How to Create an AI Receptionist for Free: 3 Honest Paths, and Exactly Where $0 Ends

An AI receptionist is a voice agent that answers your business line, greets callers, answers questions from a knowledge base you give it, books appointments into your calendar, takes messages, and routes anything it can't handle to a human. The practical difference from a human answering service, as Voiceflow's product page frames it, is that an AI agent picks up instantly and can take every call at once — there is no hold queue at lunch hour and no shift change at 5 p.m.

Free in this market means three different things, and every page ranking for this query quietly picks just one. Free-to-test: a vendor trial that expires in days. Free tier: a developer platform like Vapi that costs nothing to prototype on but meters real usage by the minute. Free software: open source, where the code costs nothing and the phone calls still don't. Knowing which free you are being offered is most of the work of not getting surprised by a bill.

Free builds make sense when the problem is real but the budget isn't yet. If you are a solo operator missing calls while you're on jobs, an appointment business losing after-hours bookings to voicemail, or a founder validating whether your callers will even talk to an AI before paying for one, the free paths below are a legitimate way to find out. The Medium tutorial that ranks first aims at exactly this crowd — consultants, coaches, and clinics tired of phone tag.

Some businesses should skip free entirely. Healthcare practices whose callers discuss patient information cannot use free tiers at all, because no free tier signs a HIPAA business associate agreement. High-volume lines — say 30-plus calls a day — blow through free ceilings in days and turn into unmonitored metered bills. Multi-location businesses need routing and reporting that free stacks don't have. And if a single missed call is worth hundreds of dollars to you, the weeks you'd spend tinkering cost more than a managed service. One more honest note: if you get three calls a week, you may need nothing at all — voicemail plus disciplined fast callbacks still works.

Here is the comparison none of the top three results makes, because each covers exactly one path: the Medium post teaches only the DIY stack, My AI Front Desk's tutorial walks only its own signup, and GetCallAgent's guide covers only paid SaaS trials. You have three genuinely different routes, with different ceilings, skill demands, and failure modes.

• ›Cost-at-volume figures are illustrative estimates assembled from the pricing anchors cited throughout this article, assuming a 3-minute average call. Platform pricing changes; verify current rates before committing.

Every AI receptionist call runs four meters at once: telephony (carrier minutes plus number rental), speech-to-text to transcribe the caller, a language model to decide what to say, and text-to-speech to say it. Free tiers bundle all four into a trial credit; when the credit runs out, each meter bills per minute. getcallagent.com lists Vapi at about $0.05 per minute — treat that figure as a partial anchor, because on developer voice platforms the model and voice costs are typically metered separately from the platform's own fee, so the all-in rate runs higher.

The strictly-$0 boundary is sharper than any competitor admits: web test calls inside a platform dashboard cost nothing, and that is where free truly lives. The free cliff is the day you attach a dialable phone number. Numbers rent for roughly a dollar or two a month, minutes start metering, and on the automation side, Make.com's free plan caps at 1,000 operations a month as of 2026. A booked call consumes three to five operations — webhook in, availability check, event creation, response out — so the free plan supports very roughly 200-300 booked calls a month before it, too, demands an upgrade.

• ›Scale check: at a 3-minute average, 100 calls a month is 300 minutes, 300 calls is 900 minutes, and 1,000 calls is 3,000 minutes. Even modest all-in per-minute rates push that last figure into hundreds of dollars a month — the zone where flat-rate pricing with no per-minute meter starts beating DIY. Treat these as illustrative estimates, not quotes.

GetCallAgent's guide deserves credit as the only competitor with a prerequisites checklist. Here is the version that prevents the failures we see most often. Gather all of it before you open a single dashboard:

• ›A one-page business fact sheet: hours, address, services, the prices you are willing to have quoted aloud, payment types, parking, and your cancellation policy. My AI Front Desk lets you paste up to 20 pages of business information; one accurate page beats twenty sloppy ones.
• ›Your top ten caller questions, each with an answer you would stand behind on a recorded line.
• ›The Google account that owns the calendar you actually keep — not a spare one.
• ›About 100 minutes of focused time, plus a second phone to make test calls from.
• ›Your escalation plan: the human number that receives transfers, and what should happen when that person doesn't pick up.
• ›A decision on your AI disclosure line. We strongly recommend using one — see the compliance section for why.

This is Path B in full — the stack the ranking Medium tutorial sketches in 490 words, rebuilt with the steps it skips: availability checking, the phone number, and real testing. Budget about 100 minutes end to end. Each step lists its time and the artifact you should have before moving on.

• ›Step 1 — Create a Vapi account and a new assistant (10 minutes). Sign up free at vapi.ai, create an assistant, give it a name and a voice, and place a test call from the browser. Artifact: an assistant that answers a web test call, even if it says nothing useful yet.
• ›Step 2 — Paste the system prompt and your business facts (15 minutes). Use the starter prompt in the next section, then append your one-page fact sheet. Keep it tight — long prompts slow responses and invite rambling. Artifact: an agent that answers your top ten caller questions correctly in web test calls.
• ›Step 3 — Define a booking tool that posts to a webhook (10 minutes). In Vapi, add a custom tool the agent calls when a caller wants an appointment, with parameters for full name, callback number, service, date, and time. Make the date parameter require the complete date including the year — the one genuinely useful tip in the Medium tutorial, and it's real: relative dates like 'next Friday' are a common failure mode. Artifact: a tool that fires a POST request with a clean JSON payload.
• ›Step 4 — Build the Make.com scenario (15 minutes). On Make's free plan, create a scenario starting with a Webhook module, copy its URL into your Vapi tool, then add a Google Calendar create-event module and map name, service, date, and time into the event. Run it once with a sample payload before trusting it. Artifact: a test webhook call that lands an event on your calendar.
• ›Step 5 — Add availability checking (15 minutes). Insert a Google Calendar search-events step before the create step; if the requested slot is busy, return a JSON response telling the agent the time is taken so it offers alternatives. The Medium tutorial explicitly defers this step, which means its build can double-book. Do not skip it. Artifact: a test call where the agent declines a busy slot and proposes another.
• ›Step 6 — Run ten web test calls (15 minutes). Happy-path booking, an ambiguous date, a price that's in your facts, a price that isn't, a transfer request, an angry caller. Fix prompt and payload issues as you go. Artifact: a written test log — every call, what broke, what you changed.
• ›Step 7 — Attach a phone number (10 minutes, and the end of strictly $0). Rent a number through the platform, typically a dollar or two a month, or import one from a telephony provider. Trial telephony accounts often restrict which numbers you can call and may play a trial notice before connecting. Artifact: a real number that rings your agent from any phone.
• ›Step 8 — Forward your business line, conditionally first (10 minutes). Use your carrier's no-answer or busy forwarding so the AI only catches calls you miss; move to unconditional forwarding (*72 with many US carriers, per getcallagent.com) only after a week of clean transcripts. Artifact: a live missed-call safety net on your real business number.

The entire prompt-engineering guidance on page one of Google today is a single bullet telling you to make the prompt friendly and focused on booking. Here is an actual starter prompt. Replace the bracketed values, paste it into your assistant, and edit from there — and read it aloud once first, because your agent will.

You are Maya, the virtual receptionist for [BUSINESS NAME], a [BUSINESS TYPE] in [CITY]. Your job is to answer questions using only the business information below, book appointments, take messages, and transfer urgent calls. Personality: warm, brief, professional. Keep every reply under two sentences unless you are reading back details.

Disclosure: after your opening greeting, always say: 'Just so you know, I'm an AI assistant — I can book you in, answer questions, or get you to a real person.' If a caller asks whether they are speaking with a machine, answer honestly.

Booking flow: collect, in this order: full name; callback number, then read it back digit by digit; the service they want; preferred date and time — always confirm the complete date, including the year, before saving. Check availability with your booking tool. If the slot is taken, offer the two nearest open slots. Read the entire booking back and get a clear yes before confirming.

Guardrails: never quote a price that is not in your business information — say 'I don't want to give you a wrong number, so I'll have someone confirm pricing.' Never give medical, legal, or financial advice. If the caller describes an emergency, say 'That sounds urgent — I'm connecting you to someone right now' and transfer immediately to [ESCALATION NUMBER]. If you don't know something, say so and offer a message or a transfer. Do not discuss other customers, staff matters, or anything unrelated to [BUSINESS NAME].

Two design notes. The disclosure line is there on purpose: transparency is good manners, regulators increasingly expect it, and callers handle it fine when it's delivered confidently. And the no-unlisted-prices guardrail exists because price hallucination is among the most expensive failure modes an unsupervised receptionist has.

The phone number is the step the top-ranking Medium tutorial skips entirely — its build never explains how a real business line reaches the agent. GetCallAgent covers the basics; here is the operator's version.

• ›Never port your main business line into a tool you are merely trialing. Porting moves ownership of the number, takes days, and is painful to undo; forwarding gives you the same behavior and reverses in seconds.
• ›If your agent sends outbound confirmations from a fresh VoIP number, expect some Spam Likely labeling at first. As of 2026, US carriers use the STIR/SHAKEN framework to attest caller identity, and unregistered new numbers start with no reputation — register the number and your caller-ID name (CNAM) with your provider if outbound matters.
• ›If you ever replace a physical line entirely with VoIP, confirm emergency-call routing. An AI receptionist should never stand between a caller and 911.

Path A — free-trial SaaS — is the right answer for plenty of readers, so here is what it actually looks like. My AI Front Desk's own ranking tutorial walks its signup flow: pick a business type, choose from 11 caller languages, write a greeting (their advice that shorter greeting phrases do better is correct and applies on every platform), paste up to 20 pages of business information, then configure text-during-call links and call-transfer scenarios. They claim setup takes under 10 minutes, and a test call is genuinely free.

The catch across this category: free means free-to-test. That walkthrough never states what the product costs once you subscribe, and per getcallagent.com, trials in this market run 7 to 14 days on plans of roughly $29 to $97 a month. Check the pricing page before you spend an afternoon training a knowledge base you might abandon — and keep a copy of everything you write, because your fact sheet and prompt port to any platform, including the DIY stack above.

Path C — open source — is the only path where the software is permanently free. You stitch together a SIP trunk for telephony, an open speech-to-text model such as Whisper, a language model, and an open text-to-speech voice, hosted on your own server. The honest accounting: telephony and hosting still cost money, and the engineering time is real — getting response latency low enough that conversation feels natural is the hard part, and tuning it can take weeks. Choose this path because you are a developer with data-control requirements or a genuine desire to learn voice AI, not because it's cheaper. Counting your hours, it usually isn't.

No competitor on page one writes this section, and it is the one that can hurt you. None of this is legal advice; it is a map of where the lines sit as of 2026.

AI disclosure. Have your agent say it's an AI at the start of every call — our starter prompt builds it in. Some states have bot-disclosure rules in specific contexts, and the regulatory direction is clearly toward transparency. Beyond the law, it's cheap insurance: callers who discover mid-call that they were fooled trust you less than callers told up front.

Call-recording consent. Federal law sets a one-party baseline, but as of 2026 roughly a dozen states — including California, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Pennsylvania, and Washington — require all parties to consent. Since you can't control where callers are standing, the safe pattern is a recording notice at the top of every call. Free tiers record by default and will not warn you about any of this.

TCPA and outbound calls. Answering inbound calls carries little TCPA risk. The moment your agent dials out — missed-call callbacks, appointment reminders — TCPA applies, and the FCC's February 2024 declaratory ruling confirmed that AI-generated voices count as artificial voices under the statute, so prior-express-consent rules apply to AI calls. Free DIY stacks ship zero consent tooling. Read /blog/tcpa-outbound-calling before any outbound use.

HIPAA. If callers discuss patient information, every vendor in your audio chain must sign a business associate agreement. Free tiers and trials do not sign BAAs — which means there is no compliant free path for a patient-facing healthcare line, full stop. The same caution applies to payments: do not let a free build collect card numbers.

Both transcripts below are illustrative — composites of common patterns, not recordings of real customers. For real production calls, listen at /call-recordings.

A good call. Agent: 'Thanks for calling Lakeside Physio, this is Maya. Just so you know, I'm an AI assistant — I can book you in or get you to a real person. How can I help?' Caller: 'I need an appointment Friday.' Agent: 'Happy to — just to confirm, that's Friday, June 19, 2026?' Caller: 'Yes, morning if you have it.' Agent: 'I have 9:30 or 11:00 open. Which works better?' Caller: '9:30.' Agent: 'Great — can I get your full name and best callback number?' The agent reads the number back digit by digit, then: 'You're booked for 9:30 on Friday, June 19, and you'll get a text confirmation in a minute. Anything else?'

Everything load-bearing happened in small moments: the disclosure in the greeting, the full-date echo, offering only genuinely open slots, the digit-by-digit read-back, and the final confirmation before saving.

A failed call. Caller: 'How much is a deep-tissue session?' Agent: 'A deep-tissue session is $99.' That price is not in the knowledge base — the agent invented it. Caller: 'Fine, book me next Friday.' Agent: 'You're booked for next Friday.' Which Friday? Nobody knows, and the event lands on the wrong week. Two configuration fixes prevent both failures: the guardrail forbidding unlisted prices, and the rule requiring the complete date with the year before saving. This is why the testing step is not optional.

Across all three ranking pages, total troubleshooting coverage is six bullets. Here are the twelve failure modes that actually take down free builds, with fixes.

• ›Awkward two-second pauses. Latency kills trust. Pick faster model and voice settings, shorten the prompt, keep replies brief — and test from a real phone, not the web demo.
• ›Hallucinated prices and policies. Add the only-from-business-info guardrail with an 'I'll have someone confirm' fallback, and audit transcripts weekly.
• ›Double-booking. Add the availability search before event creation (Step 5). The ranking Medium build defers it, so it can double-book.
• ›Relative-date misparsing. Require the complete date including the year before saving — the Medium tutorial's one real tip, made mandatory here.
• ›Webhook errors. Make.com must return valid JSON to the agent; test the scenario with sample payloads before live calls, as the Medium post also warns.
• ›Expired calendar authorization. Google OAuth lapses silently and bookings stop syncing. Re-authorize; getcallagent.com also notes Workspace admins can block third-party calendar access.
• ›Callers talking over the agent. Enable interruption (barge-in) handling and shorten the greeting — myaifrontdesk.com's advice that shorter greetings do better holds up.
• ›Accents and background noise. Test with realistic audio, and rely on read-backs for critical fields like phone numbers and dates.
• ›Voicemail loops on transfer. If the transfer target doesn't answer, the agent should take a message — not bounce the caller back and forth.
• ›Emergency misrouting. Define explicit emergency trigger words that force an immediate transfer. Never let an AI handle a true emergency.
• ›Forwarding never actually enabled. Verify with your carrier (*72 typically, per getcallagent.com) and confirm by calling your own business number.
• ›Trial credits running out mid-week. The agent simply stops answering. Set a billing alert or a recurring reminder to check the credit balance before it checks you.

Before forwarding a real line, run the 15-call test script below from a phone, not the dashboard — phone audio surfaces the latency and background-noise problems that web tests hide. Log every call: the scenario, what happened, and what you changed afterward.

Once the line is live, watch five numbers weekly. Answer rate: should be effectively 100 percent — this is software, so anything less means forwarding broke, credits ran out, or the platform went down. Containment rate: the share of calls fully handled without a human; treat 60 to 80 percent as a reasonable starting target for routine receptionist traffic — that is our recommended working range, not a benchmark from a study. Booking conversion: bookings divided by calls that wanted one.

Transfer accuracy: did transferred calls reach the right person, with enough context that the caller didn't have to repeat themselves? Hang-up rate: callers abandoning mid-call — when it rises, suspect latency or a prompt that has grown too long. None of the three ranking guides offers any measurement framework at all, and an unmeasured receptionist is just a liability with a nice voice.

• ›Happy-path booking with a specific date and time
• ›An ambiguous date ('next Friday') — does it confirm the full date?
• ›A same-day appointment request
• ›A slot you know is busy — does it offer alternatives instead of double-booking?
• ›A price that is in your fact sheet
• ›A price that isn't — does it refuse to guess?
• ›A service you don't offer
• ›An angry caller demanding a manager
• ›An immediate 'let me talk to a human' request
• ›A rambling caller who buries the request mid-story
• ›A call from a noisy environment — run it from your car
• ›A second language, if you enabled one
• ›A wrong number — does it end the call gracefully?
• ›An emergency phrase — does it transfer instantly?
• ›Hanging up mid-booking — does a half-collected booking pollute your calendar?

Run the break-even math before defaulting to free. A careful DIY build takes two to six hours, and maintenance — reading transcripts, fixing prompts, re-authorizing integrations, watching credit balances — runs roughly two to four hours a month. Price your time honestly: at $75 an hour, three monthly hours is $225 of attention before any metered minutes. The formula: monthly maintenance hours times your hourly value, plus metered costs, versus a managed flat rate. When the left side is smaller, DIY. When it isn't, stop tinkering.

Free is flatly the wrong answer when callers discuss patient information (no BAA means no path); when you are in a regulated industry with recording-retention rules; when you take more than about 30 to 50 calls a day and the meters plus QA burden compound; when you run multiple locations; when you want outbound campaigns, because free tiers ship no TCPA consent tooling; or when a single missed call is worth hundreds of dollars and weeks of tinkering cost more than a year of software.

And sometimes the honest answer is no AI at all. At a few calls a week, voicemail plus fast callbacks is free and fine. If your calls demand judgment and empathy — complex sales, distressed callers — a good human answering service still beats any AI receptionist, free or paid.

MapleVoice is not free, and we won't pretend otherwise. It is a done-for-you service: we build the agent, tune it to your industry using playbooks for 20 verticals, connect your booking system, CRM, or POS, test it, and take it live in about 48 hours — at a flat monthly price with no per-minute meter, published at /pricing. The agent answers 24/7 in under two seconds, books appointments, qualifies leads, takes orders, and transfers to a human with full context. For qualifying healthcare customers we sign a BAA, outbound runs with TCPA controls, and every call produces a recording, transcript, summary, call reason, outcome, and next step.

Who should still DIY? Builders who enjoy this, businesses with genuinely low call volume, pre-revenue founders, and anyone who wants to learn voice AI by operating it — the guide above is everything you need, sincerely. A reasonable middle path: run the free build for a month to validate that your callers will talk to an AI, then read /blog/how-to-set-up-an-ai-receptionist to see what the managed version of the same project looks like.

The ranking guides use these terms without defining them — or skip them entirely. Precision saves debugging time.

• ›Containment rate — the share of calls fully resolved by the AI without a human touching them.
• ›Warm vs. cold transfer — a warm transfer passes context and an introduction to the human; a cold transfer just rings through.
• ›STT / TTS — speech-to-text turns the caller's voice into text; text-to-speech turns the agent's reply into voice.
• ›Barge-in — the caller interrupts the agent mid-sentence and it stops talking to listen.
• ›SIP trunk — an internet phone line that connects voice software to the public telephone network.
• ›Number porting — transferring ownership of your phone number to a new provider; 24-48 hours, per getcallagent.com.
• ›CNAM — the caller-ID name displayed when you call out.
• ›Webhook — a URL one system calls to push data to another; it is how Vapi reaches Make.com.
• ›BAA — business associate agreement, the HIPAA contract a vendor must sign before touching patient information.
• ›STIR/SHAKEN — the caller-ID authentication framework US carriers use to flag spoofed and spam calls.

Frequently asked questions

The “How to…” series

Ten hands-on playbooks — real steps, real numbers, honest about the work involved.

• How voice agents work →
• Set up an AI receptionist →
• What it really costs →
• Build your own agent →
• How answering services work →
• Automate phone calls →
• Stop missing calls →
• Choose a service →
• Train your agent →

Keep reading

• How to Set Up an AI Receptionist (the Managed Path) →
• How Much Does an AI Receptionist Cost? →
• MapleVoice Pricing — Flat Monthly, No Per-Minute Meter →