AI receptionist for small business: costs and limits

I come at this as the person who reads the invoice. A small business owner hears "AI receptionist" and pictures a flat fee that replaces a salary. The pricing pages tell a different story.

Some products bill per minute, some per unique caller, some per call, and a few bill for silence. This guide translates those meters into a monthly number you can compare, and it lists the ways these systems fail so you can price the risk as well.

One boundary first. Communicate is a chat-based platform with an AI agent and a shared inbox. It does not answer phone calls, so this article is a buyer's guide to other vendors' phone products, not a pitch for ours.

Where your customers write instead of call, I will say what we do. For a broader look at voice, see our voice AI customer support guide.

What an AI receptionist actually does

A receptionist role has a few core jobs. Answer the phone in a consistent way, work out why the person is calling, take a message or book a time, route urgent matters to a human, and keep a record. An AI version does these through speech recognition, a language model that holds the conversation, and a synthetic voice, connected to your calendar and phone number.

It can handle predictable calls well. Opening hours, location, services, prices you have written down, appointment booking against an open calendar, and message taking all fit the pattern. It struggles with calls that depend on judgment, emotion, or information you have not given it.

There are two broad ways to buy. One is a finished product built for small businesses, where you answer setup questions and the vendor handles the rest. The other is a voice-agent platform that sells the components, such as speech, language model, and telephony, and expects you or a developer to build the flow.

The second looks cheaper per minute and costs more in time.

Price card: what vendors publish in October 2026

The figures below come from vendor pricing pages I read in October 2026. Prices change often, pages are sometimes ambiguous, and I flag where. Check the live page before you budget.

Retell publishes pay-as-you-go pricing for voice agents. Its pricing page lists AI voice agents at $0.07 to $0.31 per minute, with a default estimate of $0.11 per minute made of a voice engine at $0.055, a language model at $0.04, and speech synthesis at $0.015. Telephony adds $0.015 per minute on listed Twilio and Telnyx countries.

New accounts get $10 in credits.

Retell also lists add-ons per minute, including a knowledge base at $0.005, safety guardrails at $0.005, and personal data removal at $0.01. Calls are billed to the second and only connected calls count, but the page states that silence and hold time are billed. It includes 20 concurrent calls and charges $8 per month for each extra concurrent call.

Phone numbers are listed at $2 per month, and the page lists verified numbers inconsistently, so confirm that figure.

Vapi's pricing page breaks a call into components. It lists a $0.05 per minute hosting fee, a transcriber at roughly $0.0095 to $0.0099, a language model at $0.0077 to $0.0452, and a voice at $0.0146 to $0.0238. Its own sum is about $0.082 to $0.129 per minute, and a 1,000-minute estimate of $82 to $129.

That excludes carrier charges, which the page lists, for example, at $0.008 per minute for inbound calls on Twilio and $0.0055 on Telnyx, and it excludes any package fee.

Goodcall sells a finished product with flat plans. Its pricing page lists Starter at $79 a month for 100 unique customers, Growth at $129 for 250, and Scale at $249 for 500, with $0.50 for each customer above the cap. It states unlimited minutes and tokens on paid plans, and it shows annual prices at about 15% off.

The meter here is unique callers, not minutes, which changes how you should model it.

For a human-staffed comparison, I read the plans on Smith.ai's pricing page. The page lists a $300 monthly plan with 30 calls, $810 for 90, and $2,100 for 300, with overage between $8.50 and $11.50 per call, and a 10% discount for annual terms. The page describes the service as professional call answering by real people.

It lists no AI-first plan prices on the page I read, so I am not comparing an AI plan from that vendor.

Turning the meters into a monthly bill

Comparing a per-minute rate with a per-caller plan needs a common scenario. Take a small clinic, shop, or trades business that gets 300 calls a month averaging three minutes. That is 900 minutes.

This is an illustration with round numbers I chose, not a measured customer.

At Retell's default estimate of $0.11 per minute, 900 minutes costs about $99, plus roughly $2 for a phone number. At the page's published range, the same minutes cost $63 at the low end and $279 at the high end. Add telephony if your estimate excludes it, and add the per-minute add-ons you turn on.

A realistic bill sits somewhere in that spread, and the spread is the lesson.

At Vapi's summed components of $0.082 to $0.129 per minute, 900 minutes costs about $74 to $116. Carrier charges come on top, and at $0.008 per minute for inbound Twilio calls, that adds about $7. You also pay for whoever builds and maintains the call flow, which no pricing page includes.

At Goodcall, the question is how many different people call. If 300 calls come from 200 unique customers, the Growth plan covers it at $129. If they come from 120, the Starter plan at $79 would cover 100 and add $0.50 for each of the other 20, so about $89.

If the same 60 customers call five times each, Starter covers the whole month. Flat plans reward repeat callers and punish spread-out ones.

At a human-staffed service charging about $10 per call over a small plan, 300 calls cost on the order of $2,100 on the Pro plan, since 300 calls are included there. That is roughly $7 per call. Humans cost more per call and handle judgment calls better, which is the trade you are pricing.

Two caveats keep this honest. The per-minute vendors bill for time connected, including silence and hold in at least one case, so a call that sits quiet while a caller searches for a policy number costs the same as talking. And these products need your time.

Set aside hours for prompts, test calls, and tuning, and put a value on those hours.

For a broader cost model that includes chat, email, and staffing, see our AI customer support cost guide and the cost per ticket benchmark.

The hidden costs that never show up on a pricing page

• Setup and testing. Someone has to write the greeting, the routing rules, the answers, and test a dozen call scenarios.
• Calendar and CRM integration. Booking requires a clean calendar with clear availability, and some connections cost extra.
• Missed-call recovery. Every call the system fumbles still needs a human callback, which has a labor cost.
• Upgrades. Concurrency limits, extra numbers, and add-on features raise the bill as you grow.
• Review time. Someone should read transcripts or summaries weekly, especially in the first month.
• Switching costs. Moving numbers, prompts, and call history to another vendor is rarely free.

A fair comparison with a human receptionist includes these. A part-time employee also needs training and review, but a person adapts without a new prompt. The sticker price understates the total on both sides, so compare totals.

What it handles well, and what to route to a person

Sort your calls before you sort vendors. Pull a week of voicemails and call logs, or ask whoever answers the phone to tally reasons for a week. Most small businesses find that a handful of reasons cover the bulk of calls.

The practical design principle is to write the exit before the answer. For every call type on the right-hand side, decide who gets the call and how fast, then configure the system to transfer or take a detailed message. A receptionist that knows its limits is worth more than one that tries everything.

Our guides on human handoff and escalation workflows apply to phone calls as much as chat. The question is the same. What does the system hand over, to whom, and with what context?

Where AI receptionists fail

Failure is predictable, and most of it is avoidable if you plan for it. I am describing categories of failure in general, not reporting a specific incident, so you can test for each before you go live.

• Mishearing names, numbers, and addresses. Speech recognition struggles with unusual names, spelled-out emails, and phone numbers given quickly. Always read back anything the business will act on.
• Talking over the caller. Interruption handling varies. Test with a caller who pauses and restarts.
• Confident wrong answers. A language model can invent a price or a policy that was never in its instructions. Limit it to the facts you provide, and test with questions you did not cover.
• Booking errors. Double bookings, wrong time zones, and appointments outside working hours happen when the calendar rules are loose.
• Dead ends. A caller who says "let me speak to someone" and gets looped has a bad experience. Make human transfer easy and fast.
• Noisy lines. Background noise and poor mobile reception cut accuracy.
• Accents and languages. Performance differs across accents and languages, and the pricing page rarely says how.
• Silent failure. Calls that drop or a calendar link that breaks may not alert anyone until a customer complains.

Test these on purpose. Before launch, place twenty calls yourself, with different voices, from a noisy place, from a mobile, with a request the system should refuse and a request that needs a human. Fix what breaks and repeat.

The broader pattern is described in our write-up of AI chatbot failures and the guide on when not to use AI support. The lesson transfers. Keep AI away from decisions that need judgment, and give customers an easy way to reach a person.

Call recording, consent, and AI voice rules

Answering calls with software brings legal questions that small businesses often miss. I am not a lawyer, and the rules differ by place, so treat this as a list of things to check.

Recording and transcription come first. Many receptionist products record and transcribe calls to produce summaries. Laws on recording vary by country and, in the United States, by state.

Some require consent from everyone on the call. A short notice at the start of the call, such as that the call may be recorded and answered by an automated assistant, is a common practice. Ask a local lawyer for the right wording.

The second area is AI voices in telephone marketing. In February 2024 the FCC adopted a declaratory ruling in CG Docket 23-362, announced on February 8, 2024, which it described as recognizing calls made with AI-generated voices as "artificial" under the Telephone Consumer Protection Act. The FCC announcement says such calls need prior express consent of the called party unless an emergency or exemption applies.

Be careful about how far that goes. The ruling concerns calls made with AI-generated voices under a statute aimed at robocalls and unsolicited calls. An AI that answers calls customers place to you is a different situation from an AI that dials people.

If your plan includes outbound reminders, follow-ups, or marketing, that is where the consent rules bite, and you should take legal advice.

Disclosure is a third area. Several jurisdictions are moving toward requiring that people are told when they are talking to an AI, and the EU AI Act includes a transparency rule for systems that interact directly with people. Our AI disclosure guide covers the reasoning.

Telling callers up front is simple, and it avoids surprises.

Finally, treat the recordings as data you hold. Know how long the vendor keeps them, where they are stored, and who can listen. If callers describe health issues or give card numbers, ask how the system handles them.

A setup checklist for the first week

• Write down the ten reasons customers call, in order of frequency, with the approved answer for each.
• List the calls the AI must not handle, and the person or number each one goes to.
• Set business hours, after-hours rules, and holiday closures in one place.
• Clean the calendar. Block lunches, set buffer times, and define each appointment type and length.
• Write the greeting. Name the business, say that an automated assistant is answering, and give callers a way to reach a person.
• Decide how you will be notified of messages and urgent calls, and who responds by when.
• Place at least twenty test calls covering normal, noisy, confused, and angry callers.
• Read every transcript for the first two weeks. Fix the instructions after each batch.
• Set a monthly review of cost, missed calls, and complaints.

A small business can complete the first seven items in a day or two with a finished product. A platform build takes longer. Whichever route you pick, the review habit matters more than the launch.

How to decide: a short framework

Use three questions in order. They save more money than comparing rate cards.

First, what does a missed call cost you? If a missed call is a lost sale worth hundreds of dollars, such as a plumber's emergency job, after-hours coverage pays for itself quickly. If most calls are routine questions, the value is lower, and a good website and chat may do the same job.

Second, how many calls and how long? Under a few dozen calls a month, a message-taking service or voicemail with transcription may be enough. Hundreds of calls a month justify a serious look at any of the options above.

Third, how much judgment do the calls need? The more calls that depend on judgment, empathy, or expertise, the more a human service or a human receptionist earns its price. The more calls that follow a script, the more AI makes sense.

If you want a return-on-investment frame for the decision, our AI support ROI guide sets out a simple model, and the AI customer support pricing guide compares how vendors charge.

If your customers write instead of call

Not every small business needs a phone product. Many customers now start with a website chat, a text, or a message on a social app, and for those customers, answering the phone faster does not help.

Communicate covers that side. You add an embeddable chat widget to your site, connect your data sources, and the AI agent answers questions around the clock. When it cannot help, the conversation goes to your team in the shared inbox.

It works for chat and messaging only, so a business that lives on phone calls still needs a phone product. Many businesses use both, one for callers and one for web visitors.

For a look at channels beyond the phone, see our guides to SMS support, WhatsApp support, and 24/7 customer support with AI.

Three small businesses, three different answers

The cases below are composites I built to show how the framework plays out. They are not real customers, and the numbers are round.

The first is a two-person trades business whose owners are on job sites all day. Most calls are new enquiries and urgent repairs. A missed call is often a lost job.

After-hours and on-site coverage has clear value, and the calls are short and fairly scripted: name, address, problem, urgency. A message-taking AI that collects those four details and flags emergencies for immediate transfer fits well. The risk is mishearing the address, so read-back is essential.

The second is a small dental or physiotherapy practice with an existing front desk. Calls cover bookings, rescheduling, and questions about insurance and treatment. Here the AI can handle booking and rescheduling after hours and during lunch, and it should hand anything clinical or sensitive to staff.

The consent and data questions matter more, because callers may describe health details. Ask the vendor how transcripts are stored before you pick.

The third is an online shop with a phone number nobody calls much. Most contacts arrive through the website and messages. A phone receptionist would add cost for a handful of calls, and the better investment is chat on the site and a clear help page.

This is the case where I would not buy a phone product at all.

Notice that the right answer changes by business, not by vendor ranking. That is why I give no ranked list here. A ranking implies one winner, and the price card shows that the meters do not even measure the same thing.

Questions to ask in a vendor demo

• Show me the bill for 300 calls averaging three minutes, with every add-on I would need.
• What counts as billable time, and is silence or hold time included?
• What happens when the system cannot answer? Show me the transfer and the message-taking path.
• How do I see transcripts and summaries, and who else can see them?
• Where are recordings stored, for how long, and can I delete them?
• How does it handle callers with strong accents, other languages, or a bad line?
• How do I change the instructions, and how fast does a change go live?
• What does cancelling look like, and can I export my data and my number?
• What do you do about recording consent and AI disclosure in the greeting?

A vendor that answers these plainly is easier to trust. A vendor that deflects to marketing language will deflect again when something breaks.

Running the first month

Treat the first month as a test, not a launch. Keep your normal answering arrangement as a backstop for two weeks if you can, so the cost of a failure stays small.

In week one, review every call. Read each transcript or summary, note what went right and wrong, and fix the instructions that day. In week two, review a sample, but read every call that ended in a transfer or a complaint.

In week three, compare the numbers. How many calls did the system answer fully, how many went to a person, and how many callers hung up early?

In week four, decide. If the system handles most routine calls and the failures are rare and recoverable, keep it and move to monthly reviews. If failures cluster in one call type, remove that type from the AI and route it to a person.

If failures are broad, pause and reconsider the product or the route.

Pair the decision with a look at what you learned from the calls. Repeated questions suggest content missing from your website, and the live chat versus AI agent comparison can help decide where to put the fix. Guardrails such as limits on what the AI may promise are covered in our AI agent guardrails guide.

Cheaper alternatives to check first

Before you pay for any receptionist product, check whether a simpler fix covers your case. Many small businesses reach for AI when the real gap is a missing process.

• Voicemail with transcription. Callers leave a message, you read it by text or email. Cost is low and risk is low, though callers sometimes hang up.
• Call forwarding on a schedule. Route calls to a mobile during busy periods and to voicemail otherwise.
• A website that answers the top ten questions. If callers ask for opening hours or prices, put them where search and chat can find them.
• An online booking link. Many callers want to book, and a link in your voicemail greeting and text reply can move them to self-service.
• A chat agent on the website. For web visitors, an AI agent can answer the same questions without telephony.

Each of these costs less than a receptionist product and may remove a third of your calls. Do them first, then measure what remains. If the remaining calls are still costing you work or sales, the case for a receptionist product is stronger and your call data is cleaner.

A staffing comparison you can do on paper

To compare AI with a person, use your own numbers, not a generic salary. Take the hourly cost of whoever answers the phone now, including benefits and the time lost to interruptions. Multiply by the hours per month they spend on calls, then add the cost of calls missed outside those hours.

Suppose, as a round illustration, that you spend 25 hours a month on routine calls and the true cost of your time is $40 an hour. That is $1,000 a month in time. If an AI product at about $100 to $130 a month takes over most of those calls, the saving looks large.

The check is what happens to the calls it cannot handle. If you spend ten hours a month fixing its errors and returning messages, the saving shrinks to roughly $500. That is still positive, and it is an honest number.

Change one input and the answer changes. If your own time is worth more than $40, the case improves. If most calls need judgment, the saving may vanish.

Do this arithmetic with your real figures, and keep the reviewing hours in it.

Conclusion

An AI receptionist can earn its keep on routine calls and after-hours coverage, and it can embarrass you on anything that needs judgment. Price it by your own call volume, not by the headline rate. Plan the exits to humans before you plan the answers, check recording and disclosure rules, and read transcripts for the first month.

If your customers mostly write to you, see Communicate pricing and start with chat.

Frequently asked questions

What is an AI receptionist?

It is software that answers a business phone line, understands why the caller is calling, answers common questions, books appointments, takes messages, and routes urgent calls to a person. It combines speech recognition, a language model, and a synthetic voice, connected to your calendar.

How much does an AI receptionist cost?

It depends on the meter. As of October 2026, Retell lists $0.07 to $0.31 per minute, Vapi's components sum to about $0.082 to $0.129 per minute before carrier fees, and Goodcall lists $79 to $249 a month. Model your own call volume.

Is per-minute or flat pricing cheaper?

It depends on your calls. Per-minute pricing favors few, short calls. Per-caller flat plans favor repeat callers.

Use your real call count, average length, and number of unique callers, then compute each option. The cheapest rate card is not always the cheapest bill.

Do I pay for silence and hold time?

Sometimes. Retell's pricing page states that silence and hold time are billed, and that calls are tracked to the second and only connected calls are billed. Check each vendor's billing notes, since hold time can add up on calls where the caller searches for details.

How does it compare to a human answering service?

A human-staffed service costs more per call, but handles judgment and emotion. On the plans I read in October 2026, one vendor listed $300 a month for 30 calls and $2,100 for 300, with overage of $8.50 to $11.50 per call. AI is cheaper on routine calls.

Can it book appointments?

Yes, if your calendar is clean and the rules are clear. Define appointment types, durations, buffers, and working hours. Test for double bookings and time zone mistakes.

Always have the system read back the date and time it booked.

What happens when it cannot answer?

That depends on how you set it up. Good configurations transfer to a named person, take a detailed message, or offer a callback. Test these paths before launch, because a dead end is the failure that annoys callers most.

Does it work for emergencies?

Be cautious. For urgent or safety-related calls, configure an immediate transfer to a human or a clear instruction to contact emergency services. Do not rely on the AI to judge severity.

Test the emergency path more than any other.

Do I need to tell callers it is an AI?

It is a good practice and in some places a legal expectation. The EU AI Act includes a transparency rule for systems that interact directly with people. A short greeting that names an automated assistant is simple and honest.

Ask a lawyer for your jurisdiction.

Is call recording legal?

It depends on the country and, in the United States, the state. Some places need consent from everyone on the call. Many receptionist tools record or transcribe, so include a notice in the greeting and check the exact wording with a local lawyer.

Does the FCC ruling on AI voices apply to me?

The February 2024 ruling treats calls made with AI-generated voices as artificial under the TCPA, which requires prior express consent for covered calls. It concerns calls placed to people. Answering inbound calls is different, but outbound reminders or marketing need legal advice.

How accurate is it with names and numbers?

Less accurate than with ordinary words. Unusual names, spelled emails, and fast phone numbers are common error points. Configure the system to read back anything you will act on, and test with callers who have different accents and voices.

How long does setup take?

A finished product can be live in a day or two once you have written your answers and cleaned your calendar. A voice-agent platform takes longer because you build the call flow. Plan on a first month of reading transcripts and adjusting instructions.

Should I choose a platform like Retell or Vapi?

Choose a platform if you have developer help and want control over the call flow. Their per-minute rates look low, but the build and maintenance work is yours. A finished product trades some flexibility for a faster, simpler setup.

What are the hidden costs?

Setup and testing time, calendar and CRM integration, callbacks for missed calls, extra concurrency or numbers as you grow, and regular transcript review. None of these appear in a rate card, so add a line for them when you compare against a human option.

Can it replace my receptionist?

It can take over routine calls and after-hours coverage. A front desk that welcomes visitors, handles complaints, or makes judgment calls still needs a person. Many small businesses use AI to cover the gaps rather than to replace staff.

How do I test one before buying?

Place at least twenty calls yourself. Use different voices, a noisy location, a mobile phone, a request it should refuse, and a request that needs a human. Ask for a trial, read the transcripts, and check how the billing meter counted each call.

Does Communicate answer phone calls?

No. Communicate does not provide voice or phone answering. It offers an AI agent for chat and messaging, a shared inbox for human handoff, data sources, and actions.

If you need a phone receptionist, use a dedicated phone product and consider Communicate for web chat.

What about outbound calls from an AI?

Outbound calls carry stricter rules. The FCC treats calls using AI-generated voices as artificial under the TCPA, which generally requires prior express consent. Consult a lawyer before using AI to call customers, and keep consent records.

What metrics should I track after launch?

Track calls answered, calls transferred to a person, messages taken, bookings made, booking errors, and complaints. Review transcripts weekly in the first month, then monthly. Compare monthly cost with the cost of missed calls you previously lost.