9:42pm
AI voice agent for small business: answering the calls an owner misses
An AI voice agent for small business, in plain words: what it does when nobody can pick up, what it must never do, the federal and state rules for AI voices and call recording, a test of whether it is worth it, and a plan for this week.
An AI voice agent for small business answers the calls you miss while your hands are full: a caller who reaches voicemail may have booked someone else by the time you ring back. We help local business owners decide whether one should answer those calls, and we set one up when it should. Below is what it does and what it must never do, the federal and state rules for AI voices and call recording, a test of whether it is worth it for your business, and a plan for this week.
An AI voice agent is a phone voice that answers when you cannot, says at the start that it is an AI assistant, answers from your own information and books a call back for anything it cannot check. This page is written for owners whose phone rings while they are busy with the work itself: on a roof, under a sink, with a patient or a client in the chair.
An invented example: the phone rang at 9:42 on a Sunday evening. After the afternoon's storm, water was coming through a bedroom ceiling, and the caller had found three roofers on the map. Yours was the first number they tried.
You were at dinner. Four rings, then your voicemail, then the small silence where someone decides what to do next.
They did not leave a message. They went back to the map and called the next roofer on it.
In this invented example, the owner missed a chance to speak with a customer who was ready to book. A dental office with a patient in the chair, a salon owner halfway through a color and a real estate agent at a showing can miss the same call in the same way.
In short
- A call that rings out without a message can leave only a number and a time, not the reason anyone rang, and by the time you call back the caller may be talking to someone else.
- Speed decides who gets the conversation. Across 1.25 million leads, firms that made contact within the hour were nearly seven times likelier to get a meaningful conversation with a decision maker than firms that waited one more hour.
- An AI voice agent answers the calls you cannot take, asks the questions you approved and, when it is connected to your calendar, books a time. It does not close, and its first sentence says it is an AI assistant.
What does a missed call cost a small business?
A missed call costs a small business the chance to speak with a customer who was ready to book, and we could not put a figure on that cost. We could not find a published study of how fast small businesses answer their phones, in any trade. The closest careful work we found is from 2011; when we looked for something newer on November 3, 2026, we found only figures published by companies that sell phone tools.
In 2011, Harvard Business Review reported an audit of 2,241 US companies. The researchers submitted a test inquiry through each company's own website and timed how long the reply took.
The evidence
How long 2,241 companies took to answer an inquiry from their own website
Share of the 2,241 companies audited. Source: Harvard Business Review, The Short Life of Online Sales Leads (2011), reporting the audit by James Oldroyd, Kristina McElheran and David Elkington.
Cross-industry research from 2011, timing replies to inquiries submitted through a website rather than to phone calls. These percentages are not a benchmark for any one trade and should not be quoted as one. They are a reason to measure your own response times, not a measurement of them.
According to the audit, twenty-three percent of the 2,241 companies never replied at all. Among the companies that replied within 30 days, the average reply took 42 hours.
Two notes before that number gets used for anything. It is cross-industry research from 2011, not a study of your trade from this year, and the inquiries were web forms rather than phone calls. It does not tell you how fast anyone answers the phone today. What it gives you is a reason to measure your own response times, because the audited companies were spread from within the hour to never.
The same article reported a second study, across 1.25 million sales leads at 29 business-to-consumer and 13 business-to-business companies. Firms that tried to make contact within an hour were nearly seven times as likely to qualify a lead as firms that waited one more hour, and more than sixty times as likely as firms that waited a day.
The definition is the part to read closely. They counted a lead as qualified when it produced "a meaningful conversation with a key decision maker". Not an email opened, not a form completed: a conversation with the person who decides. That is what the caller at 9:42 wanted, seen from the other end: a person who could say when someone would come and look at the roof.
An invented example: a roofer calls back every missed call first thing the next morning. The caller who wanted a quote that evening already has one from someone else.

A useful first step costs nothing: open your phone's call log and mark every missed call from last week that left no voicemail and got no call back. That gives you a list of calls to investigate, not a measured loss. Some will be spam, repeat calls from the same person, or people who were never customers, and sorting them is part of the exercise.
Why do callers hang up instead of leaving a voicemail?
An invented example: a caller with three other businesses' numbers on the screen reaches your voicemail, decides not to wait for a call back, and rings the next one.
We found no study that measures how often callers hang up rather than leave a message, so this page does not print a figure. The narrower point holds without one. A missed call with no message may not tell you who called or why. The call log keeps the number and the time, but not the purpose, so a missed call from a customer can look the same as a wrong number.
That is why a phone gap is harder to recover from than an inbox gap. An unanswered email is still there in the morning with someone's name and question in it, and you can still be the one who replies. An unanswered call may leave only a number, and by the time you ring it, the caller may already be talking to someone else.
An invented example: a salon owner sees a few missed calls from unknown numbers after a busy Saturday and treats them all as spam. One of them was someone trying to book a whole wedding party.
In your numbers
What do the calls you never took add up to?
Evenings, weekends, a job site, a client in the chair. Count the ones that rang out.
Wrong numbers, sales calls and suppliers are not inquiries.
When do you usually call them back?
Conversations a year going to whoever picked up
90a year
- Calls that rang outyour 25
- 25a month
- That were a real inquiryyour 60%
- 15a month
- You never get back into a conversation withhalf of them
- 7.5a month
- Over a year12 months
- 90a year
Two things are worth knowing about these sums. The first is that they are far gentler than the research above would justify. The 1.25 million lead study found firms contacting within an hour were nearly seven times likelier to get a meaningful conversation with a decision maker than firms that waited one more hour, and more than sixty times likelier than firms that waited a day. Turning odds like that into a slider would produce a number nobody should believe, so the callback curve here is a plain judgement, and yours to disagree with. The second is that the sum stops at conversations. What one is worth depends on your trade and your prices, so that last step is yours to finish. This counts inbound calls only, and the contacts already sitting cold in your records are a different population in a different article.
Counting last week's missed calls that left a voicemail, against the total, shows how much of the problem this post is about applies to you.
What does an AI receptionist for small business do?
An AI voice agent for small business is a voice on the phone that can hold a short conversation, connected to the systems that make its answers true: your services and the area you cover, your calendar and your customer records.
As our voice agents page describes the agents we set up, they do four things. They answer the calls you cannot take, at any hour you choose. They can call new inquiries back quickly, once the consent and calling-hours checks below are met. They ask the first questions you would ask, such as what the job is, where it is and how soon it is needed, and book a time from your calendar. And they hand you the notes. How fast an agent answers, and what it can reach, depends on how it is set up and which of your systems it is connected to.
What it actually does
Four moves.
It answers the calls you cannot take, including the ones at nine on a Sunday.
Inbound calls route to the agent whenever you are unavailable, or always, depending on how you set it up. It says it is an AI assistant, answers what it can from your own information, and books a call back for the rest.
It can call new inquiries back quickly, when the checks allow.
When an inquiry lands from your site, a portal or a campaign, the agent first checks the consent the person gave and the calling hours where they are. If both pass it calls; if not, the callback goes on a person's list.
It qualifies inside a normal conversation.
What the job is, where it is, how soon it is needed and how they found you. Not a script read at someone. The questions you would ask, in the order a conversation actually goes.
It books, and then it writes it down.
When it is connected to your calendar, it offers times from your live availability and confirms one on the call. Afterwards the notes and the outcome go into your customer records, so the appointment you walk into is one you already understand.
The fourth of those is the one people underrate. When you ring back, you already know where the caller got to and what worried them, and on the phone you can hear whether the second version of an answer matches the first.
An invented example: a contractor rings back and already knows the caller wants a quote for a new water heater, asked twice about the permit and needs it before the holidays. For a dental office the same notes read: a new patient, a cracked tooth, can only come after work. For a real estate office: got as far as the taxes, asked twice about the septic, used the word soon.
Before you talk to anyone about an agent, write down the questions callers ask you most; those are the answers it must give from your own information.
Can an AI voice agent make calls as well as answer them?
An AI voice agent can also place calls; everything above is about the call you did not take. The other half is the inquiry that arrives without a call.
Some inquiries never come by phone. They arrive by a form at nine at night, from a portal or from an ad, and then wait in a list for someone to notice them. The usual answer is a follow-up task, which gets done tomorrow, or on Thursday, or not at all.
The second of the four moves is that other direction. An agent can call a new inquiry back, call one person when you tell it to, or work through a list. Each of those is an outbound call with an artificial voice, and that changes the rules.
The other direction
Calling out, and the rules it follows.
You can point it at one person and tell it to call.
It does not have to wait for a trigger. Say someone registers and agreed to a call back. You tell the agent to dial that contact, and it does, inside the calling hours, with the same questions and the same booking at the end.
It can try again on a schedule you set.
Someone who did not pick up at two in the afternoon may pick up later. You decide how many tries and when, inside the calling hours, and a request to stop ends the calls.
It can work a list, but only a consented one.
Past inquiries who agreed to this kind of call can be worked through, with the ones ready to move handed back to you. A name without that consent comes off the list before the agent sees it.
A person approves every list first.
A seasonal reminder or a new opening in the calendar can go to the people who asked to hear about it. A person approves the list and the script before the first call, and the call step cannot run without that approval.
An AI-voice call is not the same, legally, as a person dialing. Under the Telephone Consumer Protection Act, a call that uses an artificial or prerecorded voice needs the called person's prior express consent, unless it is for an emergency or falls under an exemption, and the Federal Communications Commission confirmed in February 2024 that AI-generated voices count as artificial voices. Telemarketing calls also have to respect the federal do-not-call rules and the calling hours, no earlier than 8 a.m. and no later than 9 p.m. at the called person's location. So the list an agent calls should be people who gave the right consent for that kind of call, and nobody else.
An invented example: a cleaning company's web form comes in at nine at night. Before the agent dials, it checks that the form carried the consent the company's attorney approved for an AI-voice callback and that it is still within the calling hours where the person lives. If either check fails, it schedules the call for a person instead. The same company never points the agent at an old list of people who did not ask.
Before you point an agent at any list, write one line beside each name saying how and when that person consented to be called; a name with no line comes off the list.
What decides whether an AI voice agent works on the phone?
Text chat forgives a pause. A typing indicator is a promise that something is coming, and two seconds of it costs nothing.
A phone call has no typing indicator. There is only silence, and the caller has to guess what it means: the line dropped, the other person is thinking, or nobody is there.
How long people normally leave between turns was measured in a 2009 study of conversation in ten languages, by Tanya Stivers and ten co-authors in the Proceedings of the National Academy of Sciences. It found the average gap before an answer was about two tenths of a second, and every language's most common gap fell between nothing and two tenths. The conversations were face to face, not phone calls, and the study says nothing about how callers judge a machine.
The size of the gap
How long people leave before answering a question, measured in ten languages
The grey track behind each bar is one full second. The outer two bars are the fastest and the slowest of the ten language averages. The inner two are the middle and the average of the whole dataset, which sit apart because the spread runs long. Ten languages on five continents, measured from video of ordinary unscripted conversation between two and six people. Yes or no questions were 67% of everything coded. Source: Stivers and ten co-authors, Universals and cultural variation in turn-taking in conversation, PNAS 2009.
Read what this does not say. Every conversation in it happened face to face, and not one of them was a phone call. The same paper found responses came earlier when the questioner was looking at the recipient in nine of its ten languages, reaching significance in five, which is exactly what a phone takes away. It measures people answering people, so it has nothing at all to say about what a caller will put up with from a machine. What it does establish is the size of the window. Every one of the ten languages had its commonest gap somewhere between nothing and two tenths of a second. The authors also note that speakers are sensitive to shifts of a hundred milliseconds. An agent that takes a full second has left more than twice the gap of the slowest language average in the study.
What it does show is how small the normal window is. That is why response delay is a design goal on the agents we set up: the gap between a caller finishing a sentence and the agent starting its answer should be as short as the setup allows. A long gap is something you can hear in a test call, so test it yourself rather than taking anyone's figure, ours included.
Two related checks belong in the same test call, beside the others in the legal section.
People interrupt. Someone starts answering before you have finished the question, you stop talking, and neither of you thinks about it afterwards. An agent that cannot be cut off mid-sentence talks over the caller.
Turns are short. On the phone a person says four words and waits. "Do you work Saturdays." An agent tuned to write paragraphs answers a four-word question with a speech.
On a phone, silence is all the caller hears while the agent works out its answer. Keep it short, and test it yourself.
The teardown
Watch it take the 9:42 call.
The pattern, not a recording of a real client.
What the caller heard
- Caller9:42 pm
Hi. Water is coming through our bedroom ceiling since the storm. Do you come out to Beacon?
- Agent9:42 pm
I am the company's AI assistant, and this call is recorded so nothing gets lost between now and the morning. Beacon is inside the area the company covers. What would you like to know?
- Caller9:43 pm
Can someone look at it tomorrow, and what does a repair like that usually cost?
- Agent9:43 pm
The visit I can book straight from the calendar. The cost I am not going to guess at. Nobody knows until the roofer has looked at it, and it is the kind of answer that decides who you hire. Let me book the visit, and a person will go through the cost with you there.
- Caller9:44 pm
Tomorrow morning is fine.
- Agent9:44 pm
Nine thirty is open. You are booked in, and I have sent the details to the number you are calling from.
What happened behind it
- 9:42 pm
Answered on the second ring
No hold, no voicemail, no callback form.
- 9:42 pm
Said what it was
An AI assistant, and the call is recorded. Before anything else.
- 9:43 pm
Answered from the record
The service area, read off the company's own list rather than recalled.
- 9:43 pm
Declined to invent an answer
Flagged for a person instead of guessing at it.
- 9:44 pm
Booked from the connected calendar
Offered a time from the live availability.
- 9:45 pm
Written to the records
The notes and the outcome, before dinner ended.
An invented example: a caller asks a gym's agent "Do you work Saturdays" and gets a paragraph about the class timetable and the parking. Halfway through, the caller hangs up.
When you try any phone agent, cut it off mid-sentence and ask it a four-word question. Note whether it stops, how long it leaves before answering, and whether the answer is short. Then check the rest of the list: that what it says is accurate, that it says it is an AI assistant, that it respects consent, that it books the right time, and that it hands over to a person when it should.
What should an AI receptionist never do?
The limits of an AI receptionist come first, before anyone buys anything.
It does not close. A first call carries more than words: the pause before an answer, a second voice in the background, cheerfulness that is covering something. A transcript may leave out tone, pauses and background context, and a model reading it cannot recover what was never written down.
It does not know what it cannot verify. Ask it whether the part is in stock, or whether the owner would take less for a job paid in cash. The correct behaviour is to say plainly that it does not know, and to put a call on your calendar. A voice that invents an answer to sound competent is worse than no answer, because the caller cannot check it on the phone.
It does not pretend to be a person. The agents we set up introduce themselves as an AI assistant in their first sentence, and we will not build one that claims to be a particular human being.
And it does not stand in for you with the people who matter. If the caller is a regular customer, a neighbor of a job you are on, or anyone whose relationship is the business, the agent takes a message and arranges your call back. It does not try to be the relationship.
An invented example: a caller asks a landscaper's agent whether the price includes removing the old tree. The agent says it does not know that yet and books a call for the next morning.
The limits, before anything else
Four things it never does
1
Close
Tone, pauses and the second voice in the room stay with you.
2
Guess
What it cannot check becomes a booked call.
3
Pretend
Its first sentence says it is an AI assistant.
4
Replace you
Past clients and referrals get a message and your call back.
The list of things your agent must never answer on its own, such as a price, a legal position or a fact it cannot check, belongs in its setup before it takes a single call.
What are the legal rules for AI voice calls and call recording?
None of this is legal advice, and all of it moves. The point is to know the questions exist and to take them to your own attorney. The New York and California statutes below were read on November 11, 2026 in a publisher's copy marked current as of January 1, 2026, and California's AB 2905 on November 3, 2026 in the Assembly committee's analysis; the official state pages refused our automated reader, so we name them as the citations and recommend reading them there.
An AI voice is not a loophole. In a declaratory ruling adopted on February 2, 2024 and released on February 8, the Federal Communications Commission confirmed that the TCPA's restrictions on an "artificial or prerecorded voice" cover current AI technologies that generate human voices. Such calls therefore need the called party's prior express consent, absent an emergency purpose or an exemption. The ruling also notes that an artificial-voice message must identify the business responsible for the call, and that one which is telemarketing must offer a way to opt out. We read the ruling again on November 11, 2026.
Recording is a state question. In New York the offense is eavesdropping, Penal Law section 250.05, and the kind that applies to a telephone call is wiretapping, defined in section 250.00 as "the intentional overhearing or recording of a telephonic or telegraphic communication by a person other than a sender or receiver thereof, without the consent of either the sender or receiver". Under that definition the consent of one party to the call can be enough in New York. California is stricter: Penal Code section 632 makes it an offense to record a confidential communication without the consent of all parties. A call between two states can bring in the stricter state's rule, and federal law and other states have their own provisions, so ask your attorney which rules apply to the calls you take.
An invented example: a Hudson Valley contractor takes a call from a man in Los Angeles arranging a repair at his mother's house. Plan for California's rule, and ask your attorney before you rely on anything else.
Take this to anybody, including us
Three questions that separate a product from a demo.
Ask what it says in its first sentence.
If the answer is a person's name rather than the words AI assistant, stop there. Everything else on this list depends on whether the provider treats disclosure as a feature or as a problem to be managed.
Ask what it does with a question it cannot answer.
You are listening for one specific behavior: it says so, and it books a call. If the demo never once shows the agent failing at something, then you have been shown an advertisement rather than a product.
Ask to hear a call that went sideways.
Anything sounds good reading a script to a cooperative caller. Ask for a recording where somebody interrupted it, changed their mind halfway, or went quiet.
Some states already require the disclosure. California's Assembly Bill 2905, Chapter 316 of the Statutes of 2024, requires a natural voice announcement to say whether an automatic dialing-announcing device's prerecorded message uses an artificial voice, defined to include a voice generated using artificial intelligence. It took effect on January 1, 2025, and it applies to that kind of device, so ask your attorney whether your calling setup is one.
Our policy is simpler than the map. The agent says at the start of every call that it is an AI assistant. Where the call is recorded, it also says so at the start, before anything else is said. If a caller does not want to be recorded, the agent stops recording or offers a call back from a person, depending on how the owner and the attorney set it up. Disclosure and recording consent are two different things, and the script handles both.
If you plan any outbound calling, have your own attorney review the script and the calling setup before the first call goes out.
The script itself
Ask and we will send you in writing the exact opening our agents use. One line says it is an assistant. The other says the call is recorded. Give those to your attorney rather than a blank page.
One page, no charge, and yours to use whoever ends up building the thing.
The system
What it is connected to.
An agent is only as good as what it can reach. This is every hop between the ring and the appointment you walk into.
Scroll to follow the chain
- The call: Your own number, any hour
- The agent: An AI assistant, given limits
- Your services: What it may state
- Your calendar: Live availability, not a guess
- The record: Your records, with the notes
- The handoff: Time booked with a person
Where does an AI voice agent go wrong?
An AI voice agent goes wrong in three ways: a phone agent that works technically still loses you the call, each a setup or follow-through problem:
- It is set up to answer everything. An agent that never hands off ends up standing between a serious caller and you. The setting that matters most is the point at which it stops talking and books.
- Nobody told it what it may not say. With no limit on price, condition or anything legal, it fills the silence fluently, and you hear about it from the client rather than from the log.
- Nobody reads the calls. The record of what callers asked, in their own words, sits unread, and the mistakes in it keep repeating.
Three ways it fails
Each one is a setup problem.
It is set up to answer everything.
Set the hand-off first: the point at which it stops talking and books.
No limits were written down.
Write the limits on price, condition and anything legal before the first call.
Nobody reads the calls.
In the trial a named person reads every call; after it reads cleanly, a weekly reading.
An invented example: a salon's agent is set up with last spring's opening hours and keeps offering callers a Saturday slot the salon no longer opens. Nobody reads its calls, so nobody notices for weeks.
For the trial, a named person, usually the owner, reads every call, decides what gets fixed, and can switch the agent off. After a failure that reached a caller, such as a wrong time booked or a question it should have handed over, the agent stops taking that kind of call until the setup is fixed. Only after the trial reads cleanly does that move to a weekly reading of calls from start to finish.
Who an AI voice agent fits, and who it does not
An AI voice agent fits owners whose calls come in while they cannot pick up: trades on job sites, a dental or accounting office whose front desk is with someone, salons, spas and gyms with their hands busy, real estate agents at showings, and anyone whose callers ring in the evening or at the weekend. It fits best where a caller's first questions repeat and the next step is a booked time.
It does not fit an office where a person already answers almost every call and after-hours calls are rare. It fits poorly where nearly every caller is a past client or a referral who expects you by name, or where the first call is usually a hard or private conversation.
Those owners do better without an agent: a person on a shared rota for the phone, a voicemail greeting that says when you call back and then keeps that promise, or a short text back to a missed call where the caller has agreed to texts.
An invented example: a home inspector who works only from referrals and calls everyone back within the hour gets little from an agent. A clear greeting does the job.
Who this is for
A good fit, a poor fit, and what to do instead
A good fit
Calls that come while you cannot pick up: job sites, a patient or a client in the chair, showings, evenings and weekends. First questions that repeat, and a booked time as the next step.
A poor fit
An office where a person answers almost every call, or callers who are nearly all past clients and referrals expecting you by name.
Instead, for a poor fit
A shared phone rota, a greeting that says when you call back and keeps that promise, or a text back where the caller agreed to texts.
Either way
Whatever answers the phone says what it is, and a person owns what happens next.
Is an AI receptionist worth it for a small business?
Whether it is worth automating is a test you run on one job, and "leave it alone" is a real answer. The job here is answering a call you cannot take. These are the screening questions of our AI assessment, our own method rather than a published standard; the post on what to automate first and what to leave alone sets it out in full.
Start by sorting the job into clicks, the steps a system can repeat, and decisions, the judgement you are paid for. Only a decision that is settled and approved, such as the opening sentence or the times it may offer, becomes a default the agent reads on every call. A judgement that is not settled, a sensitive decision or an exception stays with a person.
Then three questions. Does the job happen often enough that a failure is noticed within a week? Can the rule be written so a new hire could follow it? Does a wrong answer land where a person sees it first? Then the stop question, which asks what breaks if the job stopped: if nothing breaks, stop doing it rather than automate it. Last, ask whether the phone system or booking tool you already pay for has a setting that does it.
In the assessment we score each job for impact and risk from 1 to 5, each with a reason. Four checks set a floor under the risk: does the result bind the business, reach a customer, feed another system people act on, or risk damaging trust if it is wrong? One yes means the risk is at least 3, and two or more mean at least 4. Answering a caller reaches a customer, so the risk is at least 3 and a person stays on it. In practice that means the agent may say only what a person approved in advance, and anything that binds the business, such as a price, an offer or an outbound list, waits for a person's approval. The step that places an outbound call cannot run until that approval exists. In our method, a job whose impact scores 1 or 2 is left alone at any risk, and a person keeps doing it.
When we interview an owner for the assessment, we use the two-minute test: write one job in two minutes under input, do, output and the person's check, hand it to a colleague who does not know it, and note where they stall. Where they stall is a decision you have not written down yet.
An invented example: the manager of an accounting office writes up "answering a call after six" in two minutes. Her assistant stalls at "what do I say about appointments in tax season", so that default gets written and approved before anything is built.
Before anything is set up
Is answering your missed calls worth automating?
Each question can send the job to the tray. What survives keeps a person on it, because a caller is a customer.
Does anything break if the calls go unanswered?
Yes, carry on
If the answer is No, so leave it alone
Stop worrying about it
If nothing breaks, nothing needs building.
Is it missing from the tools you already pay for?
Missing, carry on
If the answer is One does it, so leave it alone
Switch the setting on
A phone system's own feature comes first.
Can the rule be written for a new hire?
Yes, carry on
If the answer is No, so leave it alone
Settle the rule first
Each stall in the two-minute test is a missing default.
Does the impact score 3 or more out of 5?
Yes, carry on
If the answer is No, so leave it alone
Saves too little
A person keeps answering.
Set it up with a person on it
Risk at least 3: the agent says only what a person approved, and outbound calls wait for an approved list.
An invented example, not dataThe accounting office's assistant stalls at tax-season appointments, so that default is written first.
Running the two-minute test on answering a missed call, and marking where your colleague stalls, is the first piece of work.
What should an owner do first before buying an AI voice agent?
What to do first is one job at a time, on a suggested trial schedule rather than a fixed one.
- This week. Ring your own number from a phone your office has not saved, read the missed calls in your log, and run the two-minute test on answering them.
- The next two weeks. Write down and approve the defaults: what the agent may answer, what it books, what it never says, and the sentence it opens with, that it is an AI assistant and, where it records, that the call is recorded. Then try one direction only, such as evening and weekend calls on one line, with a named person reading every call.
- Then. Write the decision down: keep it, change it, or leave it alone, with the reason in one line and the name of the person who owns it.
An invented example: a heating and cooling company starts with weekend calls only, reads every call, and adds callbacks to new form inquiries only after the weekend calls read cleanly and the attorney has approved the consent wording on the form.
A suggested trial schedule
One job, one direction, a decision written down
This week
Ring your own number, read the missed calls, run the two-minute test.
The next two weeks
Approve the defaults and the opening line; one direction on one line, every call read.
Then
Keep it, change it or leave it alone, with the reason and the owner written down.
A schedule to start from, not a deadline: the reading decides when it moves on.
Common questions about an AI receptionist
What is an AI voice agent for a small business?
An AI voice agent for a small business is a phone voice that answers the calls you cannot take, holds a short conversation from your own information and calendar, says it is an AI assistant and books you to call back. It suits any small business whose callers ask the same first questions: a contractor, a salon, a dental office, a real estate office.
Will callers know they are talking to an AI?
Yes. The agents we set up say they are an AI assistant in their first sentence, and we will not build one that pretends to be a particular human being.
What happens when it cannot answer something?
When an AI voice agent cannot answer something, it says so and books a call with you instead of guessing. It does not invent a price, a legal position, or a fact about a job.
Does it take over the front desk?
No. It takes the repeatable first contact off the desk: answering calls nobody can pick up, calling back new inquiries where the consent and calling-hours checks are met, asking the first questions you approved, and booking a time. The people at the desk keep the conversations where judgement and relationship decide the outcome, and get the notes of every call the agent took.
Does it work with my calendar and my CRM?
An AI voice agent can work with your calendar and your CRM when it is connected to a calendar and a customer system it supports and has the permissions to read and write them. Connected that way, it offers times from your live availability and writes the notes and the outcome back. The appointment booking side uses the same connections. Check the booked times in the trial like everything else.
Is it legal to have an AI make outbound calls?
An outbound call by an AI voice can be legal when the checks for that call are met. Under the TCPA an AI voice counts as an artificial voice, so the called person's prior express consent is needed unless an exemption applies, and telemarketing also has to respect do-not-call rules and calling hours. The call's purpose, the person called and the states involved all matter, so have your attorney review the setup before the first outbound call.
Can we talk about an AI voice agent on a call if my business is outside New York?
We are based in New York, and the call works for an owner anywhere. The rules for AI voices and call recording depend on the states a call involves, as the legal section above says.
Ten questions
What did you learn? Check it in two minutes
Ten questions on what this page taught. It takes about two minutes.
Take the quiz
1 of 10
In the 2011 study, what counted as a qualified lead?
- A.An opened email
- B.A completed web form
- C.Having a meaningful conversation with the decision maker
- D.A booked appointment
Show the answer
C. Having a meaningful conversation with the decision maker The study defined a qualified lead as a meaningful conversation with a key decision maker; see "What does a missed call cost a small business?".
2 of 10
What does a missed call with no message usually leave you?
- A.The caller's name and question
- B.A number and a time, but not why they called
- C.Nothing at all, not even the number
- D.A recording of what they wanted
Show the answer
B. A number and a time, but not why they called The call log keeps the number and the time, not the purpose; see "Why do callers hang up instead of leaving a voicemail?".
3 of 10
What should the agent do when asked something it cannot verify?
- A.Give its best guess
- B.Change the subject
- C.Say it does not know and book a call with you
- D.Hang up politely
Show the answer
C. Say it does not know and book a call with you It says it does not know and puts a call on your calendar; see "What should an AI receptionist never do?".
4 of 10
In the 2009 study of ten languages, about how long was the average gap before an answer?
- A.About two tenths of a second
- B.About two seconds
- C.About five seconds
- D.About half a minute
Show the answer
A. About two tenths of a second About two tenths of a second, measured face to face, not on the phone; see "What decides whether an AI voice agent works on the phone?".
5 of 10
What did the FCC's ruling of February 2024 confirm?
- A.AI calls need no consent
- B.Only text messages are covered
- C.Recording calls is banned
- D.AI-generated voices count as artificial voices under the TCPA, so prior express consent is needed
Show the answer
D. AI-generated voices count as artificial voices under the TCPA, so prior express consent is needed The ruling brought AI voices under the TCPA's artificial-voice rules; see "What are the legal rules for AI voice calls and call recording?".
6 of 10
Under New York's definition of wiretapping, whose consent can be enough to record a call?
- A.Nobody's: recording is always banned
- B.One party to the call, though a caller in a stricter state such as California brings in that state's rule
- C.Only a court's
- D.Every party's, in every state
Show the answer
B. One party to the call, though a caller in a stricter state such as California brings in that state's rule The New York definition turns on the consent of the sender or the receiver, and California requires all parties; see "What are the legal rules for AI voice calls and call recording?".
7 of 10
Which decisions become defaults the agent reads on every call?
- A.Only settled, approved ones, such as the opening sentence or the times it may offer
- B.Every decision the owner makes
- C.Whatever the agent decides on its own
- D.None: the agent improvises
Show the answer
A. Only settled, approved ones, such as the opening sentence or the times it may offer An unsettled judgement, a sensitive decision or an exception stays with a person; see "Is an AI receptionist worth it for a small business?".
8 of 10
Answering a caller reaches a customer. What does that set in our scoring?
- A.Nothing: risk stays at 1
- B.The agent may call anyone it likes
- C.A risk of at least 3, so a person stays on it
- D.An automatic impact of 5
Show the answer
C. A risk of at least 3, so a person stays on it One yes on the four checks means the risk is at least 3; see "Is an AI receptionist worth it for a small business?".
9 of 10
Before an agent dials a new form inquiry, what has to be checked?
- A.Nothing: a form is permission
- B.Only that the phone number is valid
- C.That the owner is awake
- D.The consent for that kind of call and the calling hours where the person lives
Show the answer
D. The consent for that kind of call and the calling hours where the person lives If either check fails, a person calls instead; see "Can an AI voice agent make calls as well as answer them?".
10 of 10
What is the first thing to do this week?
- A.Buy an agent
- B.Call your own line from a phone the office has never saved
- C.Call a list of past leads
- D.Rewrite the website
Show the answer
B. Call your own line from a phone the office has never saved Hear what a stranger hears before anything else; see "What should an owner do first before buying an AI voice agent?".
How to check your own missed calls this week
The simplest check needs nobody, us included. It is a call to your own number on a Sunday evening, from a phone your office has not saved, sat through the whole way a stranger would sit through it: the rings counted, the greeting heard all the way to the beep. What you hear on that call is the first thing someone with water coming through the ceiling, or a toothache, hears from your business.
You can see the alternative on the home page. The AI voice agents page has the rest: what the agent is connected to, and what it does with a call once it has taken one. The same argument at the other end of the day is on the website chat side, where the customer messaging you late at night never picks up the phone at all.
An invented example: the owner of a small moving company rings his own line from his daughter's phone and hears a greeting that still names a manager who left.
In the story at the top, the caller at 9:42 did not try again on Monday. Ring your own number this Sunday and hear what they heard.
A voice agent is set up for each business, against its own number and its own calendar, so there is no shared demo line to ring.
Ask us about an AI receptionist for small business.
Sources
- Harvard Business Review, "The Short Life of Online Sales Leads", by James Oldroyd, Kristina McElheran and David Elkington, March 2011, on hbr.org: the audit of 2,241 companies and the study of 1.25 million leads. Read November 11, 2026; kept for the argument, as the closest careful work we found.
- Federal Communications Commission, Declaratory Ruling FCC 24-17 under the Telephone Consumer Protection Act, a federal law, CG Docket No. 23-362, adopted February 2, 2024, released February 8, 2024. Read November 11, 2026.
- Code of Federal Regulations, 47 CFR 64.1200(c)(1), the calling hours for telephone solicitations, in the current eCFR. Read November 11, 2026.
- New York Penal Law sections 250.00 and 250.05; California Penal Code section 632. Read November 11, 2026 in FindLaw's copies; the official pages refused our automated reader.
- California Assembly Bill 2905 (Low), Chapter 316, Statutes of 2024, effective January 1, 2025, as described in the Assembly Committee on Privacy and Consumer Protection's analysis of April 23, 2024. Read November 3, 2026; the official page refused our automated reader on November 11, 2026.
- Tanya Stivers and ten co-authors, "Universals and cultural variation in turn-taking in conversation", Proceedings of the National Academy of Sciences 106(26), 2009, on PubMed Central. Read November 11, 2026; kept for the argument.
- Our own capability source: the AI voice agents page; our method: the AI assessment.
Sources checked November 11, 2026. None of this is legal advice.
Written with an AI assistant and checked by a person.
Keep reading
All three leads look the same. Two are worth your morning. 
July 31, 2026
All three leads look the same. Two are worth your morning.
Website chatbot for a small business: the question that arrives at 11:40pm 
July 12, 2026
Website chatbot for a small business: the question that arrives at 11:40pm
Workflow automation for small business: follow up every quote, then wire the next job 
July 13, 2026
Workflow automation for small business: follow up every quote, then wire the next job
