Let an AI receptionist finish three jobs on its own: the greeting, the message, and a booking taken from a calendar you control. Make it hand off the moment money, a complaint, an emergency, or a question it was never briefed on comes up, by live transfer to a person, or by a callback promise with a time attached. You set that line in your own firm, and you brief the machine so it holds while you are out of the office.
Start from your own phone records
Pull three counts from your own phone system for one month: calls answered, calls that went to voicemail, calls that rang out. Those three counts are the whole input, and they are the only ones that describe your firm. The missed-call percentages that open most articles on this subject have no primary source behind them, so use your own counts.
Google ended the chat and call history features in Business Profile on July 31, 2024. Your profile still receives calls. What ended is the record of who called and when. Google's help page says past call history was downloadable only until August 30, 2024, so that history is gone.
Write one line per kind of call
For one week, log every call by kind in the words you would use out loud: a new enquiry, a question from an existing client, a supplier or a delivery, a sales call you do not want taken at all, anything urgent. Then write one line per kind with two halves. What may the receptionist do on its own, and what triggers a handoff.
This is the same list-making step we argue for in how to pick the one workflow worth automating first. The phone is a good first workflow because the same calls come in all day, and a bad answer costs you a customer at the exact moment they were willing to buy.
What it finishes on its own
Four jobs contain no judgment: the greeting, a name and a number, the reason for the call in the caller's words, and a booking into slots you chose and own. A machine does all four at lunch, after hours, and on the day your front desk is short.
Everything past those four is a handoff: money in any form, including a quote, a discount and a disputed invoice; a complaint; an emergency; a caller in distress; a regular who expects to be recognised; and any question that is not in the briefing. Firms often leave that last trigger out of the brief, and callers ask questions nobody wrote an answer for.
Exhibit 1
The receptionist finishes the greeting, the message and the booking, and hands off money, complaints, emergencies and every unbriefed question
Brief it in your own words
Dictate the brief as though you were training a temporary receptionist, then edit the transcript. Real hours, real closures, the real address, and prices only if you want prices quoted to strangers. Give it the three questions every new caller must be asked before the call ends: name, number, and the reason for the call. Write out the answers to your repeat questions in plain sentences, which is the work in answering repeat customer questions with AI; a phone brief is that list spoken aloud, one question per answer, no clauses.
Then read what it actually said. Read every transcript for the first two weeks, then once a week. A month of unread transcripts is a month of callers being told your hours are wrong.
Tell callers they are speaking to software
Say it in the first sentence of the call: you are speaking to an automated assistant for the firm, and a person can take a message or arrange a callback. Pew asked 5,119 U.S. Adults in February 2026 whether they use AI chatbots and 49 percent said they do, up from 33 percent in 2024, and in the same survey 71 percent said they think the increased use of AI will make their personal information less secure. Those are the general public, not your callers; Pew did not survey people who call small service firms. Its customer-service figure measures awareness, 65 percent of the 11,004 adults surveyed in December 2022 correctly identifying that chatbots use AI, and it says nothing about preferring a human.
There is no single national disclosure rule, so be careful with anyone who sells you a compliance script. California's bot law covers a bot communicating "online", which the code defines as a public-facing internet website, web application, or digital application, and it needs intent to mislead in order to incentivize a sale; a phone call is not inside those words. Utah's duty sits at Utah Code 13-77-103: a business using generative AI to talk with a customer must say it is AI and not a human when the customer asks a clear and unambiguous question, and people in regulated occupations must speak the disclosure at the start of a verbal interaction in high-risk cases. For most firms that is a duty triggered by a question, not a standing announcement. Colorado wrote a disclosure duty into law in 2024 requiring a business to tell each consumer they are interacting with an AI system, excused where it would be obvious to a reasonable person; the effective date moved twice and SB 26-189 repealed and reenacted that part, so the current rules apply to consequential decisions made on and after January 1, 2027. The February 1, 2026 and June 30, 2026 dates still printed around the web belong to superseded editions.
The FCC's February 2024 ruling holds that AI-generated voices count as an "artificial or prerecorded voice" under the Telephone Consumer Protection Act, and it requires prior consent from the person you call. That is about calls you place with an AI voice. Answering the phone is not dialing out, and the ruling does not reach the calls your business receives.
For most firms the disclosure is your choice to make, and we would make it. A caller who finds out by asking the assistant directly starts the call annoyed, and Utah gives that caller the right to ask.
Recording is a separate rule from disclosure
Telling someone they are talking to software does not authorize recording the call. Florida is an all-party-consent state, and its statute makes recording lawful only when every party gave prior consent; recording without it is a third-degree felony. Read the statute as the worked example, then check your own state, because one-party-consent states exist. If your receptionist records, the consent prompt has to come before the recording starts.
Health information makes the vendor a business associate
If callers describe health matters to your receptionist vendor, you have handed patient health information to a third party. HHS guidance is plain that the rule is a written contract giving satisfactory assurances the vendor will safeguard the information, and it lists a third-party chatbot handling symptom assessment, medical reminders and appointment scheduling for a provider as an example of a business associate. The test is a signed business associate agreement. A vendor page calling itself HIPAA compliant is not the test, and no product carries compliance as a property. Our note on ChatGPT in a medical practice covers the same distinction on the text side.
Four numbers each week
Read them against your own month of phone records, since anyone else's benchmark describes a different firm.
Exhibit 2
Calls answered, complete messages, bookings that held and hang-ups are the four numbers that say whether the receptionist is working
An answered call with an incomplete message is a missed call that cost you money. A booking that does not hold is worse than no booking, because it consumes a slot and a client's morning. Hang-ups in the first seconds of the greeting tell you the announcement is badly worded, a wording change you can make yourself.
What it does well and what it drops
It is good at the calls nobody wants. An enquiry after closing time, a third caller asking where you are, and a sales call to end politely at once. It never has an off day, and it asks the same three questions every single time.
It drops the calls where tone carries the meaning. A caller in distress, a negotiation about a bill, a regular who expects to be recognised by name, an accent it cannot parse, a line that breaks up in the middle of a number. Those need a person on the line. Add to them every question you never wrote down, because the assistant will improvise there and sound confident doing it.
The live transfer is the part to design
The line sits where the judgment starts, and building the handoff is a design job of its own. When the trigger fires, the assistant should pass what it already knows: the name, the number, the reason for the call, and which rule sent it to you. The caller should not repeat everything to a second voice moments later, and if your vendor cannot show you that transfer carrying a written summary, treat the handoff as a voicemail with extra steps. The principle is the one in the approval step: an automation acts alone inside a boundary, and asks a person at the edge of it. Write the callback promise with a time in it, "before 4pm today", and keep the promise on the day.
The cost against a human answering service
The three vendors we looked at price the same job on three units, all as shown 2026-11-03. Smith.ai sells an AI receptionist as a monthly subscription plus a per-call rate: the Free plan is $0 a month with 25 calls and $3.00 for each call after that, and Pro is $150 a month with a $2.00 per call rate, which the page describes as about two to three calls a day. It bills month to month. Ruby sells live human receptionists by the minute inside a monthly plan, its smallest phone plan at 50 receptionist minutes for $250 a month, then 100 minutes for $395, 200 minutes for $720 and 500 minutes for $1,725. Goodcall prices per unique customer per month, its Starter plan at $79 a month for 100 unique customers and 79 cents each after, with no minute metering at all.
Those prices are not comparable as printed, and no vendor will tell you yours. Ruby's smallest plan works out to $5.00 a minute of live receptionist time, which is our arithmetic on $250 buying 50 minutes, and it shows why the plan counts minutes and not calls: your bill depends on how long your calls run. A caller who rings back five times costs Goodcall nothing extra and costs Smith.ai five calls. To compare them you need your own average call length and your own call count, which is what your own month of phone records gives you. Building from parts is billed by usage, and OpenAI's rates are only part of that bill: $32.00 per million audio input tokens and $64.00 per million audio output tokens for its larger realtime model, $10.00 and $20.00 for the smaller one, and $0.05 per minute for one live voice model. The phone number, the integrations and the monitoring are not in that figure, and nobody maintains them for free.
Exhibit 3
An AI receptionist answers every ring, a human answering service answers conversations, and voicemail answers nothing
Questions to put to a vendor, and the sign it worked
Ask five questions before you point a number at it. Where does the audio go. How long is it kept. Will it train their model. Can you export your transcripts in a format you can read after you leave. What happens when their service is down, since a receptionist that fails takes your phone with it.
Judge it after a week where you never touched the phone. Every caller got an answer, the right people got the messages, and the bookings held. Setting the call list, the brief and the handoff rule is the part of this that we do with firms, on the AI enablement page.