You build an AI receptionist by wiring a phone line into a real-time voice pipeline (speech-to-text, an LLM and text-to-speech) and giving that pipeline tools that book, look up and transfer inside your own systems. The phone part is easy. The hard part is making it answer in under a second, hand off cleanly and write correct data into your calendar or CRM.
This guide covers the stack layer by layer, what each option really costs in 2026, and a five-question test that tells you whether to buy a subscription or build your own.
What is an AI receptionist?
An AI receptionist is a voice agent that answers your business line, understands what the caller wants and completes or routes the request without a person.
An AI receptionist answers calls in natural speech instead of a keypad menu. It greets the caller, works out the reason for the call, answers questions from your own information, books or changes appointments, takes messages and transfers to a person when the call needs one.
It differs from an IVR in one way that matters: the caller talks the way they would to a person. There's no "press 2 for billing." That only works if the agent responds fast enough to feel like a conversation, which is why the stack below is built around latency.
How does an AI receptionist work? The 6-Layer Front Desk Stack
Every AI receptionist, bought or built, runs on the same six layers; the quality of each layer decides how natural and how useful the calls feel.
We call this the Front Desk Stack. Vendors hide it behind a dashboard, but it's the same six layers underneath every product on the market.
Layer | What it does | Typical components | What breaks when it's weak |
1. Telephony | Gets the call to your agent | Your existing number with call forwarding, a SIP trunk, or a number from the platform | Dropped calls, one-way audio, no caller ID |
2. Media transport | Moves live audio between caller and agent | WebRTC media server such as LiveKit, SIP bridge | Robotic audio, delay that makes callers talk over the agent |
3. Speech in | Turns speech into text and decides when the caller has finished | Streaming STT, voice activity detection, turn detection model | Agent cuts callers off or waits awkwardly |
4. Brain | Decides what to say and what to do | LLM, system prompt, knowledge base, tool calls | Wrong answers, made-up policies, loops |
5. Speech out | Speaks the reply | Streaming TTS, a chosen or cloned voice | Flat or slow voice, mispronounced names |
6. Business systems and handoff | Acts on the call | Calendar, CRM or practice software, SMS, warm transfer, call logs | Double bookings, lost messages, angry transfers |
Layers 1 and 2: getting the call in
Most businesses keep their current number and forward it, either all the time or only after hours and on overflow. A SIP trunk connects the call to the media server, which keeps audio flowing both ways in real time. ITU-T G.114 sets 150 ms one-way as the comfortable limit for a phone conversation, so the media path can't waste time.
Layer 3: listening and knowing when to talk
Streaming speech-to-text transcribes while the caller is still speaking. Turn detection then decides whether a pause means "I'm done" or "let me find my policy number." LiveKit's audio turn detector commits a turn after as little as 0.3 seconds of silence when it's confident, and waits up to 2.5 seconds when it isn't. That single setting decides whether your receptionist feels polite or pushy.
Layer 4: the brain
The LLM gets a system prompt with your hours, services, prices, policies and tone, plus tools it can call: check availability, book a slot, look up an order, transfer to Sarah. Keep the knowledge base short and specific. A receptionist that answers 20 questions well beats one that half-answers 200.
Layer 5: the voice
Streaming text-to-speech starts talking before the full sentence is generated. Pick a voice that fits your brand and test it on your own business names, street names and medical or legal terms.
Layer 6: doing the actual work
This is where AI receptionists earn their cost or lose it. Booking into Google Calendar is easy. Writing a new patient into a practice management system, checking insurance fields, or creating a lead with the right owner in Salesforce takes real integration work. Warm transfer matters just as much: the agent should brief the person picking up so the caller never repeats themselves.
Why AI receptionists matter now: the numbers
A human front desk costs thousands a month for one shift, while AI options cost tens to hundreds of dollars and answer around the clock.
Metric | Value | Source (date) |
Median receptionist wage, US | 38,010ayear(18.27 an hour) | BLS Occupational Outlook (May 2025) |
Receptionist jobs, US | 947,500, projected to fall 2% from 2025 to 2035 | BLS (2025) |
Share of a 168-hour week one full-time hire covers | About 24% (40 hours) | Calculation |
Smith.ai AI Receptionist entry plan | $150 a month for 75 calls | JustCall pricing review (Oct 2026) |
Ruby human receptionists entry plan | $250 a month for 50 minutes | JustCall pricing review (Oct 2026) |
LiveKit Cloud phone agent, default models | About $0.048 a minute | LiveKit pricing calculator (Nov 2026) |
Natural gap between turns in human conversation | Around 200 ms | Stivers et al., PNAS (2009) |
The pattern is clear. A person covers one shift. Software covers every hour, every overflow spike and every holiday, at a fraction of the wage bill.
How much does an AI receptionist cost? Build vs buy at real volume
At 600 three-minute calls a month, buying costs $155 to $1,010, a human costs about $3,168 in wages, and a custom LiveKit build costs about $136 in platform usage plus your build and upkeep.
Price lists are hard to compare because one vendor bills per call, another per minute and another per unique caller. Here's one scenario on equal footing: 600 calls a month, 3 minutes each, so 1,800 minutes.
Option | How it bills | Monthly cost at 600 calls | What you don't control |
Smith.ai AI Receptionist | Per call | About $1,010 | Custom prompting only on Enterprise; vendor says it is not HIPAA compliant |
Goodcall | Per unique caller | About $155 | Prompt depth, integrations beyond its list |
Rosie | Included minutes | $299 (2,000-minute plan) | Booking and transfers only from the $149 plan up |
Ruby (humans) | Per minute | Above its largest published plan (500 minutes for $1,725) | Quality varies by receptionist |
One in-house receptionist | Salary | About $3,168 in wages only, one shift | Coverage after hours, sick days, turnover |
Custom build on LiveKit Cloud | Per minute usage + plan | About $86 usage + $50 Ship plan | Nothing: you own prompts, data, voice and integrations |
Vendor figures come from JustCall's published comparison (prices read Oct 2026, 3-minute average call). LiveKit usage uses its calculator default of $0.0479 a minute before plan allowances. Wages use the BLS May 2025 median.
The custom row looks cheapest, but it leaves out two real costs: building it and running it. A useful formula for your own numbers:
If that total beats your best subscription quote within 12 to 18 months, and you need what the subscription can't do, build. If not, buy.
The 5-Question Build vs Buy Test
Answer five yes/no questions; one or fewer "yes" means buy, two or three means a managed or hybrid build, four or five means build your own.
Does the agent need to write into a system the vendors don't support? Practice management software, a custom ERP, a dispatch board, or a CRM with strict field rules.
Do you have compliance terms a vendor won't sign? A HIPAA BAA, data residency in a specific region, call recordings kept on your own storage, or zero data retention at the model provider.
Do your callers need languages or accents most tools handle poorly? Urdu, Arabic, code-switching like Hinglish, or heavy regional accents.
Is volume high enough that per-call pricing hurts? As a rule of thumb, above 1,000 calls a month the gap between per-call plans and usage pricing gets large.
Is the receptionist part of your product? You resell it, white-label it, or need to own the voice, prompts and data as an asset.
Count the "yes" answers.
0 to 1: Buy. A $50 to $300 subscription will serve you better than an engineering project.
2 to 3: Managed build. Build on an open framework like LiveKit, but let a team that runs voice agents in production own deployment, monitoring and upgrades.
4 to 5: Build your own stack, with in-house or partner engineers who have shipped real-time voice before.
Top AI receptionist options compared
Subscription tools win on speed to launch, human services win on empathy, developer platforms win on flexibility, and custom LiveKit builds win on control and unit cost.
Here's an honest view of the main routes, each with what it does best.
1. Custom build on LiveKit with RTC LEAGUE
An AI receptionist built on LiveKit's open-source Agents framework, connected to your phone line and your own systems. RTC LEAGUE designs, builds and can run the stack on LiveKit Cloud or your own servers, with source code handed over.
Best for: Businesses that answered "yes" to two or more questions above, especially clinics, multi-location brands and companies that need multi lingual or deep system writes.
2. Smith ai
A well-known service with both an AI receptionist and human receptionists, priced per call, with strong legal-market integrations.
Best for: Law firms and professional services that want a human safety net behind the AI.
3. Goodcall
An AI receptionist billed per unique caller, so long calls cost the same as short ones.
Best for: Small businesses with repeat callers and longer calls.
4. Rosie
A low-cost AI answering service that trains itself from your website and Google business profile.
Best for: Solo operators and small shops that mainly need messages taken and questions answered.
5. Ruby
A live virtual receptionist service with real people on every call.
Best for: Businesses where every caller should reach a person and the budget allows per-minute human rates.
6. Developer platforms (Vapi, Retell)
Hosted voice-agent platforms where your team configures prompts, voices and tools through an API and dashboard.
Best for: Product teams with developers who want to ship fast and accept the platform's limits on media control and hosting.
How to build an AI receptionist in 7 steps
Scope the calls first, then build, test with real recordings, and launch on overflow before you route every call.
Pull 50 real calls. Listen and tag each by reason. Most businesses find 5 to 8 reasons cover 80% of calls.
Write the call map. For each reason: what the agent says, what data it needs, what tool it calls and when it transfers.
Set up telephony. Keep your number and forward it, or bring a SIP trunk. Start with after-hours and overflow only.
Build the voice pipeline. STT, turn detection, LLM and TTS in one session. Target a reply within 800 ms of the caller finishing.
Connect the tools. Calendar first, then CRM or practice software, then SMS confirmations and warm transfer.
Test with recordings and simulations. Replay real calls, noisy lines and interruptions. LiveKit now prices audio simulations at $0.05 a turn on paid plans, which is cheap insurance.
Launch in stages. After-hours first, then overflow, then all calls. Review transcripts weekly for the first month.
How to pick the right AI receptionist: a checklist for any vendor
Test speed, accuracy, handoff and compliance on your own calls before you sign anything.
Call it yourself 10 times. Does it reply within a second of you finishing?
Interrupt it mid-sentence. Does it stop and listen?
Ask something it shouldn't know. Does it admit it and offer a transfer, or invent an answer?
Ask for a person. Does the person get a summary before picking up?
Book, then reschedule, then cancel. Does your calendar match every time?
Ask for the BAA or DPA, data retention terms and where recordings are stored.
Price it at your real volume and real call length, not the entry plan.
Final take
If your calls are simple and your volume is under a few hundred a month, buy. Goodcall, Rosie and Smith.ai will get you live this week for less than a day of a receptionist's wages.
If the receptionist has to work inside your systems, meet strict compliance terms, speak your customers' languages or become part of your product, build it on an open stack you own. That's where a custom LiveKit build pays for itself, and where most subscription tools stop.







