Every missed call is a customer deciding whether to try someone else. In US home services, call data shows businesses miss 14–27% of inbound calls, and 85% of those callers never call back (Thoughtly). Dental practices, clinics, law firms and property managers face the same problem in a different shape: the phone rings during a treatment, at 7 p.m. or on Saturday morning, which is precisely when customers have time to call.
An AI receptionist closes that gap. It picks up every call within a ring or two, answers routine questions, books and reschedules appointments directly in your calendar, captures new leads in your CRM and hands anything sensitive to a person together with a short summary. This is no longer an experiment: inbound calls make up 52.1% of the voice AI agent market, which Grand View Research sizes at $3.5 billion in 2026 (Grand View Research).
This guide explains what an AI receptionist actually does, what it costs compared with human options, how to roll it out without irritating your callers, and which mistakes cost the most.
What an AI receptionist does, and what it should not do
Think of it as a front-desk colleague who never takes a break but works strictly from an approved playbook. A production-grade setup usually covers six jobs:
- Answering 24/7 and on overflow. After hours, at lunchtime and when every line is busy, nobody hears a voicemail greeting.
- Answering routine questions from an approved knowledge base: opening hours, address and parking, services, insurance or payment options, preparation instructions.
- Booking, rescheduling and cancelling directly in your calendar, practice-management system or field-service software, with real availability in view.
- Capturing and qualifying leads: name, callback number, reason for the call, urgency and how they found you, written into the CRM rather than onto a sticky note.
- Routing and escalation: emergencies, upset callers, key accounts and anything outside the script go to a person by warm transfer, or as a callback task with the transcript attached.
- Following up by text: booking confirmations, directions, links to intake forms, reminders.
What it should not do matters just as much. It should not invent prices, discounts or policies; hallucinated promises are one of the most common failure modes of voice agents. It should not give medical, legal or financial advice. And it should not try to “save” a call that clearly needs a human. A fast, polite handoff is a feature, not a failure.
| Task | AI handles alone | Human stays involved |
|---|---|---|
| Opening hours, directions, services | Yes | — |
| New booking or reschedule | Yes, within booking rules | Exceptions such as double slots or special equipment |
| Price questions | Only approved price ranges | Custom quotes |
| Emergencies (pain, gas smell, flooding) | Recognise and escalate immediately | Always |
| Complaints and billing disputes | Capture details, create a task | Always |
| Existing clients asking about their case | Basic status, if integrated | Anything that requires judgement |
How it works under the hood
A typical AI receptionist is a pipeline of four parts. Your phone number or SIP trunk delivers the audio; a speech-to-text model transcribes the caller; a large language model decides what to say and which tools to use; and a text-to-speech voice speaks the answer. Newer speech-to-speech models such as OpenAI’s Realtime API merge the middle steps, but they still need telephony, orchestration and integrations around them (OpenAI pricing).
Two technical facts shape what callers experience:
- Latency. In human conversation the gap between turns is around 200 milliseconds (Stivers et al., PNAS). Well-tuned cascaded voice agents respond in roughly one second (DevToolLab). That is acceptable, but every extra system lookup during the conversation adds delay, so the design of the tools matters.
- Integrations. The tools, such as “check availability”, “create booking”, “look up customer” and “create callback task”, are what make the receptionist useful. An agent that cannot see your calendar is a voicemail with good manners.
Quality is the concern buyers raise most often. In LangChain’s 2026 survey of teams building agents, quality and reliability were the top barrier (32%) and latency came second (about 20%) (State of Agent Engineering). Both are engineering problems with known remedies, namely testing, guardrails and monitoring, rather than reasons to wait.
What it costs: AI versus human options
The prices below are published list prices or figures from the cited sources. Check current vendor pages before you sign anything.
| Option | Typical price | What you get |
|---|---|---|
| In-house receptionist | $48,000–63,000 per year (AIRA) | One person, business hours, holidays and sick days to cover |
| Live answering service | $235–1,640 per month (AIRA) | Human agents, usually priced per call or per minute |
| Smith.ai human receptionists | $300/month for 30 calls, then $8.50–11.50 per call (Smith.ai) | Managed human service |
| Budget AI receptionist (Rosie) | $49 / $149 / $299 per month for 250 / 1,000 / 2,000 minutes (Rosie) | Self-serve setup |
| Managed AI receptionist (Smith.ai AI) | From $150/month for 75–300 calls, about $1.67–2.00 per call; Enterprise $500 (Smith.ai) | Managed AI with human backup options |
| Custom-integrated voice agent | Roughly $0.08–0.31 per minute in platform, model and telephony fees (Vapi, Retell AI), plus setup and ongoing care | Your scripts, your integrations, your data |
The raw telecom layer is cheap: a US number on Twilio costs $1.15 per month, inbound minutes $0.0085 and call recording $0.0025 per minute (Twilio). In a custom build you pay mainly for integration work, testing and maintenance.
When is a $49 tool enough? If you mostly need a polite message-taker after hours and booking happens elsewhere, start there. When do you need an integrated agent? When the value lies in booking directly into a practice-management or field-service system, checking customer records, triaging emergencies and reporting on outcomes, and when you work in several languages or under EU data rules. We compare the build platforms in detail in Vapi vs Retell vs ElevenLabs vs Bland.
A five-minute ROI sanity check
You can estimate the money at stake quickly. Multiply monthly inbound calls by your missed-call rate, by the share of callers who will not call back, by your booking rate and by the average value of a booking. Here is a worked example; replace every input with your own figures:
| Input | Example value |
|---|---|
| Inbound calls per month | 500 |
| Missed-call rate | 15% → 75 missed calls |
| Callers who never call back | 85% → about 64 lost callers |
| Share who would have booked | 40% → about 25 bookings |
| Average value per booking | $300 |
| Revenue at stake per month | about $7,600 |
Even if the AI recovers only half of that, the comparison with a running cost of a few hundred dollars a month is easy to make. For a fuller model that includes staff time and churn, read how to calculate the ROI of AI automation.
Where AI receptionists pay off fastest
- Dental practices and clinics: booking, rescheduling, reminders and insurance questions. In the US, patient data means you need a HIPAA-ready setup and a business associate agreement; Vapi, for example, sells HIPAA support as a $2,000/month add-on (Vapi).
- Home services (HVAC, plumbing, electrical, roofing): emergency triage and job booking. After deploying voice agents through ServiceTitan, Riley Plumbing raised the share of calls that turned into bookings from under 50% to 80% (ServiceTitan).
- Law firms: intake and qualification of new matters and appointment setting, strictly without legal advice.
- Real estate and property management: viewing requests, maintenance tickets and tenant questions.
Venture investment points the same way: among Y Combinator voice-agent startups, 69% sell to businesses and 18% to healthcare (a16z).
Step-by-step rollout
“A voice bot in ten minutes” is a demo, not a deployment. A production rollout takes weeks because the real work lies in integrations, testing and edge cases. A sequence that works:
- Audit your calls. Listen to 50–100 recent recordings and sort them: new bookings, existing customers, routine questions, sales, emergencies, spam. This tells you what to automate first and what success looks like.
- Define scope and handoff rules. Write down what the agent may do alone, what it must escalate and to whom, including phone numbers and hours.
- Build the knowledge base and guardrails. Approved answers, price ranges, booking rules (durations, buffers, service areas) and a list of forbidden topics.
- Integrate. Calendar or practice-management system, CRM or field-service software, SMS. Every call should end with a structured record: who called, why, what happened and what happens next.
- Write the opening line. Tell callers they are speaking with an AI assistant and, if you record, say so. In the EU this has been mandatory under Article 50 of the AI Act since 2 August 2026 (AI Act, Article 50). Our guide to legal AI phone calls covers consent, disclosure and recording in detail.
- Test with a fixed evaluation set. Run at least 50 scripted test calls: accents, noisy lines, spelling of names, phone numbers and dates, interruptions, long silences, callers who change their mind.
- Launch in stages. Start with after-hours and overflow, then extend to daytime once the numbers hold.
- Review weekly. Listen to a sample of calls, read the failed transcripts and update the knowledge base. Most long-term quality comes from this habit.
KPIs to track from day one
| Metric | What it tells you |
|---|---|
| Answer rate | Share of calls picked up; it should approach 100% |
| Booking rate | Share of booking-intent calls that end with an appointment |
| Containment rate | Calls resolved without a human |
| Transfer success | Escalations that actually reached a person |
| Early hang-ups (under 15 seconds) | Callers rejecting the experience; check the greeting and latency |
| Data accuracy | Names, numbers and dates captured correctly |
| Cost per booked appointment | The number to compare with your previous setup |
Eight questions to ask any provider
Whether you buy a ready-made product or commission a custom build, these questions separate a demo from a system you can run your front desk on:
| Question | Why it matters |
|---|---|
| Which calendar, CRM or practice-management systems do you write to directly? | Without a write integration, someone still re-types every booking |
| How does the agent hand over to a person, and what does the person see? | A warm transfer with a summary saves the caller from repeating everything |
| What happens when two calls arrive at once, or twenty? | Concurrency limits decide whether peaks are covered |
| Where are recordings and transcripts stored, and for how long? | GDPR, HIPAA and your own retention policy depend on it |
| Can we export prompts, configuration and call data if we leave? | You should own what you paid to build |
| How are changes tested before they go live? | A fixed test set prevents a fix in one place from breaking another |
| Which languages are supported on the phone, not just in text? | Phone audio is harder than chat; test your customers’ languages |
| What is billed: minutes, calls, attempts, transfers, add-ons? | The headline per-minute price is rarely the full price |
The export question deserves special attention. Prompts, call flows and evaluation sets are real intellectual property, and some no-code builders make them hard to take with you. Make sure the contract states that they belong to you and can be exported in a usable format.
Common mistakes
- Choosing by the demo voice. A beautiful voice that cannot book into your system loses to a plain one that can.
- No escalation path. A caller stuck in a loop with a bot is worse off than one who reached voicemail.
- Letting the agent improvise prices and policies. Restrict pricing answers to approved ranges and say so in the instructions.
- Hiding that it is an AI. In the EU it is now a legal obligation; everywhere else it is a question of trust.
- Skipping data-capture tests. Mistakes in names, addresses, numbers and dates are among the most frequent failures. Test them deliberately and have the agent read critical details back to the caller.
- Ignoring concurrency. Platforms limit simultaneous calls: Retell includes 20 concurrent calls and charges $8 for each additional one, and Vapi charges $10/month per extra line (Retell AI, Vapi). A Monday-morning peak can exceed a small plan.
- Set and forget. Without ongoing review, quality drifts as your services, prices and staff change. Our guide to AI quality control for 100% of calls shows how to monitor human and AI calls with the same pipeline.
Conclusion
An AI receptionist is one of the best-proven uses of AI for small and mid-sized businesses. The problem, missed calls, is measurable; the cost per minute is low; and the result, booked appointments, shows up in the calendar within weeks. What separates a gimmick from a working system comes down to three things: integration with the tools where your bookings live, clear escalation rules, and a weekly quality review.
If you want to know whether it would pay off in your business, we can look at your call volumes and booking flow together. Book a free 15-minute AI audit, or see how we build voice agents for inbound calls.