A missed call is often a lost customer. CallRail's 2025 benchmark of 1.1 million leads found that home services businesses miss 14% of their calls, and that up to 85% of callers who are not answered do not call back. An AI answering service is the cheapest way to make sure every call gets picked up.
Most pages that rank for this topic are written by a provider and end with their own product at number one. This one is about the decisions you have to make whichever provider you pick, including the parts that go wrong.
How an AI answering service works
Under the hood, every call goes through the same loop, many times a second:
- Your number forwards the call to the service, or the service gives you a new number.
- Speech recognition turns what the caller says into text.
- A language model decides what to say or do next, using your instructions and the information you gave it.
- A synthetic voice speaks the reply, and the agent can take actions: book a slot, send a text, transfer the call, log a message.
- After the call you get a summary, a transcript and usually a recording.
The quality you hear depends on two things you control: how good your instructions and information are, and how clearly you defined what the agent may and may not do. The rest of this guide is about those two things.
What it costs: three billing models, one worked example
Headline prices are hard to compare, because providers count different things. Published plans from the providers' own pages, checked on 18 October 2026:
| Plan | Price per month | Billing unit and what is included |
|---|---|---|
| Smith.ai AI, Free | $0 | Per call: 25 calls, then $3.00 per call |
| Smith.ai AI, Pro | $150 | Per call: 75 calls, then $2.50 per call |
| Goodcall Starter | $79 | Per unique caller: 100 unique callers, then $0.79 each |
| Goodcall Growth | $129 | Per unique caller: 500 unique callers, then $0.26 each |
| Rosie Professional | $49 | Per minute: 250 minutes |
| Rosie Scale | $149 | Per minute: 1,000 minutes |
Which model suits you depends on your calls:
- Many short calls from different people: per-minute plans tend to win.
- Long calls, such as quotes or intake: per-call or per-caller plans tend to win.
- Lots of repeat callers, such as customers checking on a job: per-unique-caller plans tend to win.
- Spam: ask how it is counted. Smith.ai says known spam numbers are filtered before the AI answers and do not count; Goodcall says spam and wrong numbers do not count as customers.
Where AI answering fails, and how to cover it
Every provider demo sounds smooth. Real calls are messier. These are the places AI answering most often goes wrong, and the cheap fix for each.
| What goes wrong | Why | How to cover it |
|---|---|---|
| Wrong phone number or address in the message | Digits and street names are easy to mishear | Make the agent read back every number, name and address, and text the caller a confirmation |
| Caller talks over the agent | Turn-taking on the phone is hard for software | Keep the agent's answers short; long scripted paragraphs invite interruptions |
| Noisy line: a job site, a car, a crying baby | Speech recognition degrades with background noise | If the agent fails to understand twice, it should offer a text-back or a transfer instead of guessing |
| Confident wrong answer | The model fills gaps when your information is missing | Tell it in writing to say "I'll have someone confirm that" for anything not in your knowledge sheet |
| Price or arrival-time promises | Callers push for a number | Ban quotes and time promises explicitly; give it the exact sentence to use instead |
| Caller who wants a human and keeps asking | The agent is told to resolve calls itself | Make "talk to a person" an instant transfer or a guaranteed call-back |
None of these is a reason to avoid AI answering. They are reasons to read the first week of transcripts yourself. Ten minutes a day for a week will show you which of these your callers actually trigger.
The knowledge sheet: what to give the AI before day one
Most providers say the agent can learn from your website. That is a start, but websites rarely say what a receptionist needs to know. Before you go live, write one page with:
- Business name exactly as it should be said, plus the greeting.
- Opening hours, service area (cities or ZIP codes) and what you do not serve.
- Your services in plain words, and the jobs you do not take.
- What to say about price: a call-out fee if you publish one, otherwise the exact sentence for "someone will confirm the price".
- The questions to ask every new lead, in order (name, number, address, what they need, how urgent).
- Booking rules: which slots it may offer, how long each job type takes, how much notice you need.
- Your ten most common questions with the answers you would give.
- What counts as urgent for your business, and who gets called.
- Things it must never do: promise arrival times, diagnose problems, give safety advice beyond "call 911".
Update the sheet every time a transcript shows a question it could not answer. After a month, the agent knows your business far better than your website does.
When the AI should hand the call to a person
An AI agent should know when to stop. Write these rules down and give them to the provider:
| Situation | What the agent should do | Why |
|---|---|---|
| Caller asks for a person | Transfer during hours; after hours, promise a call-back by a set time | Callers who are refused a person hang up and call a competitor |
| Urgent problem (your own list) | Reach the on-call person, then the backup, then flag it at the top of the summary | A single unanswered attempt is not cover |
| Upset or confused caller | Apologise once, stop the script, transfer or arrange a call-back | Scripts make upset callers angrier |
| Question not in the knowledge sheet | Take a message; do not guess | A wrong answer costs more than a call-back |
| Gas smell, smoke, fire or medical emergency | Tell the caller to hang up and call 911 or the gas utility first | Safety comes before booking the job |
For a full escalation ladder with time limits, see our after-hours answering service guide. For the vendor questions to ask about transfers and hand-offs, see the AI receptionist buyer's guide.
Recording consent and telling callers it is AI
AI answering services record and transcribe calls by default. That creates two questions most buying guides do not mention.
1. Recording consent
US states differ. Most follow a one-party consent rule, but in a group of states every party to the call must agree to a recording. A July 2024 survey by the law firm Kilpatrick lists 13: California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Oregon, Pennsylvania and Washington. When the caller and the business are in different states, the safer approach is to follow the stricter rule.
The practical fix is simple: start every call with a short notice, such as "This call may be recorded so we can help you." Check that your provider lets you add it to the greeting.
2. Telling callers it is AI
In February 2024 the FCC ruled that AI-generated voices count as "artificial" voices under the Telephone Consumer Protection Act. That ruling is about outgoing calls, such as robocalls, which need the called person's prior consent. An answering service that picks up calls customers make to you is a different situation, but if your agent also makes outgoing calls, such as reminders or follow-ups, those rules apply.
Separately from the law, saying "I'm the virtual assistant for [business]" in the greeting tends to help: callers adjust how they speak, and anyone who wants a person can ask for one straight away.
AI or a live answering service?
| Question | AI answering | Live answering service |
|---|---|---|
| Typical starting price | $0 to $150 a month on the plans above | $250 to $300 a month on published entry plans (Ruby, Smith.ai Starter) |
| Night and weekend calls | Same price as daytime | Same on 24/7 plans; check yours |
| Repeat questions and bookings | Strong and consistent | Strong |
| Upset callers and unusual problems | Weak; should hand over | Better |
| Many calls at once | Answers all of them | Depends on staffing |
For most small businesses, AI answering with a clear handover rule is the best value, and a live service is worth paying for when a large share of calls are stressed, complex or high-value. A hybrid, AI first with a person behind it, gives you most of both. Trade-specific comparisons: plumbing, HVAC and contractors.
Setting it up in an afternoon
- Pull last month's call log: number of calls, distinct callers, average length, share after hours.
- Price that pattern on at least two billing models, using the worked example above.
- Write the knowledge sheet and the handover rules.
- Add the recording notice and, if you choose, the AI disclosure to the greeting.
- Set call forwarding: all calls, after hours only, or when nobody answers. You keep your number.
- Run test calls yourself: a new lead, a booking, a noisy line, a misspelled street name, a caller demanding a person, an urgent call.
- Read every transcript for the first week and update the sheet.
If you only want to stop losing callers who hang up, start with missed call text back. If you want calls answered, booked and handed over by your own rules, that is what we build: our AI receptionist answers in your business name, the AI voice agent handles the conversation and transfers, and the appointment booking agent puts jobs straight into your calendar.
Frequently Asked Questions
It is software that answers your business phone with a voice agent instead of a person. It greets callers in your business name, answers common questions from information you give it, takes messages, books appointments and transfers calls to your team when your rules say so. It works around the clock, usually by forwarding your existing number to it.
On published plans checked in October 2026: Smith.ai AI Receptionist has a free plan with 25 calls a month and a $150 Pro plan with 75 calls; Goodcall starts at $79 a month for 100 unique callers; Rosie starts at $49 a month for 250 minutes. The same call volume can cost very different amounts because providers bill per call, per unique caller or per minute.
Often, yes, especially on longer or unusual calls. Many businesses choose to say so in the greeting. It sets expectations, and callers who want a person can ask for one straight away. Check the rules where you operate, because disclosure and recording requirements vary.
Recording rules depend on the state. In a group of US states every party must consent to a recording, so the safe default is a short notice at the start of every call, such as "this call may be recorded". This article is not legal advice; ask a lawyer if you operate in several states or a regulated field.
For most small businesses it replaces voicemail and the overflow, not the receptionist. It is strongest on repeat questions, bookings and after-hours calls, and weakest on upset callers, complex jobs and anything that needs judgment. Those calls should go to a person.
Sources
- CallRail releases benchmark report (missed call rates by industry) — CallRail, 15 January 2025
- Smith.ai AI receptionist pricing — retrieved 18 October 2026
- Goodcall pricing — retrieved 18 October 2026
- Rosie pricing — retrieved 18 October 2026
- Smith.ai virtual receptionist pricing (live plans) — retrieved 17 October 2026
- Ruby pricing (live plans) — retrieved 17 October 2026
- Wiretap Laws in the United States — Kilpatrick, 29 July 2024
- FCC: AI-Generated Robocalls Illegal Under the TCPA — Cooley, 15 February 2024