AI Phone Answering for Small Business: What It Costs and How to Set It Up

A one-van plumbing business takes roughly ten calls on a busy day and answers maybe four of them. The rest arrive while someone is under a sink with wet hands, and the person on the other end is working down a list of three plumbers. By the time the callback happens at seven in the evening, the job is booked elsewhere. Nothing about that is a technology problem — it is a staffing problem that no small business can afford to solve by hiring a receptionist for $45,000 a year to handle forty calls a week. AI phone answering for small business exists precisely for that gap.

It has quietly become the cheapest fix available. A voice agent picks up on the second ring, finds out what the caller wants, books the job or takes the details, and texts you a summary. The interesting part in late 2026 is not that the technology works, but that the price has fallen to the point where a trade business can run one for less than its monthly fuel bill. What has not fallen is the amount of nonsense written about it, so this guide separates the parts you can verify from the parts vendors would like you to believe.

The numbers on missed calls are worse than you think, and worse sourced

Search for missed-call statistics and you will meet the same figure everywhere: 62% of calls to small businesses go unanswered. It is cited by dozens of vendor blogs, usually credited to a company called OnceHub, and none link to the underlying study. Treat it as folklore.

There is one genuinely useful dataset. The business phone app Quo, previously OpenPhone, analysed 16.7 million missed calls logged by its own customers over a three-month window and counted how many got an outbound call back within 48 hours. The answer was 31%. Just over two-thirds of missed business calls were never returned at all, and of the ones that were, nearly all the callbacks happened within the first 24 hours — if it did not happen the same day, it never happened. Broken down by sector, between 59% and 79% of missed calls went permanently unreturned, with healthcare the best performer at a 41% callback rate and consumer goods down at 22%.

That study has an obvious limitation: it only sees businesses that had already bought a phone system with call logging, a self-selecting group, and it measures callbacks rather than lost revenue. Use it as a direction, not a benchmark. The number that actually matters is yours, and you can get it this afternoon. Open your call log or VoIP dashboard, count the inbound calls for the last two weeks, how many went to voicemail, and how many of those you called back the same day. Multiply the gap by your average job value, then halve it to allow for time-wasters. That figure is the budget for everything below.

What AI phone answering actually means for a small business

Three different products get sold under the same label, and knowing which one you are buying prevents most of the disappointment.

Missed-call text-back is the simplest. When a call goes unanswered, an automation fires an SMS within a few seconds: Sorry we missed you — this is Dan at Brightwater Plumbing — what’s the job and which ZIP code? No AI is strictly required, although most tools now add a model to handle the reply. It is cheap, it is nearly impossible to get wrong, and it recovers the subset of callers who are happy to text.

An AI voice agent actually answers the phone and holds a conversation. It can ask qualifying questions, check a calendar, book a slot, and transfer to a human when it hits something it cannot handle. This is what people mean when they say AI receptionist, and it is the one with real upside and real failure modes.

An AI layer on your existing phone system sits in the middle: your current business number, your current app, with an agent that answers when nobody picks up.

Most businesses should run text-back and a voice agent together, because they catch different people. The caller who hangs up after two rings never reaches the agent, and the text-back is the only thing that reaches them.

What it costs in October 2026

The figures below come from each vendor’s own pricing page on 20 October 2026, not from comparison blogs, several of which are badly out of date. Check the live page before you commit.

The easy tier: a phone app with the agent built in. Quo charges $19, $33 or $47 per user per month on monthly billing, dropping to $15, $23 or $35 if you pay yearly. Every plan includes one number, auto-replies and the Sona agent that answers calls around the clock. The catch is in the metering: each plan comes with 1,000 automation credits, an answered call costs 100 credits, so the base plan covers ten AI-answered calls a month before you buy more credits, in tiers running up to $199 a month. For a business fielding forty missed calls a month, the headline price is not the real price.

The middle tier: a voice agent platform. ElevenLabs sells conversational agents on a minute bundle — $6 for 75 minutes, $22 for 275, $99 for 1,238 — with a free tier of 15 minutes to test, extra minutes at $0.08, and LLM usage billed separately on top. Telephony is not included; your Twilio or equivalent bill arrives separately.

Retell prices per minute and shows its working: $0.055 for the voice infrastructure, $0.015 for text-to-speech on its standard voices (ElevenLabs voices cost $0.040 to $0.10), roughly $0.024 to $0.025 for a mainstream model, and $0.015 for telephony. Its own calculator lands on $0.11 a minute for a default configuration, against a quoted range of $0.07 to $0.31 depending on what you switch on. Numbers are $2 a month, twenty concurrent calls are free, and there is $10 of credit to test with. Watch the add-ons, which are individually trivial and collectively not: knowledge base and denoising at half a cent a minute each, PII removal at a cent.

Vapi charges $0.05 a minute for hosting and passes the model, speech and telephony costs through at provider rates — Deepgram transcription around $0.0095 a minute, OpenAI between $0.0077 and $0.0452, ElevenLabs voices $0.0146 to $0.0238, Twilio inbound at $0.008. Its free tier gives one number and four concurrent calls; the $29 Core plan gives five numbers and ten.

Run the arithmetic on a realistic load. Forty answered calls a month at three minutes each is 120 minutes. On Retell’s default configuration that is about $13 of usage plus $2 for the number. On ElevenLabs the $22 bundle covers it with room to spare. Even the most expensive sensible configuration keeps you under $50 a month, which is roughly one recovered job a year in most trades.

The DIY tier: Twilio and your own glue. A local US number is $1.15 a month, inbound calls are $0.0085 a minute, outbound $0.014, and SMS is $0.0083 per segment before carrier surcharges of $0.0035 to $0.005 depending on the network. The real cost here is not the metering, it is the registration described below. Build this tier only if you already run workflows in n8n and want call data alongside everything else.

Setting it up in an afternoon

The sequence below assumes the middle tier, because that is where most businesses land.

Keep your existing number. Do not publish a new one. Your current number is on your Google Business Profile, your van, and three years of invoices. Set conditional call forwarding so that calls divert to the agent’s number after four rings or when the line is busy — every mobile carrier and VoIP provider supports this, usually as “forward on no answer”. The agent becomes a safety net rather than a gatekeeper, which also means a bad week with the agent cannot cost you the calls you would have answered anyway.

Write the brief before you touch the platform. The single biggest determinant of quality is the instruction you give the agent, and it should read like a note to a competent temp rather than a prompt. Name the business and what it does. List the five things callers ring about and what to ask for each. State the service area as a list of ZIP codes. Give the opening hours and what “emergency” means. Most importantly, state what the agent must never do: quote a price, promise a time slot that is not in the calendar, or claim a tradesperson is on the way.

Decide the three outcomes. Every call ends in one of: booked, details captured for callback, or transferred to a human now. Define the transfer trigger explicitly — water coming through a ceiling, a caller who asks twice for a person, a caller who sounds angry — and give it a real number to transfer to. The agent that cannot escape to a human is the one that generates complaints. This is the same discipline that makes automated support ticket triage work: the routing rules matter more than the model.

Connect the calendar, not a form. An agent that says “someone will be in touch” has converted a phone call into a lead, which is a downgrade. An agent that books a real slot has converted it into revenue. Most platforms connect to Google Calendar, Cal.com or a scheduling tool directly; the pattern is the same one behind automated demo scheduling, just with job types instead of sales calls.

Add the text-back path. Configure an SMS that fires on any unanswered or abandoned call, containing a booking link and one question. “Sorry we missed you — book here, or reply with your ZIP code and what’s wrong” outperforms “we’ll call you back” by a wide margin, because it gives the caller something to do while they still have the problem in front of them.

Then test it badly on purpose. Call from a highway. Call with a radio on. Mumble a ZIP code. Ask for a price. Ask something off-topic. Ask it whether it is a robot. Read every transcript from the first fifty calls — not a sample, all of them — and fix the brief rather than the model. Two hours of transcript reading in the first two weeks is worth more than any amount of platform configuration.

Close the loop afterwards. A booking the agent made is still a booking that can no-show, so feed it into the same reminder sequence as everything else; the mechanics are covered in the guide to automating appointment reminders. If you also want the agent’s call summaries scored for lead quality before they reach you, the approach in AI lead qualification applies directly to voice transcripts.

The compliance part nobody puts in the demo

Two separate regimes apply, and the one that bites first is the boring one.

Text messaging registration. If you send automated SMS to US numbers from a standard ten-digit number, you must register under A2P 10DLC before carriers will deliver reliably. Through Twilio, brand registration is a one-time $4.50 for sole proprietor and low-volume brands, campaign vetting is a non-refundable $15 whether or not it passes, and the campaign then carries a monthly fee of $1.50 to $10 depending on use case. A standard brand adds $41.50 of secondary vetting, bringing registration to around $46. Vetting can take a week or several, so start early. Skip this and your texts do not bounce visibly — they just quietly fail to arrive, which is worse.

Voice disclosure. The FCC’s February 2024 declaratory ruling put AI-generated voices inside the TCPA’s existing definition of “artificial or prerecorded” voice, so the consent and identification rules that already governed robocalls apply to AI calls too; under 47 CFR 64.1200 an artificial-voice message must identify the responsible business at the start and provide a phone number. A September 2024 FCC proposal would add an explicit “this call uses AI” disclosure, but as of late 2026 it has not been adopted and is not law, and the proposal was limited to outbound calls in any case.

Inbound answering sits in a grayer area federally, but several states do not draw that line — California’s artificial-voice disclosure requirement and Utah’s consumer-interaction disclosure rules are both worth reading properly, and Utah’s text is not restricted to outbound. Call recording adds a third layer, since two-party consent states require an announcement before the recording starts.

The practical default costs you nothing: have the agent say what it is in its first sentence, announce recording if you record, and offer a human on request. “Hi, you’ve reached Brightwater Plumbing — I’m the automated assistant, and I can book you in or put you through to Dan” is one sentence and removes most of the exposure. If you operate across state lines or in a regulated trade, spend an hour with a lawyer rather than an afternoon with a vendor’s compliance blog; the ones surveyed for this article contradicted each other on basic specifics.

Where these setups actually fail

The failure is almost never voice quality. In order of how often it shows up in transcripts:

The agent quotes a price. Someone asks what a new boiler costs, the model produces a confident number, and you have a customer who believes they were quoted $1,900. Forbid pricing in the brief and give the agent a sanctioned deflection instead.

The handoff has nowhere to go. A transfer rule pointing at a mobile that is switched off is worse than no transfer rule, because the caller has now waited through two systems. Test the transfer path weekly.

Nobody reads the transcripts. The agent repeats the same misunderstanding ninety times before anyone notices, because the dashboard shows a healthy answer rate.

The metering surprises you. Per-minute pricing looks trivial until a confused caller stays on for nine minutes, or an agent with no maximum call duration sits in silence on an open line. Set a hard call length cap and a monthly spend alert on day one.

Measurement stops at “calls answered”. The number that justifies the spend is cost per booked job: total monthly platform and telephony spend divided by jobs that exist because the agent answered. Track answer rate, booking rate, transfer rate and the share of callers who hang up in the first ten seconds — that last one is the tell that the opening line is too long or too obviously synthetic.

Start with the cheapest tier that covers your call volume, keep your existing number, read the first fifty transcripts, and raise the budget only when the cost-per-booked-job figure makes the decision for you. The businesses that get burned here are the ones that bought a voice agent for the novelty; the ones that profit from it are the ones who counted their missed calls first.