music on the line is often transcribed as speech, and the agent responds to it.

What an AI receptionist genuinely does well today

AI receptionists are good at high volume, structured, repetitive calls where the right answer already exists somewhere in your systems. That is a narrower description than the marketing, and it still covers most of what a front desk does all day.

Two of these deserve to be called out because they are the ones that consistently pay for the whole deployment.

The first is the simultaneous caller. Almost every business measures missed calls as calls that rang out, and almost none measure calls that arrived while the line was busy. If you have one person on the phones, your second caller has been getting voicemail all year and you have no record of it. Ask your phone provider for a report of calls received while all lines were occupied. The number is usually a shock.

The second is the after-hours leasing, booking or new customer inquiry. Somebody who calls at 7pm about a service they want to buy is the highest intent caller you will get all day, and they are calling from a list. If they get a recording, they call the next number on it, and you never learn the call happened.

Where AI receptionists still fail

They fail on ambiguity, on emotion, on bad data, and on anything requiring discretion. Those four categories cover nearly every bad call you will read in a transcript review.

The failure that costs the most

The expensive one is the confidently wrong answer, and it is almost never the model making things up. It is your data.

An agent told that a unit is available will say it is available. An agent pointed at a website listing a price that changed in March will quote the March price in December. An agent given a services page that does not mention that you stopped doing a thing will keep booking people in for it. In every transcript review we have run, the wrong answers trace back to a source document rather than to invention.

This has a useful implication. Deploying an AI receptionist forces you to find out where your business's authoritative answers actually live, and in a lot of companies the honest answer is in one long serving employee's head. That is worth knowing regardless of whether you buy anything.

What it cannot do at all

• Judgment. Deciding whether to waive a fee, make an exception, or believe a story. These are not hard for a person and they are not available to software.
• Reading distress accurately. It can catch explicit words. It cannot hear that somebody is frightened and being polite about it.
• Anything requiring physical presence or eyes on a situation.
• Genuine negotiation, where the answer depends on what you are prepared to concede.
• Building a relationship. Some businesses run entirely on the fact that a specific person answers the phone. Automating that is value destruction dressed as efficiency.

The integration question decides what you actually bought

Whether the agent can read from and write to the systems you already run is the single question that separates a working deployment from an expensive way of taking messages. Ask it on the first call, before the demo, because the demo is designed not to raise it.

There are three levels, and vendors across all three describe themselves the same way.

A concrete way to test it in a demo: ask the salesperson to book an appointment into your own calendar, live, on the call, and then refresh your calendar while they are still talking. If that cannot be arranged within the sales process, it is telling you something about how the integration works in production.

Real integration, Zapier, and the difference

A native API integration means the agent queries your system during the call and gets an answer in time to use it. A Zapier or Make automation means an event fires after the call and something happens a few seconds or minutes later.

Both are useful and they are not interchangeable. Post-call automation is fine for creating a CRM record, sending a confirmation, or notifying a team channel. It cannot tell a caller whether Thursday at 2pm is free, because by the time it runs the caller has hung up.

When a vendor lists thousands of integrations, that number is almost always the Zapier catalog. Ask which integrations are native and live during the call. It is usually a short list: Google Calendar, Outlook, one or two CRMs, and whichever vertical system they built for first.

The systems that matter, by industry

If you run one of these, name it in your first email to any vendor and make the answer part of the shortlist criteria.

Some AI receptionist vendors publish vertical integration lists on their own sites. My AI Front Desk, which now presents itself as Frontdesk AI, names Yardi, RealPage, AppFolio, Entrata, Buildium, SimplePractice, TherapyNotes and Jane App among its integrations. Whether a named integration does what you need is a separate question from whether it exists, so ask what fields it reads and what records it can create.

The question behind the question

If nothing integrates, you can still get value, but you should buy accordingly. A no-integration agent priced like a message taking service, replacing voicemail on an after-hours line, is a reasonable purchase. The same product priced as an AI employee is not.

And if you run a system nobody integrates with, which is common in property management and in older medical practices, your realistic options are a custom build against whatever API exists, a middleware layer, or accepting message capture for now. Find that out in week one of shopping, not in week six of implementation.

Choosing by business size: solo, 5 to 20, 20 to 200

The right AI receptionist for a one person business is a different product, bought a different way, from the right one for a company with three locations. Most bad purchases in this category are a business buying the tier above or below where it actually sits.

Solo operator and very small teams

Your problem is that you cannot answer the phone while you are doing the work, and the work is what you are paid for. Every missed call is a customer who called somebody else.

Buy a packaged product, self-serve, month to month, and set it up yourself. The published entry tiers in this category sit in the tens of dollars: Rosie publishes $49 a month for 250 included minutes, and My AI Front Desk publishes a $20 a month basic tier with voice minutes metered separately through a credit system. Those are real published numbers from their own pricing pages in August 2026, and they establish the shape of the market rather than a recommendation.

What matters at this size: does it book into your calendar, does it text the caller a confirmation, can you change the script from your phone, and can you turn it off instantly. What does not matter: enterprise reporting, single sign on, or an integration roadmap.

What to avoid: paying anybody a four figure setup fee. At this size the setup is a form.

5 to 20 staff

This is the size where the argument gets interesting, because you do have someone answering the phone and the question feels like whether to replace them. It usually is not.

At 5 to 20 people the phone is answered by somebody who also does three other jobs. They are on a call when the second call comes in. They are at lunch. They leave at five. The AI is not replacing that person, it is covering the hours and the concurrency they cannot, and taking the repetitive third of their day so they can do the rest of the job properly.

Frame the purchase that way and the requirements change. You need one genuine integration, usually the calendar or the industry system. You need a clean warm transfer during office hours so the AI can hand a live caller to a real person without them repeating themselves. You need transcripts somebody actually reads. You probably do not need a custom build.

The mistake at this size is buying on price per minute and discovering the product cannot write to the one system your business runs on. The integration constraint should pick the vendor. Price should break the tie.

20 to 200 staff

Above about twenty people this stops being a purchase and becomes a project, and the difference is not the software. It is that you now have multiple numbers, multiple locations, existing routing rules, an existing answering service contract, a compliance position, and staff who need to be told what is happening before it happens.

Things that appear at this size and not below: per location knowledge, so the agent does not quote one branch's hours to another branch's caller. Role based access to recordings. Data retention rules. A named owner who reviews transcripts weekly. A migration plan for the answering service contract, which almost certainly has a notice period.

The right sequencing is one location, one call type, four weeks, with the existing service still running behind it. Then a second call type. Then a second location. Companies that switch everything over on a Monday spend the following month unable to tell which of nine changes caused the problem they are looking at.

On cost, expect a setup or implementation component as well as a run rate, and expect the implementation to be most of the first year's spend if there is real integration work. A vendor that quotes only a monthly number for a twenty location rollout has not understood the job.

By industry: match the automation to your call mix

The case for an AI receptionist is decided by what your calls are actually made of. A business whose calls are ninety percent scheduling has a very different answer from one whose calls are ninety percent discretionary advice, and the split is visible in a week of listening.

Property management

Property management is one of the strongest fits because the calls are relentless, repetitive, and mostly not urgent to anyone except the person calling. Maintenance intake, leasing inquiries, rent and lease questions, contractor coordination.

Maintenance intake is the best first candidate in almost every portfolio, because the script barely varies. Who is calling, which unit, what has failed, how urgent it is, and whether a contractor can get in. Five fields, every time. An agent that captures all five identically and writes a work order beats a person typing up a voicemail the next morning, and it removes the most common source of bad data in a property management system.

The after-hours case is easier still, because you are not comparing the agent against your team. You are comparing it against voicemail. A tenant with no hot water at 10pm who reaches a recording calls back angry in the morning, or triggers the emergency line for something that was not an emergency.

The founder of Augment AI Studio also operates My Getaways, a short-term property management company whose inbound line is answered by an AI phone agent he built. That is where our view of this comes from: reading transcripts of real tenant and guest calls rather than watching demos.

Two hard limits in this vertical. Anything involving gas, fire, flooding, structural damage or resident safety goes to a human immediately, and the escalation path gets tested before launch. And anything touching fair housing stays factual: published availability, published pricing and published screening criteria can be read out, discretionary judgments about applicants cannot.

Dental practices

Dental has the highest ratio of automatable calls to total calls of any vertical we have looked at, and the clearest single point of loss: the new patient who calls while the front desk is already on the phone.

A new dental patient is a multi-year relationship. A practice that misses one in eight new patient calls because one person cannot answer two lines is losing more than any software costs, and the loss is invisible because nobody records the call that was never answered. That is the argument, and it does not depend on replacing anyone.

Good first automations: booking and rescheduling against the live schedule, filling cancellations, answering hours, location, parking, new patient paperwork and what to bring, and taking after-hours calls with an emergency escalation rule for dental trauma and severe pain.

The dangerous one is insurance. Callers ask do you take my insurance and expect a yes or a no. The honest answer is usually complicated, plan specific, and depends on a verification your team has not run yet. An agent that answers yes because your website lists the carrier will generate a patient who arrives expecting coverage they do not have. Restrict the agent to confirming which carriers you are in network with and booking the verification, never to confirming a benefit.

On systems, the question is whether it writes to Dentrix, Eaglesoft, Open Dental or Curve, and whether the vendor will sign a business associate agreement. Both answers should arrive before the demo.

Medical practices

Medical is the same call shape as dental with a much shorter leash. Scheduling, directions, hours, forms, and routing are safe. Anything clinical is not, and the line has to be drawn in the configuration rather than in a policy document.

Off limits without argument: symptom questions, triage, dosage, drug interactions, whether the caller should come in, and reading out results. The agent should not attempt any of these even to say something cautious, because a cautious wrong answer is still a wrong answer and it is now recorded.

The specific risk to design for is the caller who does not announce that they are having an emergency. People describe serious symptoms conversationally and apologetically. Build an explicit rule set that hears chest pain, difficulty breathing, bleeding, and self harm language and routes immediately, and test each phrase yourself before launch.

On compliance, if the agent handles protected health information you need a business associate agreement with the vendor, and you need to know whether their subprocessors, meaning the speech recognition, model and voice providers, are covered by it. Several platforms advertise HIPAA support at higher tiers. Ask for the agreement in writing, ask what data is retained and for how long, and ask where recordings are stored.

Home services: HVAC, plumbing, electrical, roofing

Home services has the highest value per missed call of the high volume verticals, and the most dangerous emergency profile. Both facts should shape the deployment.

The value case is simple. A homeowner with a failed furnace in January calls three companies. The one that answers gets the job. Most of those calls arrive outside office hours or while every technician and the office manager are on other calls, which is exactly the gap software fills.

Good first automations: capturing a quote request with address and job type, giving job status for a booked appointment, booking a standard service call, and confirming or rescheduling.

The hard no is safety. A caller reporting a gas smell, a burning smell, sparking, water coming through a ceiling, or anything involving a hazard gets a human on the line or an instruction to call the utility or emergency services, immediately, with no intake questions first. Write those trigger phrases out, test all of them, and re-test after every prompt change.

One honest warning about this vertical: it is crowded with near identical products, and a lot of them are the same underlying platform with different branding. Judge on the integration with ServiceTitan, Jobber or Housecall Pro, and on what happens on an emergency call, because on everything else the products are hard to tell apart.

Law firms

Legal has the highest value per call and the tightest constraints on what can be said, which makes intake the automation and advice the boundary.

What works: capturing a new matter inquiry with the caller's name, contact details, matter type, jurisdiction, key dates and the opposing party, then either booking a consultation or routing to the attorney who covers that practice area. Capturing the opposing party matters because it feeds your conflict check, and an intake that misses it creates work rather than saving it.

What must never happen: the agent offering an opinion on the merits, estimating a likely outcome, quoting a fee, giving a limitation period, or saying anything that could read as accepting representation. Those are unauthorized practice of law risks and they are not worth any efficiency gain. The script should state plainly that the assistant is taking details and an attorney will advise.

Two practical notes. Confidentiality: decide before launch where transcripts live, who can read them, and how a potential conflict is handled if the caller turns out to be adverse to an existing client. And speed: legal intake is a race, so the value is in the agent answering at 8pm and booking a consultation for 9am, not in saving a receptionist ten minutes.

Professional services and everything else

Accountancies, agencies, insurance brokers, consultancies and B2B service firms have low call volume and high call value, which inverts the usual argument.

You are not automating to absorb volume. You are automating so that the one important call that arrives while everyone is in a meeting reaches a human being quickly, with context, instead of reaching voicemail. The valuable functions here are identification, qualification and routing, plus a booked callback with a real time on it.

Be careful about tone. In relationship businesses the phone being answered by a person is part of the product, and there are clients who will read an AI answering as a signal about how much you value them. The safe pattern is the AI answering only when a human genuinely cannot, and saying so plainly.

If your total inbound volume is under fifty calls a month, read the section further down on when an AI receptionist is the wrong answer before you spend anything.

The pricing models, and what each one hides

There are eight ways AI receptionists are priced and they are not comparable to each other without doing arithmetic on your own call data. A per minute rate and a per conversation rate can differ by ten times on the same volume, in either direction, depending on how long your calls run.

Every figure in the table below was read from the vendor's own published pricing page in August 2026. They are examples of each model, not recommendations, and prices in this category change often. Check before you buy.

GoTo Connect also markets an AI receptionist feature. We were not able to load a GoTo pricing page to verify a figure, so we are not publishing one.

Reading a per minute quote properly

Two things make per minute pricing misleading. The first is rounding. RingCentral publishes that call time is rounded up and billed in 30 second increments, which is normal and which means a hundred 40 second calls cost you the same as a hundred 60 second calls. Ask about the increment.

The second is what a minute contains. On an unbundled developer platform the advertised rate may cover only the orchestration layer, with the model, the voice, the transcription and the phone line each billed on top. That is a transparent way to price a toolkit, and it is not a number you can compare against a packaged product's monthly fee.

What an AI receptionist costs in practice

For a small business, a working AI receptionist costs between $50 and $500 a month. For a mid-sized business with a real integration it costs between $500 and $3,000 a month plus a setup or build cost. Those ranges hold across almost every product we have looked at, and the variable that moves you between them is integration depth, not call volume.

The monthly figures in that table are our own synthesis from published vendor pricing plus what implementation work of this kind costs. They are ranges to sanity check a quote against, not a price list.

The costs that do not appear in the quote

• Telephony. Phone numbers, inbound minutes and porting. Small, but never zero, and sometimes billed by a different company. Retell publishes $2.00 a month per phone number as an indication of the scale.
• The integration build. If your system is not on the vendor's native list, somebody is writing code. This is usually the largest single line item in a mid-market deployment.
• Your own time. Writing the knowledge base, deciding the escalation rules, and testing them. Budget a full day for a simple deployment and several days for a complex one.
• Transcript review. Someone has to read calls every week, at least at first. This is a real recurring cost and skipping it is how deployments quietly degrade.
• Add-ons that are priced separately. Denoising, PII redaction, extra concurrency, extra knowledge bases and quality assurance features are all separately metered on some platforms.
• The cost of a bad call. Not on any invoice, and the reason the never-automate list exists.

The arithmetic to run before you talk to anyone

Four numbers decide whether this is worth doing, and you can get three of them from your phone provider this week.

• Total inbound calls per month.
• Calls that went unanswered, including calls that arrived while your lines were busy. Ask specifically for the busy count, because most reports hide it.
• The share of calls arriving outside your office hours.
• The value of a new customer, and roughly what proportion of new callers become customers.

Then the calculation. Suppose you take 600 calls a month and 12 percent go unanswered. That is 72 lost calls. If one in five of those was a new customer inquiry, and one in three of those inquiries would have converted, you are losing about 5 customers a month. At an average customer value of $600 that is $3,000 a month against a product costing a few hundred.

Those percentages are illustrative and yours will be different. The point of writing them down is that the arithmetic is usually either obviously yes or obviously no, and the businesses that get burned are the ones that never ran it. If your numbers do not support it, do not buy it.

What a human virtual receptionist costs, for comparison

Human virtual receptionist services in the US publish rates that cluster between $1.75 and $2.60 per receptionist minute, and every one of them steps the rate down as volume rises. That is the number an AI receptionist is competing against, and it is the cleanest way to understand why the category exists.

All figures below were read from each company's own published pricing in August 2026. AnswerConnect's figures come from the pricing PDF published on its own site, because its pricing page is behind a form.

Two of these are worth reading twice. Smith.ai publishes both products side by side, and the AI receptionist runs at $1.67 to $3.00 per call against $8.50 to $11.50 per call for its live agents. Abby Connect goes further and prices the gap inside a single currency, publishing that an AI minute consumes half the allowance of a human minute. When two companies that sell both models price them that far apart, they are telling you what the cost difference actually is.

The same volume, priced both ways

Normalizing to a monthly minute figure makes the choice concrete. These are published list prices for the nearest matching plan, in August 2026.

The ratio widens as volume grows, which is the whole economic story of the category. Human services have a labor cost that scales linearly and can only be discounted so far. Software does not.

What the table does not capture is quality, and that is not a rhetorical hedge. A US based human receptionist handling a distressed caller, an unusual request or an angry customer will outperform any current AI on that call by a wide margin. The correct conclusion is not that software is cheaper so buy software. It is that paying $2 a minute for a human to read out your opening hours is a poor use of $2, and paying it to handle your hardest calls may well be a good one.

Four ways to buy one, and how to pick a route

There are four routes to a working AI receptionist and they suit very different businesses. Picking the route before you pick a vendor removes most of the confusion, because the shortlists barely overlap.

Packaged products

These are the products most people should look at first. Goodcall, Rosie and My AI Front Desk are three that publish their pricing openly, which is itself a useful filter, and all three are configured through a browser rather than through a sales process.

The strength is that you can be live the same day and cancel next month. The limitation is that you get the integrations they built, the conversation design they chose, and the failure handling they decided on. If your business is a normal shape, that is fine. If it is not, you will hit the edges quickly.

Phone platform features

If you are already on RingCentral, Nextiva, Dialpad, Zoom or Quo, look at what they offer before you look anywhere else. The number does not have to move, the billing is already in place, and the transfer to a human works because it is the same system.

Be careful about names inside these product lines. Several platforms sell both a classic auto attendant and an AI receptionist, and the two are described in similar language on the marketing pages. Read the support documentation, not the pricing page.

Human services with an AI tier

Smith.ai, Posh and Abby Connect all sell AI and human answering, and Ruby includes AI features inside a human led service. If your main worry is that the AI will mishandle an important call, this route puts a person behind it from the first day, at a published price, with no build.

The trade is that these are answering services first. They are excellent at capturing and escalating, and they are generally not the route to an agent that writes work orders into your property management system.

Developer platforms and custom builds

Retell AI, Vapi, Bland and Synthflow are platforms, not products. They give a developer the pieces: telephony, speech, model orchestration, function calling, and a way to deploy. Somebody still has to design the conversation, write the integrations, build the escalation logic and maintain it.

This route is correct when your systems are non-standard, when the workflow you need is specific to how you operate, or when the value is in an integration nobody has built. It is the wrong route if your requirement is answer the phone after five and book people in, because you will spend $10,000 recreating a $99 product.

The honest test: write down the one thing you need that a packaged product cannot do. If you cannot fill that sentence in, do not commission a build.

Sixteen questions to ask before you sign

Ask these in order, in writing, before the demo. A vendor's willingness to answer the awkward ones is more informative than the answers.

• 1. Is a human being ever on the line, and under what circumstances? If the answer is only if you pay for it, find out what that costs per call.
• 2. Which of my systems can it read from during a call, and which can it write to? Ask about the specific system by name, not the category.
• 3. Which integrations are native and live during the call, and which run afterwards through Zapier or a webhook? These are different products.
• 4. What happens when it does not know the answer? The good answer is that it says so and transfers or takes a message. The bad answer is a description of how rarely that happens.
• 5. What happens when a caller asks for a person? It should transfer on the first request, without an attempt to talk them out of it.
• 6. Can I read every raw transcript and listen to every recording, and can I export them? If transcripts are summarized, aggregated, or only available on a higher tier, that is a serious mark against.
• 7. How is the knowledge base built and how do I update it? Ask how long a change takes to go live and whether you can make it yourself.
• 8. What is the measured response latency on a call that requires a lookup? Then test it yourself rather than accepting the number.
• 9. How does it handle interruptions, and can I hear a recording of a call where the caller interrupted it?
• 10. What is the escalation path for an emergency, how is it triggered, and can I test it before launch?
• 11. What exactly is the billing unit, and what counts as one of them? For per-conversation pricing, ask whether a wrong number counts. For per-minute, ask the rounding increment.
• 12. What is the overage rate? A surprising number of published price pages omit it.
• 13. Where is call data stored, how long is it retained, who has access, and can I delete it? For healthcare, ask for the business associate agreement in writing and ask whether it covers subprocessors.
• 14. What is the contract length and the notice period? Month to month is common in this market, so a twelve month lock is worth questioning.
• 15. If I leave, what do I take with me? Recordings, transcripts, the knowledge base, the phone number. Ask about porting specifically.
• 16. Can I have a reference customer of roughly my size, in roughly my industry, who has been live for at least six months?

One more, which is not a question but a test. Call the vendor's own main line during business hours and see what answers. Companies selling AI receptionists that route their own inbound calls to voicemail are telling you something.

What to never let an AI receptionist handle

Write this list before you launch, configure it as hard rules rather than as guidance in a prompt, and test every line on it yourself. A deployment is defined as much by what it refuses to handle as by what it handles.

The common thread is discretion. Anywhere the right answer depends on judgment, on circumstances, or on what your business is prepared to do in this particular case, a language model will produce something fluent and it will not be your decision.

Disclosure, recording and compliance in the US

Three legal questions come up on every deployment: do you have to tell callers it is AI, can you record the call, and what happens to the data. None of what follows is legal advice, and every one of these should be checked with your own counsel before launch, but knowing the shape of the questions makes that a much shorter conversation.

Do you have to disclose that it is AI?

For inbound calls there is no single federal rule requiring it, and you should do it anyway. Disclosure costs you nothing, it prevents the worst kind of complaint, and in practice it improves calls: people who know they are talking to a machine speak more clearly and ask simpler questions.

A growing number of states have passed or proposed bot disclosure requirements, and the rules vary in scope and in which industries they cover. Rather than track a shifting set of state statutes, most businesses take the safe position of disclosing in the first sentence and moving on. Something like this is enough: you have reached the automated assistant for X, I can book appointments and answer questions, and I can put you through to someone if you need.

Outbound is a different legal universe. In February 2024 the FCC confirmed that AI generated voices count as artificial voices under the Telephone Consumer Protection Act, which brings AI outbound calling under the consent rules for artificial and prerecorded voice calls. TCPA penalties run from $500 to $1,500 per call with no aggregate cap. If anyone proposes that your AI receptionist also makes outbound calls, that is a conversation with your lawyer before it is a conversation with a vendor.

Recording

Federal law permits recording with the consent of one party to the call. A number of states, including California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania and Washington, require all parties to consent. If you take calls from anywhere in the country, the practical answer is to give a recording notice on every call regardless of where the caller is.

Most AI receptionist products record and transcribe by default, because that is how they work. That means you have made a recording decision whether or not you thought about it. Check the default, check whether it can be turned off per call, and put the notice in the opening line alongside the AI disclosure.

Health data and other regulated categories

If the agent touches protected health information you need a business associate agreement with the vendor, and you need to know whether it covers their subprocessors. An AI receptionist typically involves at least three of them: the speech recognition provider, the language model provider and the voice provider. An agreement that covers the vendor but not the model provider is not the protection you think it is.

Ask four specific questions. Will you sign a business associate agreement. Which subprocessors handle audio or transcripts, and are they covered. Is call data used to train models, and can that be turned off. What is the retention period and can I set it. Some platforms sell PII redaction as a separately metered feature, which is worth knowing before you price the deployment.

The same shape of question applies to financial services, insurance and anything under a state privacy statute. The answers differ, the questions do not.

How to run a pilot without risking the business

Point the agent at one call type, on a separate number, for four weeks, with the old arrangement still running behind it. That single sentence prevents most of the ways this goes wrong.

The temptation is always to switch the main number over on a Monday and see what happens. Resist it. Not because the software will fail, but because if anything does go wrong you will have changed nine things at once and will not be able to tell which one caused it.

Before you start

• Pick one call type. After hours is almost always the right first choice, because the thing you are replacing is voicemail and the bar could not be lower.
• Write the never-automate list, configure it as hard rules, and test every trigger phrase yourself out loud on a real call.
• Define the required outcome for the call type you chose. A booking in the calendar, a ticket in the system, a structured message with named fields. If you cannot state the outcome, you cannot tell whether the pilot worked.
• Set up the human fallback and call it. Know exactly who the transfer reaches at 2pm and at 2am, and what happens if they do not pick up.
• Tell your team before it goes live, and tell them what to do when a customer mentions it. Staff finding out from a customer is a bad start.
• Record your baseline: calls per month, unanswered calls, busy signals, after hours volume. You cannot show an improvement against a number you never took.

The four weeks

Thirty calls a day generates roughly twenty minutes of transcript reading. That is the entire cost of running this properly, and it is the part people skip.

Rules that stop a pilot becoming an incident

• Never route your main number to an untested agent. Use a separate number, or forward only outside office hours.
• Always keep a working path to a human, and test it weekly rather than once.
• Give one named person ownership. Deployments without an owner degrade in about six weeks, because knowledge goes stale and nobody notices.
• Change one thing at a time, and write down what you changed and when. Prompt changes have side effects on calls you were not thinking about.
• Re-test the emergency triggers after every prompt change, without exception.
• Keep the exit available. Month to month is normal in this market. Do not sign a year for a pilot.

How to tell whether it is actually working

Read the transcripts. Every vendor dashboard in this category reports metrics that look like performance and are not, and the gap between the two is where bad deployments hide.

Containment rate deserves particular suspicion, because it is the metric vendors optimize for and it is directly opposed to your interests on a hard call. A high containment rate can mean the agent resolved everything, or it can mean the agent would not let anyone through to a person. Only the transcripts distinguish the two.

The six things to look for in a transcript review

• Calls where the agent gave a specific answer that was wrong. Prices, availability, hours, coverage and dates are the usual culprits, and the fix is nearly always the source data rather than the agent.
• Calls where the caller asked for a person and did not get one within two turns.
• Calls that ended with no outcome. No booking, no ticket, no callback promised, no message.
• Calls under about twenty seconds. Almost all of these are hang-ups, and a rising share of them is the earliest warning sign you will get.
• Calls where the agent asked the same question three times. That is speech recognition failing on one field, and it needs a transfer rule rather than a better prompt.
• Repeat callers within seven days. The same number calling three times in a week nearly always means the first call failed.

The two numbers worth tracking every month

Outcomes per hundred calls, and missed calls.

Outcomes per hundred calls means completed bookings, tickets raised or qualified leads captured, divided by total calls answered, times one hundred. It is the only metric that measures the thing you bought. Track it monthly and it will tell you when a knowledge base has gone stale, because it falls before anything else does.

Missed calls means calls that rang out, hit voicemail, or arrived while every line was busy. Compare it to your pre-launch baseline. If missed calls have not fallen, the deployment has not done the one job it exists to do, regardless of what any other number says.

Everything else is diagnostic. Those two are the scoreboard.

Will your callers mind talking to an AI?

Most callers do not mind, provided they get what they called for quickly and can reach a person when they want one. Those two conditions carry almost all of the goodwill, and violating either of them produces the complaints.

Some of the fear here is misplaced. Your callers have spent twenty years being routed through phone trees, held in queues, and told their call is important to us. The bar they are actually comparing you against is not a warm human conversation. It is a menu and a hold tone, or a voicemail nobody returns.

What genuinely annoys people, in the transcripts we have read, is a short list.

• Being trapped. Asking for a person and being redirected back into the conversation. This is the single biggest source of anger and it is entirely a configuration choice.
• Repeating themselves. Giving their details to the agent and then having to give them again to the human they are transferred to. If the handoff does not carry context, do not call it a warm transfer.
• Being deceived. An agent that pretends to be a named human being, or that dodges the question when asked, is a much bigger problem than one that says up front what it is.
• Long pauses. Callers assume the line has dropped and hang up. This reads as rudeness rather than as a technical issue.

The disclosure question is worth settling on the side of honesty. Say what it is in the first sentence. The people who care will ask within ten seconds anyway, and being caught evading the question is far worse than the disclosure ever was.

One caveat on tone. In relationship-led businesses, where clients pay partly for access to a specific person, an AI answering the main line reads as a downgrade in service no matter how good it is. In those businesses the AI should only pick up when a human genuinely cannot, and it should say so.

When an AI receptionist is the wrong answer

There are six situations where the honest recommendation is do not buy one, and a vendor will not tell you about any of them.

The last one is worth expanding, because it is the most common way this goes badly. Businesses that remove the human before they have read a month of transcripts discover the failure modes through customer complaints instead of through a review process. The sequence that works is add capacity, measure it, then decide about staffing. The sequence that does not is remove capacity and hope.

How Augment AI Studio approaches this

We start with your call data, and we will tell you when the answer is no.

Kevin Musprett, who runs Augment AI Studio, also operates My Getaways, a short-term property management company whose inbound line is answered by an AI phone agent he built. That is where this page comes from: reading transcripts of real calls from real customers, finding out which ones the agent got wrong, and fixing them. Not from watching demos.

The method is the same one described above, because it is the one that survives contact with production. Work out what the calls are actually costing before building anything. Start with one call type where the alternative is voicemail. Integrate properly, so that a call ends with work completed rather than a message in an inbox. Write the never-automate list first and test every item on it. Then read the transcripts, every week, and keep fixing the source data that the wrong answers keep tracing back to.

If your numbers do not support it, we will say so. There are businesses on this page that should not buy an AI receptionist from us or from anyone else, and the section above says which ones.

Frequently asked questions