AI Receptionist — Definition, the Three Meters It Is Billed On, and Why the Same 300 Calls Cost $129 on One and $500 on Another

Quick answer: An AI receptionist is a voice bot with a job description. It answers the number on your door or your Google listing, says who it is, listens to the caller in their own words, and then does one of four things: answers from a knowledge base, takes a message and texts it to someone, books a slot in a connected calendar, or transfers the call to a human with what it has learned so far. Under the surface it is the pipeline our voice bot entry describes, speech-to-text, a language model, text-to-speech, on a clock. What this entry adds is the part a buyer meets first and understands last: the meter. Four vendors we read on 3 September 2026 bill the same work three different ways. Retell AI charges by the minute ($0.07 to $0.31, with a calculator default of $0.11). Vapi charges $0.05 a minute for its platform and passes model costs through at cost. Smith.ai charges by the answered call ($2.00 down to $1.67 on its Pro plan). Goodcall charges a flat $79 to $249 a month per agent for a quota of unique callers, with minutes unlimited. Put 300 four-minute calls through each and the bill runs from $129 on Goodcall to $500 on Smith.ai, with Retell at $152 between them and Vapi's $60 platform fee below all three because it excludes every model cost, before anyone has judged whether the calls were handled well. Two of the four pricing pages, Smith.ai's and Goodcall's, argue in their FAQs that per-minute billing is misaligned with the customer; Retell's FAQ answers the per-minute objection head-on; Vapi's says nothing about it. The argument is the most useful thing on those pages.

What the term covers, and what it is sold against

The product is defined by the alternative. A business with one phone line and no one free to answer it loses the call to voicemail, and the caller, who is often a new customer, calls the next listing. Smith.ai's pricing page puts the sales argument in one line: "By the time you call them back, they've likely found someone else." Goodcall builds its entire meter around the caller served rather than the minute spent. Everything an AI receptionist does is downstream of that: it exists to convert a would-be missed call into a message, a booking or a transfer, at any hour, for less than a person.

Four kinds of product carry the label, and the meter usually tells you which one you are looking at.

The first two are products; the third is a kit; the fourth is a feature. A buyer who has "AI receptionist" in a search box usually wants one of the first two and is quoted all four. Our best voice AI platforms list covers the fourth row from our reviews' verification of vendor pricing pages and documentation (it says itself that no phone-line stress testing was performed); this entry reads the first three, because that is where the meters differ.

How a call moves through it

flowchart LR
 A[Caller dials your number] --> B[Telephony<br/>Twilio, Telnyx, SIP, or a vendor-issued number]
 B --> C[Speech-to-text<br/>transcript, one utterance at a time]
 C --> D[Language model<br/>system prompt + knowledge base]
 D --> E{Which job?}
 E -->|Question| F[Answer from knowledge base]
 E -->|Message| G[Capture name, number, reason<br/>send SMS or email to staff]
 E -->|Booking| H[Calendar tool<br/>check availability, hold, confirm]
 E -->|Transfer| I[Ring a person<br/>hand over with a summary]
 F --> J[Text-to-speech<br/>spoken reply]
 G --> J
 H --> J
 J --> C
 I --> K[Call leaves the AI<br/>agent fee stops on some meters]

Figure 1. An inbound call as an AI receptionist handles it. The loop from speech-to-text through the model to text-to-speech repeats on every caller turn; the four jobs on the right are what distinguish a receptionist from a general voice bot. The transfer branch matters to the bill: Retell's FAQ says the agent fee stops once a call is transferred and only the telephony fee continues.

Two things in the diagram decide whether callers stay on the line. The first is the silence timer. On a phone line the platform has to decide when you have finished talking, and it does so with a clock: Twilio's TwiML Gather reference, one common way to wire speech into a call, documents a speechTimeout that when set to auto "stops recognizing speech at the first pause in speech." A caller who pauses to check a date gets talked over; a timer set long enough to avoid that makes every reply feel late. The turn-taking entry covers the trade-off. The second is latency, the gap between the caller's last word and the bot's first. In our own Scenario C test of Voiceflow, a five-turn appointment-confirmation agent on Twilio measured about 1.8 seconds typical and 2.6 seconds at the 95th percentile across 25 test calls; Wati's Astra Voice measured about 750 milliseconds median round trip on short utterances. Both are inside what a caller tolerates. Neither is what a receptionist tolerates in a person, and the difference is why a good AI receptionist fills the gap with a short acknowledgment rather than silence.

The failure that our testing recorded is not the model's. Of Voiceflow's 25 calls, the two that escalated did so because speech recognition heard "no" as "yeah," and recognition accuracy fell from 92 percent on clear audio to 78 percent under simulated line noise. A receptionist that books appointments on a misheard "no" is worse than voicemail, which is why the booking branch in the diagram reads the slot back before it confirms, the pattern our confirmation message entry describes for chat.

The three meters, with the same 300 calls on each

The prices below are from each vendor's pricing page, read on 3 September 2026, monthly billing where a choice exists, and they are a snapshot, not a recommendation. The worked scenario is 300 answered calls a month averaging four minutes, from about 240 distinct phone numbers, which is a plausible month for a clinic, a law office or a home-services firm with one line.

Three observations survive the arithmetic.

The two managed products argue their meter is the honest one, and the argument is about incentives, not price. Smith.ai's FAQ says it bills by the call "because unlike per-minute services, our success is aligned with your success," and that "per-minute services work opposite to your needs; they want to maximize your call length while you want to minimize your call time." Goodcall's FAQ says minutes and tokens "are extremely limited indicators of the actual value provided to you by your agent" and "can create misaligned incentives," and that its unique-customer count means "a single client who calls 10+ times in a month would only count as 1." Retell's FAQ does not argue against anyone; it answers the obvious per-minute question directly: asked whether you are charged during silence or hold, it says "Yes. Billing covers the entire duration of the call because the speech-to-text engine remains active and listening throughout." Vapi's pricing FAQ covers trials, credits and volume discounts and takes no position on meters. None of these is a pricing trick. They are three different theories of what a receptionist call is worth, and you should pick the one that matches your calls.

The meter changes what a bad build costs you. Per minute, a rambling script or a slow model is billed to you at the vendor's rate; a caller left on hold for a transfer is billed too, until the transfer completes. Per call, a spam call or a wrong number costs the same as a booked appointment; Smith.ai's page says every answered call counts toward the quota by default, that it filters calls from "20 million+ known spammers" before the agent picks up, and that customers can strike up to 10 percent of a cycle's calls as spam. Per unique caller, Goodcall says robocalls, blocked numbers and "callers who never say anything to your agent" do not count, and a regular who calls twelve times is one unit. If your calls are short and your callers repeat, per-caller is cheap; if your calls are long and your callers are new, per-minute punishes the length and per-call punishes the volume.

The cheapest number on the table is not a finished product. Vapi's $60 excludes every model cost, and Retell's $152 assumes you write the prompt, wire the calendar and set the transfer rules yourself. Smith.ai's $500 buys a finished receptionist with a Quality Studio for test calls and a live agent network the AI can transfer to, though the page does not say what that transfer costs. Our voice bot buyer's guide says to run the arithmetic at your real volume and again at double; add a third pass at half your volume, because a receptionist's bill is decided as much by how many calls it should not have taken as by how many it did.

The six numbers to hold it to

A receptionist is a support metric problem with a phone attached, and the vocabulary already exists. Ask a vendor for these, in this order, and ask which ones the dashboard actually reports.

1. Answer rate. Calls answered by the agent as a share of calls that reached the number, after spam filtering. This is the number the product is sold on, and the one most dashboards show first. It is also the only one that gets better by doing less.
2. Containment. Calls that ended with the caller's need met with no transfer and no callback, the phone version of the containment rate we track for chat. A receptionist that takes a message on every call has an answer rate of 100 percent and a containment rate of zero.
3. Transfer with context. The share of transferred calls that reached a person with the caller's name, number and reason attached. This is the warm transfer test, and our reviews say it is the half of handoff nobody measures.
4. Booking completion. Bookings the agent confirmed that exist in the calendar and were not cancelled within a day. Read back before confirming; the misheard "no" is a booking error, not a speech error, once it is in the calendar.
5. Message accuracy. Sample twenty messages a week against the recording. Name, number, reason, urgency. This is the check most owners never run and the one that decides whether staff trust the tool.
6. Latency and interruption. Median and 95th-percentile silence before the first word of each reply, and how often the bot talks over the caller. Our Voiceflow figures above are a reference point; ask for the vendor's on a real call, not a demo.

The chatbot first response time and average handle time entries cover the two clocks a support team already runs; a receptionist adds the two above and removes the queue.

The rules that reach it, on both directions of the call

An AI receptionist mostly answers calls, and the well-known US robocall law mostly governs placing them. Keeping the two directions apart is the whole job here.

Outbound. On 2 February 2024 the FCC adopted Declaratory Ruling FCC 24-17, released 8 February and effective on release, which confirms "that the TCPA's restrictions on the use of 'artificial or prerecorded voice' encompass current AI technologies that generate human voices" and that "callers must obtain prior express consent from the called party before making a call that utilizes artificial or prerecorded voice simulated or generated through AI technology." Where "these robocalls introduce an advertisement or contain telemarketing," the ruling says the Commission's rules require "prior express written consent," and it notes that artificial or prerecorded voice messages must "state clearly the identity of the business" at the beginning and, for advertising, offer an opt-out. The Discussion's closing sentence limits its own reach: the requirements apply "to any AI technology that initiates any outbound telephone call." So the moment your receptionist product places a callback, a reminder or a follow-up in a synthetic voice, you are in that text, and the consent has to exist before the dial, "absent an emergency purpose or exemption" in the ruling's own words; it names the Commission's exemptions by rule number and does not enumerate them, and neither do we. The appointment reminder guide covers the consent record for reminders; a voice reminder needs the same record and a stricter one for anything promotional.

Inbound disclosure. No federal rule we read requires an AI receptionist to announce itself to a caller. At least one state does, in a specific way. Utah Code 13-77-103, effective 7 May 2025, says a supplier using generative AI "in connection with a consumer transaction shall disclose to the individual that the individual is interacting with generative artificial intelligence and not a human, if the individual asks or otherwise prompts the supplier about whether artificial intelligence is being used," and 13-77-104 gives a safe harbor to a system that "clearly and conspicuously discloses" at the outset and throughout that it is generative AI, "is not human," or "is an artificial intelligence assistant." Regulated occupations in a "high-risk" interaction, which the chapter defines to include collecting health or financial data or giving medical, legal or financial advice, must disclose "verbally at the start of a verbal interaction." Fines run to $2,500 per violation. Other states have their own texts and we have not read them for this entry; the design answer that satisfies the strictest one we did read costs nothing: open with "This is an AI assistant for [business]" and never deny it when asked.

Recording. Three of the four products in the table document call recording on their pages: Smith.ai lists "Recording, transcription & summaries" on its free plan, Retell's plan table has an "Opt-out Recording & Transcription" row, which reads as on unless turned off, and Goodcall's dashboard shows call recordings to team members. Call-recording consent law is state by state in the US and some states require every party's consent, so the opening line should also say the call is recorded where your state requires it. Health data. A medical or dental front desk that lets the agent take a patient's name, reason for the call and appointment is handling protected health information, and needs a business associate agreement with the vendor. Retell's compliance page, linked from its pricing page, says its BAA is available for self-signing, that "a signed BAA is required before transmitting PHI through Retell," and that "there is no additional fee to sign"; Vapi prices a HIPAA add-on at $2,000 a month on either plan. Same obligation, a $24,000-a-year difference in how it is priced. Our healthcare chatbot guide has the four questions to ask before any of this is switched on.

What our fifteen reviews record about AI receptionists

Nothing under that name. The word "receptionist" appears in none of the fifteen platform reviews. Only Voiceflow's review reports a counted batch of test calls, the 25-call appointment-confirmation run whose latency and recognition figures this entry uses; Wati's reports a latency measurement from calls without a call count. Eleven of the fifteen mention a voice agent, voice bot, phone channel or telephony, and the pattern is a tiering one. Voiceflow ships voice across paid tiers with telephony brought from Twilio; Wati ships Astra Voice with cloning from a 60-second sample on Pro and Business; Chatbase ships voice and native telephony from Standard at $150 a month; BotPenguin ships voice agents from King at $99 a month, which our best-list records as the cheapest self-serve entry in our catalog; Blip ships an enterprise IVR and voice surface used by Brazilian banks; Botpress locks voice behind Enterprise; Intercom has a phone channel with AI voice specifics "not deeply documented" in our reading; Typebot, Landbot and Tars ship none, and Manychat's reviewers ask for one. No review tested message-taking, transfer with context or a booking made by voice, and the search strings in Sources will reproduce that absence. For a receptionist specifically, the dedicated products in the meter table are outside our catalog, and this entry does not rank them.

The one test to run before the trial ends

The voice bot buyer's guide owns the evaluation procedure. The receptionist-specific addition is one afternoon's work: write the four jobs down first (which questions it may answer, what a message must contain, which calendar it books into, who it transfers to and when), then make twenty calls yourself from a noisy place. Say "no" when it expects "yes." Ask whether it is a person. Ask it to spell your name back. Ask for a price it should not know. Score the six numbers above by hand, and walk the transfer path with the person who will receive the calls, because the handoff design guide applies to phones with one change: the receiver has seconds, not a transcript, to catch up.

Related terms

• Voice bot — the pipeline an AI receptionist runs on, the four things called a voice bot, and why voice is harder than chat.
• Interactive voice response — the menu tree the receptionist replaces, and the parts of it that survive.
• Warm transfer — the handoff with context attached, which the receptionist's transfer branch has to reproduce in seconds.
• Turn-taking — the silence timer and interruption problem that decides whether callers stay.
• Chatbot containment rate — the resolution metric that answer rate is not.
• Business associate agreement — the contract a medical front desk needs before a vendor hears a patient's name.

FAQ

What is an AI receptionist?

A voice agent that answers a business's incoming calls and performs front-desk jobs: answering questions from a knowledge base, taking a message and sending it to staff, booking or changing an appointment in a connected calendar, and transferring the call to a person with a summary. Technically it is a voice bot; commercially it is sold against the missed call.

How much does an AI receptionist cost?

It depends on the meter. On 3 September 2026, Retell AI listed $0.07 to $0.31 a minute all-in with a calculator default of $0.11; Vapi listed $0.05 a minute for its platform with model costs passed through; Smith.ai listed $2.00 a call on a $150-a-month plan falling to $1.67 at 300 calls; Goodcall listed $79 to $249 a month per agent for 100 to 500 unique callers with unlimited minutes. Three hundred four-minute calls from 240 callers cost $129 on Goodcall's Growth plan, $152 on Retell's default with a Retell number, and $500 on Smith.ai's Pro plan; the product included is different in each case.

Can an AI receptionist book appointments?

Yes, when it is connected to the calendar and given a tool to check availability, hold a slot and confirm. The failure to design for is a misheard answer: in our Voiceflow test, the two escalations out of 25 calls came from speech recognition hearing "no" as "yeah." Have the agent read the slot back before it confirms, and measure booking completion against the calendar rather than against the transcript.

Do I have to tell callers they are talking to an AI?

Under federal law as we read it, not for an inbound call. Utah requires disclosure if the caller asks and gives a safe harbor to an agent that announces itself at the outset and throughout, and regulated occupations there must disclose verbally at the start of high-risk interactions. Other states have their own rules, which we have not read for this entry. Opening with "This is an AI assistant for [business]" satisfies the strictest text we did read and costs nothing.

Can an AI receptionist call people back?

Most products can, and that is where the FCC's February 2024 ruling applies: an AI-generated voice is an "artificial voice" under the TCPA, so an outbound call needs the called party's prior express consent absent an emergency purpose or an exemption, prior express written consent if it contains telemarketing, and must identify the business at the start. Inbound answering is not covered by that text.

Is an AI receptionist HIPAA compliant?

A product is not compliant on its own; a deployment is. A medical or dental front desk handling patient names and reasons for calling needs a business associate agreement with the vendor. Retell's compliance page offers a self-signed BAA at no additional fee and requires it before any PHI is sent; Vapi prices a HIPAA add-on at $2,000 a month. Our healthcare chatbot guide covers what has to be true before the agent is switched on.

Which chatbot platform in your reviews makes the best AI receptionist?

Our reviews cannot say. None used the term, and only Voiceflow's review reports a counted batch of test calls, on an appointment-confirmation flow rather than an inbound front desk. Seven reviewed platforms record some kind of voice surface, and our best voice AI platforms list ranks five of them (Voiceflow, Chatbase, Intercom, BotPenguin, Botpress) from vendor-page verification rather than phone tests. The dedicated receptionist products in this entry are outside our catalog.

What is the difference between an AI receptionist and an AI answering service?

Mostly the seller's vocabulary. "Answering service" is the older industry's term and usually implies a finished, managed product, often with humans available; "AI receptionist" is used by managed products and self-serve platforms alike. The products behind each label differ mainly in whether a person is ever on the line.

Sources

• Retell AI, AI Phone Agent Pricing (retellai.com/pricing), read 3 September 2026: "$0.07-$0.31 / min for AI Voice Agents"; the calculator's default "Cost Per Minute $0.11" with its components "LLM Cost $0.04/min, Retell Voice Infra $0.055/min, TTS Cost $0.015/min, Telephony Cost $0.00/min"; the component table (Retell Voice Infra $0.055/minute; platform, Cartesia, OpenAI, Minimax and Fish voices $0.015/minute, ElevenLabs $0.040/minute; LLM rows from $0.003 to $0.32/minute; Twilio and Telnyx call rates $0.015/min; Retell phone numbers $2.00/month; concurrency "Free for first 20," $8.00 each after); the plan table's "HIPAA / BAA," "Custom BAA" and "Opt-out Recording & Transcription" rows (the checkmarks that assign them to a plan did not survive our text read) and the "SOC 2 certified, HIPAA-ready" line; and the FAQ answers on transfer ("Once the call is transferred, the AI voice agent fee stops. Only the telephony fee continues"), silence ("Yes. Billing covers the entire duration of the call") and concurrency. The $84 and $372 figures are 1,200 minutes at the two ends of the published range; $152 is 1,200 × ($0.11 + $0.015) + $2. Retell AI, Security and compliance (docs.retellai.com/general/compliance, linked from the pricing page footer), read 3 September 2026: "A signed BAA is required before transmitting PHI through Retell," the BAA "available for self-signing," and "There is no additional fee to sign these agreements."
• Vapi, Pricing (vapi.ai/pricing), read 3 September 2026: Build plan "Calls $0.05 / min," "Model Provider Cost (STT, LLM, TTS) … At cost ($0 if you bring your own API key)," "Call Concurrency 10 included + $10 / line / mo," "Call history 14 days," "Add-on: HIPAA $2000/mo," "Add-on: Zero Data Retention $1000/mo"; Scale plan "Annual Contract," "Volume-based per minute pricing."
• Smith.ai, Plans & Pricing for AI Receptionist (smith.ai/pricing/ai-receptionist), read 3 September 2026: Free "$0 mo 25 calls/mo ($3.00/call after that)"; Pro "$150/mo $2.00/call" with the call-volume steps 75, 150 (−10%) and 300 (−17%) and the feature-guide rates $2.00, $1.80, $1.67 and extra-call rates $2.50, $2.30, $2.17; Enterprise "$500/mo $1.67/call" at 300 calls; the FAQ passages on billing by the call, on spam ("Every call answered by your AI Receptionist counts toward your call quota by default," the 10 percent removal allowance, "20 million+ known spammers"), on transfer to "a live North America-based receptionist," and the sales line "By the time you call them back, they've likely found someone else"; "Recording, transcription & summaries" in the Free plan's feature list. The $500 figure is the price the page displays at the 300-call step; 300 × the rounded $1.67 rate is $501.
• Goodcall, Pricing (goodcall.com/pricing), read 3 September 2026: monthly prices Starter $79, Growth $129, Scale $249 "per agent" with annual prices $66, $108, $208 shown under a "15% OFF" toggle; "unlimited minutes and tokens"; 100, 250 and 500 "unique customers monthly" with "$0.50/customer after"; the FAQ definition of a unique customer (a caller "with a unique phone number that calls and interacts (i.e actually says something to agent and does not quickly hang up)"), the exclusion of "Robo calls, numbers you block, or callers who never say anything," and the incentive passage quoted in the body; the "performance viewer" permission level with access to "call recordings"; the "14 days" free trial in the page description.
• Federal Communications Commission, Declaratory Ruling FCC 24-17, CG Docket No. 23-362, Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts, adopted 2 February 2024, released 8 February 2024, "effective upon release" (¶11); read in the Commission's text release on 3 September 2026. Quoted: ¶2 (the "encompass current AI technologies that generate human voices" confirmation and "absent an emergency purpose or exemption"), ¶5 with its footnote 13 (prior express consent; prior express written consent for advertising or telemarketing, citing 47 CFR 64.1200(a)(2), (3)), ¶6 with its footnote 23 (the 64.1200(b)(1) identification requirement), ¶9 (identification, opt-out, and "any AI technology that initiates any outbound telephone call"). The inbound-answering statement in the body is our reading of that scope sentence and of the TCPA provisions the ruling cites, all of which concern calls made or initiated by the caller; it is not a holding of the ruling.
• Utah Code Title 13, Chapter 77, Generative Artificial Intelligence — Consumer Disclosures and Enforcement, effective 7 May 2025, sections 13-77-101 and 13-77-102 as amended by Chapter 95 of the 2026 General Session, sections 13-77-103 through 13-77-106 as enacted by Chapter 465 of the 2025 General Session; read in the Legislature's code PDF on 3 September 2026. Quoted: 13-77-103(1)(a) and (2)–(3), 13-77-104(1), 13-77-105(4)(a) ($2,500 per violation), and the "high-risk artificial intelligence interaction" definition in 13-77-101(5).
• Twilio, TwiML Voice: Gather reference for the <Gather> verb (page modification date 22 June 2026), read 3 September 2026: "If you set speechTimeout to auto, Twilio stops recognizing speech at the first pause in speech."
• Chatbotscape review corpus (the fifteen platform reviews listed at /reviews), searched 3 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. The absence claim: grep -li 'receptionist' sample-reviews/*-review.md returns 0 files. Voice surface mentions: grep -liE 'voice agent|voice bot|voicebot|phone channel|telephony' sample-reviews/*-review.md returns 11 (blip, botpenguin, botpress, chatbase, intercom, landbot, manychat, tars, typebot, voiceflow, wati); grep -liE 'test calls' sample-reviews/*-review.md returns 1 (voiceflow); grep -liE 'voicemail|take a message|message.taking' sample-reviews/*-review.md returns 0. Figures quoted: voiceflow-review.md Scenario C ("25 test calls," "~1.8 seconds typical, ~2.6 seconds 95th-percentile," "92%" and "78%" recognition, "23 of 25 calls (92%) without escalation," the "no" as "yeah" note); wati-review.md line 211 ("~750ms median round-trip"); chatbase-review.md line 220 (Standard tier, $150/mo); botpenguin-review.md line 223 (King tier); botpress-review.md line 513 (Enterprise-only voice); intercom-review.md line 229; blip-review.md line 210; typebot-review.md line 234; landbot-review.md line 219; tars-review.md line 392; manychat-review.md line 510. The "cheapest self-serve entry" statement is from best/best-voice-ai-platform.md.
• Ahrefs Keywords Explorer, US overview and volume-by-country, queried 3 September 2026 — the demand, difficulty, CPC, global-volume, parent-topic and country-split figures in this entry's keyword note.
• Chatbotscape evaluation methodology, including the six-scenario protocol under which the Voiceflow Scenario C figures were recorded. /methodology (continuously updated).