Is an AI receptionist worth it for a small business? A decision guide with checked numbers

An AI receptionist is software that picks up inbound calls, speaks in a generated voice, and follows a scripted job: answer common questions, qualify the caller, book into a calendar, transfer to a person, or text a summary to the owner. It is not a full replacement for someone who can negotiate, diagnose a complex job, or calm an angry customer.

Where it tends to work: after-hours overflow, busy periods when nobody can pick up, and high-frequency questions that already have written answers on your website or intake form. Where it fails: callers who are upset, jobs that need a site visit before any quote, edge cases the vendor never trained on, and anything that requires a privacy or clinical judgment call.

Cleland & Co. does not resell AI receptionist software. This guide is how to decide, test, and buy one without treating the brochure as evidence.

Vendor pages for this category often open with the same scare numbers. Some have a narrow origin. Several do not. Use the table as a filter, then measure your own line.

Measure your own missed calls before you believe anyone else's. A 2016 sample of 85 businesses is not your Tuesday afternoon.

Speed still matters when a caller or form lead is waiting. The MIT/InsideSales 2007 study and the 2011 HBR audit both point at faster first contact raising the odds of a real conversation — not at a guaranteed booked job. For a fuller reading of those studies and the common misquotes, see the Cleland & Co. post on speed-to-lead statistics (speed-to-lead statistics for local service businesses).

Fit signals: you already lose live answers during jobs or after hours; a large share of calls are routine; ticket size is high enough that recovering a few missed jobs per month would matter; and you can name a person who will review transcripts and take transfers the same day.

Poor fit signals: almost every call needs custom scoping; callers expect a known person by name; you cannot write down hours, service area, and booking rules clearly enough for software to follow; or you have no human fallback when the system is stuck.

• Contractors and HVAC: strong candidate for after-hours and overflow when the crew is on a job. Keep emergency and complex estimate calls on a fast transfer path.
• Salons and appointment businesses: strong when booking rules are fixed. Weak when every request is a negotiation.
• Dental and medical offices: possible for hours, directions, and non-clinical scheduling only. Patient privacy and confidentiality obligations apply; get counsel before recording or automating clinical intake — this post is not compliance advice.
• Restaurants: useful for hours, wait questions, and simple reservations if your process is already scripted. Poor when the line is mostly custom catering or conflict resolution.

Ignore the vendor calculator until you have a week or a month of your own call log.

• Export inbound call detail from your carrier or phone system for the last 30 days: time, answered or missed, duration, and voicemail if available.
• Count missed calls by hour and by day. Separate business hours from after hours.
• Note how long it usually takes you to call back, and how often a missed caller becomes a booked job when you do reach them.
• Estimate only with your numbers: missed calls that look like real jobs × your close rate on callbacks × your average job value. Leave the scary industry averages out of the sheet.
• Decide what share of those missed calls are routine enough for software versus judgment calls that must reach a person.

If you cannot get a 30-day export, keep a paper or spreadsheet log for two weeks: date, time, answered yes/no, and whether the call would have been bookable from a short script. Writing it down changes behavior, so treat the first few days as noisy.

Treat the pilot as a rehearsal, not a demo. The same pattern that stalls broader AI work applies here: if success was never falsifiable, nobody knows when to keep or kill the tool. That logic is spelled out in why AI pilots stall after the demo.

Write these down before the vendor turns anything on:

• Scope: after-hours and overflow only for the first 30 days. Do not put every daytime call on the system until after-hours and overflow pass.
• Correct booking rate: pass if at least X% of bookings the system claims match the calendar and the caller's request on a weekly audit of recordings or transcripts. Set X from what you would accept from a junior human; if you cannot name X, you are not ready.
• Transfer accuracy: pass if urgent or off-script calls reach a human within your chosen time window, and fail if more than Y% of transfers drop or misroute.
• Caller complaints: fail the pilot if complaint volume about the phone experience rises above a pre-agreed count, or if you hear a pattern of callers hanging up when they realize it is automated and you have no recovery path.
• Owner time: pass only if the hours you spend fixing bad bookings, listening to failed calls, and rewriting prompts stay below a weekly cap you set on day one.
• Human fallback: name the person who takes transfers, reviews the daily summary, and can disable the system without waiting on the vendor.
• Stop condition: if any fail metric trips for two consecutive weeks, turn the system off for daytime overflow and keep only a narrower after-hours path — or cancel.

A pilot that only ever sees quiet Tuesdays has not been tested. Include a busy day, a holiday week if one falls in the window, and at least one deliberately awkward test call.

No vendor names and no price league tables. Ask every candidate the same questions and write the answers next to each other.

• Can a caller reach a human on demand, and how is that transfer warmed (context passed) versus cold?
• Does it book into the calendar you already use, with rules for buffers, staff, and service types?
• Are calls recorded? What consent notice plays, and have you checked the recording laws for the states you serve?
• How long are recordings, transcripts, and caller data retained, who can access them, and how do you delete them on request?
• What does the system say when it does not know — transfer, take a message, or invent an answer?
• Is pricing per minute, per seat, per call, or flat with overages? Model your own volume; do not trust a generic savings claim.
• If you leave, can you export recordings and transcripts, and how fast does the number and greeting revert?

Google may call the number on your Business Profile on a customer's behalf to book, check wait times, or confirm price and availability, and it may also call to verify business details. Those calls are documented in Google Business Profile Help; availability varies by region, and several US states are excluded for some features. Before you buy, have the vendor demonstrate what happens when an automated Google-style caller reaches the AI receptionist: does it transfer cleanly, book correctly, or loop? Details and scam distinctions are in why Google may be calling your business.

This section is informational, not legal advice. Confirm current rules with counsel before you enable outbound AI voice follow-ups.

On February 8, 2024, the FCC released Declaratory Ruling FCC 24-17 confirming that AI technologies that generate human voices count as an "artificial or prerecorded voice" under the Telephone Consumer Protection Act. Outbound calls that use those voices generally need prior express consent of the called party (written consent when the call is telemarketing), plus identification and disclosure information for the initiating party, and opt-out methods when the message is an advertisement or telemarketing. The ruling does not ban AI voice calls outright; it places them under the existing TCPA framework for artificial or prerecorded voice.

Separately, on August 7, 2024 the FCC adopted a Notice of Proposed Rulemaking (FCC 24-84), published in the Federal Register on September 10, 2024 (2024-19028), that would define "AI-generated call" and add further consent and beginning-of-call disclosure requirements. As of 2026-09-26, that proposal had not been adopted as a final federal rule. Existing TCPA obligations from the February 2024 ruling still apply to covered outbound AI voice calls.

Inbound answering is a different fact pattern from outbound AI follow-up. If a vendor pitches "the AI will call every missed lead back," treat outbound consent, identification, and opt-out as a gate, not a feature checkbox.

An AI receptionist is one way to cover the phone. Sometimes a simpler fix is enough:

• Missed-call text-back so the caller gets a same-minute reply with a booking link.
• A shorter voicemail greeting and a same-day callback routine with an owner of the queue.
• A live answering service for overflow only.
• Online booking that removes the need for routine scheduling calls.

If the real problem is slow follow-up on form leads rather than missed rings, start with response-time discipline and the evidence in the Cleland & Co. speed-to-lead statistics post before you buy voice automation. Phone coverage also sits next to how customers find you in AI answers; see AI SEO for small business for the broader operating picture.