An AI receptionist for small business is a voice agent that answers your phone line, books appointments into your calendar, takes messages and routes urgent calls to a person, 24 hours a day. At current usage rates, the raw model and phone costs for 600 three-minute calls a month come to roughly $50 to $115, against about $4,500 a month for one full-time human receptionist in the US. The real decision is not the price per minute. It is whether the agent can book cleanly into your systems, hand off at the right moment and tell callers it is an AI where the law requires it.
Small businesses lose calls at lunch, after 5 pm, on weekends and when two lines ring at once. A full-time receptionist covers about 40 of the 168 hours in a week; the rest goes to voicemail.
This guide covers the monthly maths, what the technology can and cannot do, calendar and CRM booking, disclosure rules across six markets and setups for six verticals.
Key Takeaways
- Usage is cheap, integration is not. Model and telephony costs run about 3 to 6 US cents per minute at October 2026 rates; the work and the risk sit in calendar, CRM and handoff logic.
- Think coverage, not replacement. One human covers about 24 percent of the week. The best-performing setups use AI for after-hours, overflow and routine calls, and keep a person for judgment calls.
- Disclosure is now law in several places. EU AI Act Article 50 applies from 2 August 2026, Maine requires a clear notice when a bot could pass as human, and the FCC treats AI voices as "artificial" for outbound calls.
- Vertical rules matter more than features. A dental office needs a HIPAA business associate agreement; an HVAC line needs a hard-coded emergency script; a law firm needs intake that never gives advice.
- Build when your booking system is unusual. Buy a packaged service for a standard calendar; build a custom agent when you need deep integration with practice, field service or reservation software.
What Is an AI Receptionist for Small Business?
An AI receptionist is software that picks up a phone call, listens, understands what the caller wants and responds in a natural voice. Modern versions run on speech-to-speech models, so the caller can interrupt, change their mind or ask a follow-up question without pressing keys.
A typical call passes through four layers:
- Telephony. A phone number from a carrier or cloud provider forwards the call into the agent.
- The voice model. It converts speech to meaning, decides what to say and speaks the reply.
- Tools. The model calls functions such as "check availability", "book slot", "create lead" or "transfer to human".
- Your systems. Calendar, CRM, practice management or field service software, where the booking actually lands.
A phone tree asks the caller to fit its menu; an AI receptionist fits the caller's sentence to your systems. It is a narrower job than a full agentic AI system for customer service, which can refund, cancel and change accounts. A receptionist mostly answers, books, captures and routes.
What It Can and Cannot Do Today
Voice models have improved fast, but they still have clear limits.
| Task | AI handles it well | Keep a human, or add a guardrail |
|---|---|---|
| Hours, location, parking, prices on your list | Yes, from an approved knowledge base | Prices that change daily |
| New appointment or reservation | Yes, if it can read and write your calendar | Complex multi-resource bookings |
| Reschedule or cancel | Yes, with caller verification | Late-cancel fee disputes |
| Lead capture (name, number, job details) | Yes, and it never forgets to ask | High-value deals that need rapport |
| Call routing and warm transfer | Yes, by intent and urgency | When nobody is available to take it |
| Upset or grieving callers | Poorly | Always route to a person |
| Medical, legal or safety advice | No | Hard-coded refusal and transfer |
| Payments over the phone | Avoid | Use a secure payment link by SMS |
The single biggest failure is a dead-end handoff. If the agent says "let me transfer you" and the call rings out, the caller blames your business, not the software. Every setup needs a fallback: a callback promise with a time, an SMS to the owner and a ticket in the CRM.
If the receptionist also answers questions about your services, ground it in a reviewed knowledge base rather than the open web. The same retrieval method that powers a RAG chatbot for enterprise search keeps a phone agent from inventing prices or policies.
How Much Does an AI Receptionist Cost?
There are three ways to pay, and each moves the cost to a different place.
| Pricing model | What you pay | Who carries the usage risk | Fits |
|---|---|---|---|
| Packaged subscription | A flat monthly plan with a minute or call allowance, then overage | Shared | Standard calendars, low setup effort |
| Voice platform plus your own setup | Per-minute platform fee plus model and phone costs | You | Teams with a developer, moderate volume |
| Custom build | A one-off build fee, then raw usage and hosting monthly | You | Unusual systems, several locations, strict data rules |
Packaged plans hide the underlying rates. Knowing them lets you convert any quote into an effective cost per minute.
The raw usage rates, as of October 2026
- Phone line. Twilio charges $0.0085 a minute to receive calls on a US local number and $1.15 a month per number, plus $0.0025 a minute to record.
- Voice model. OpenAI lists gpt-realtime at $32 per million audio input tokens and $64 per million audio output tokens, and gpt-realtime-mini at $10 and $20. Cached input is $0.40 and $0.30 per million.
- Token conversion. OpenAI's Realtime cost guide counts caller audio at 1 token per 100 milliseconds (600 tokens a minute) and assistant audio at 1 token per 50 milliseconds (1,200 tokens a minute). It also notes that each turn re-sends the conversation so far.
Worked example: one three-minute call
Assume the caller speaks for 1.5 minutes, the agent speaks for 1.5 minutes, and the call has about ten turns that re-read roughly 20,000 tokens of earlier context from cache.
| Cost line | gpt-realtime-mini | gpt-realtime |
|---|---|---|
| Caller audio: 900 tokens | 900 x $10/M = $0.009 | 900 x $32/M = $0.029 |
| Agent audio: 1,800 tokens | 1,800 x $20/M = $0.036 | 1,800 x $64/M = $0.115 |
| Cached context: 20,000 tokens | 20,000 x $0.30/M = $0.006 | 20,000 x $0.40/M = $0.008 |
| Phone line and recording: 3 min | 3 x ($0.0085 + $0.0025) = $0.033 | $0.033 |
| Total per call | about $0.084 | about $0.185 |
| 600 calls a month, plus one number | about $52 a month | about $112 a month |
One trap is worth flagging. If the context is not served from cache, those 20,000 tokens bill at the full audio input rate: 20,000 x $32/M = $0.64 per call on the larger model. Long calls and bloated instructions push costs up quickly, so keep system prompts tight and end idle sessions.
These are industry estimates from published rates. Hosting, monitoring and support add a fixed monthly amount on top.
AI Receptionist vs Human Receptionist: The Monthly Maths
A fair comparison prices the human fully, then compares coverage, not just cost.
United States
The US Bureau of Labor Statistics puts the median receptionist wage at $18.27 an hour, or $38,010 a year, as of May 2025. In private industry, wages make up 70.0 percent of total employer compensation costs, with benefits the other 30.0 percent (June 2026).
- Loaded hourly cost: $18.27 / 0.70 = $26.10
- Full-time year: $26.10 x 2,080 hours = $54,288
- Monthly: $54,288 / 12 = $4,524
- Coverage: 40 of 168 weekly hours = 24 percent
United Kingdom
The National Living Wage is £12.71 an hour from 1 April 2026, and employer National Insurance is 15 percent above £5,000 a year. Employers also pay at least 3 percent of qualifying earnings into a workplace pension.
- Salary at 37.5 hours a week: £12.71 x 37.5 x 52 = £24,785
- Employer NI: 15% x (£24,785 minus £5,000) = £2,968
- Pension: 3% x (£24,785 minus £6,240) = £556
- Total: £28,309 a year, about £2,359 a month, at the legal minimum wage
The three options side by side (US, 600 calls a month)
| Option | Monthly cost | Hours covered | Notes |
|---|---|---|---|
| One full-time receptionist | about $4,524 | 40 of 168 | Best for upset callers and judgment |
| Custom AI receptionist | about $52 to $112 usage, plus hosting and the build amortised | 168 of 168 | Example: a $20,000 build over 24 months adds $833 a month |
| Hybrid: human by day, AI after hours and overflow | Human cost plus usage | 168 of 168 | The setup most reviewers report working best |
The value case usually comes from missed calls, not salary. Here is an illustration with assumptions you should replace with your own numbers: if 15 percent of 600 calls go unanswered (90 calls), and 30 percent of those callers would have booked a $150 job, that is 27 lost jobs, or $4,050 a month. Recovering even half of that pays for most AI setups several times over.
How Calendar and CRM Booking Works
Booking is where AI receptionists succeed or fail. A clean flow has six steps:
- Identify intent and service. "I need a cleaning" maps to a 60-minute hygiene slot, not a 30-minute check-up.
- Look up real availability. The agent queries the calendar or scheduling system for open slots that match the service length, staff member and location.
- Offer two or three options. Fewer choices shorten calls and reduce errors.
- Hold the slot. A short reservation prevents two callers taking the same slot at the same moment.
- Confirm and write back. The booking goes into the calendar and the contact record goes into the CRM with the call summary and recording link.
- Send a confirmation. An SMS or email with the time, address and a reschedule link closes the loop.
Three engineering details separate reliable systems from demos:
- Idempotent writes. If the network blips and the agent retries, the system must not create two bookings.
- Caller verification before changes. Rescheduling should require the phone number on file plus one more detail, such as a date of birth or postcode.
- A source of truth. The calendar in your practice or field service software wins. The agent should never keep its own shadow copy.
Google Calendar and Microsoft 365 have mature APIs. Industry software varies, so check early whether your platform exposes a booking API; that answer often decides build versus buy. Multi-step flows like these are what AI agent development services are designed for, because the agent has to act in other systems, not just talk.
Disclosure Rules: What You Must Tell Callers
Rules differ by market and by call direction. Inbound calls (customers ringing you) carry fewer rules than outbound calls (your AI ringing them for reminders or follow-ups).
| Market | Key rule | What it means for an AI receptionist |
|---|---|---|
| United States (federal) | The FCC's February 2024 declaratory ruling treats AI-generated voices as "artificial" voices under the TCPA | Outbound AI calls need the same prior express consent as robocalls |
| Maine | Chapter 241 on communications with consumers via AI bars chatbots that may mislead a reasonable consumer into thinking they are human, unless a clear and conspicuous notice is given | State it is an AI at the start of the call |
| California | The BOT Act, Business and Professions Code section 17941, requires disclosure when a bot is used to mislead people in order to sell | Disclose on sales and lead calls |
| Utah | The Artificial Intelligence Policy Act requires disclosure when a caller asks, and up front in regulated or high-risk interactions | Answer "yes, I am an AI" truthfully every time |
| European Union | AI Act Article 50 requires people to be told they are interacting with an AI system, at the latest at the first interaction, from 2 August 2026 | Disclose in the greeting; the Digital Omnibus on AI, Regulation (EU) 2026/1744, moved high-risk deadlines but not this one |
| United Kingdom | The ICO's PECR guidance on telephone marketing requires specific consent for automated marketing calls | Inbound is fine; outbound automated marketing needs opt-in |
| Canada | The CRTC's rules for automatic dialing-announcing devices ban solicitation calls by these devices without express consent | Outbound AI sales calls need express consent |
| Australia | The ACMA enforces the telemarketing industry standard and Do Not Call rules, including calling hours | Outbound calls must respect hours and identify the business |
| UAE | Cabinet Resolution No. 56 of 2024 on telemarketing requires approval, licensed local numbers and records of marketing calls | Keep AI to inbound service unless outbound is approved |
Call recording is a separate question: several US states require every party's consent to record. The simplest compliant greeting covers both: "Hi, you have reached Smile Dental. I am the practice's AI assistant, and this call is recorded. How can I help?"
This is not legal advice. Have counsel check your script for each state or country you serve.
Vertical Setups That Work
Every industry has one rule that matters more than anything else.
| Vertical | Top call types | Must integrate with | The non-negotiable rule |
|---|---|---|---|
| Medical office | Appointments, refills, test results, directions | Practice management or EHR scheduling | Never give clinical advice; a BAA with every vendor |
| Dental office | New patients, cleanings, emergencies, insurance questions | Dental practice software | Triage pain and swelling to same-day slots |
| Law firm | New matter intake, existing client updates | Legal practice software and conflict check | Intake only, never legal advice |
| Salon or spa | Bookings, rebooking, cancellations | Salon booking platform | Deposits and no-show policy read out at booking |
| HVAC or home services | Breakdowns, quotes, maintenance plans | Field service software | Gas smell or carbon monoxide: tell the caller to leave and call emergency services |
| Restaurant | Reservations, hours, large parties, takeaway | Reservation and POS system | Never confirm a dish is allergen-safe |
Medical and dental offices
A receptionist for a clinic handles protected health information the moment a caller gives a name and a reason for the visit. In the US, every vendor in the chain that touches that data is a business associate under HIPAA, and HHS expects a business associate agreement to be signed before any data flows. Check the telephony provider, the voice model provider and the hosting provider separately. If you are building a patient app around the agent, pair it with a team that has done healthcare app development before.
Law firms
The ABA's Formal Opinion 512 on generative AI tools (July 2024) reminds lawyers that confidentiality, competence and supervision duties apply to AI tools. In practice, the receptionist should collect intake facts, run them past a conflict check before details are discussed, book a consultation and avoid any statement that sounds like advice.
Salons, HVAC and restaurants
Salons gain most from rebooking: offer the next appointment on every call and read out the deposit policy. HVAC lines need a safety script for gas, carbon monoxide and flooding that the model cannot override, with the job sent to the on-call technician by SMS. Restaurants should let the agent take reservations and send takeaway callers an ordering link, with large parties and allergy questions going to a manager.
Build vs Buy: Which One Fits?
| Factor | Buy a packaged AI receptionist | Build a custom AI receptionist |
|---|---|---|
| Time to live | Days | About 10 to 16 weeks for a voice agent with integrations |
| Upfront cost | Low | A project fee (estimate below) |
| Monthly cost | Plan fee plus overage | Raw usage plus hosting and support |
| Integrations | Popular calendars and CRMs | Any system with an API, including industry software |
| Data control | Vendor's cloud and retention terms | Your choice of region, retention and model |
| Custom logic | Limited to the vendor's settings | Any triage, routing or verification rule |
| Best for | One location, standard booking | Several locations, regulated data, unusual systems |
Estimated custom build cost. A voice receptionist with calendar and CRM integration, a reviewed knowledge base, warm transfer, SMS confirmations, a call dashboard and testing typically takes 200 to 480 engineering hours. At $40 to $100 an hour, that is $8,000 to $48,000 as an industry estimate. The main swing factors are the number of systems it books into, the number of languages and how strict your data rules are.
Go-Live Checklist
Use this list before the agent takes a real call.
- Greeting discloses AI and recording in one sentence.
- Scope is written down: what it answers, books and refuses.
- Knowledge base is reviewed by a person and dated.
- Booking is tested for double bookings, retries and time zones.
- Handoff works in and out of hours, with a callback promise if nobody answers.
- Emergency script is hard-coded for your vertical.
- Verification protects reschedules and account details.
- Data terms are checked: region, retention, training use, and a BAA or DPA where needed.
- Security testing covers callers trying to talk the agent into breaking its rules. The controls in this guide to AI agent security apply to voice too.
- Measurement is live: answer rate, booking rate, transfer rate, hang-ups in the first 10 seconds and cost per call.
For the wider business case and the metrics that matter, see these benefits of AI in customer service, with evidence and costs.
What Reviews and Owners Say
Owner reviews and forum threads on AI receptionists tend to agree on a few points. Owners who wanted to stop missing calls are usually happy; owners who expected callers not to notice it is an AI are usually disappointed. The common complaints are robotic greetings that make callers hang up, slow transfers, wrong service lengths in bookings and surprise overage bills.
The fixes are practical: a short, warm greeting that admits it is an assistant, fast handoffs with a callback promise, and reading the overage terms before signing. Weigh reviews from businesses with your call volume and booking software above star ratings.
How Pixel Web Solutions Can Help
Pixel Web Solutions designs and builds custom AI receptionists and voice agents that book directly into the systems a business already runs. Its AI chatbot development services cover voice and phone agents, knowledge grounding, CRM and calendar integration, warm transfer and the disclosure scripts each market needs. The team operates from Ras Al Khaimah, UAE and Madurai, India, and works with clients in the US, UK, Canada, Australia and the Gulf. Custom engineering starts from $40 an hour, and every project begins with a scoping call that maps your call types and booking system. To discuss a receptionist for your business, contact Pixel Web Solutions.
Frequently Asked Questions
How much does an AI receptionist cost per month?
Raw usage for 600 three-minute calls a month costs roughly $52 to $112 at October 2026 rates for model and phone line. Packaged services add their margin, hosting and support into a monthly plan, and custom builds add a one-off project fee, estimated at $8,000 to $48,000.
Is an AI receptionist worth it for a small business?
It usually is when you miss calls after hours or at busy times and each booking is worth more than a few dollars. It is weaker when most calls need judgment, empathy or negotiation, so many businesses run AI for overflow and after-hours and keep a person during the day.
Do I have to tell callers they are speaking to an AI?
In the EU, from 2 August 2026, yes, under AI Act Article 50. In Maine, yes, when callers could think they are talking to a human. In Utah, when asked and in regulated interactions. Elsewhere it is still the safest practice, and outbound AI calls in the US need TCPA consent.
Can an AI receptionist book appointments into my calendar?
Yes, if your calendar or booking software has an API or a supported integration. Google Calendar and Microsoft 365 are straightforward. Practice management, legal, salon, field service and reservation platforms vary, so confirm API access before choosing a product.
Is an AI receptionist HIPAA compliant for a medical or dental office?
Only if every vendor that handles patient data signs a business associate agreement and the setup encrypts, logs and limits access to that data. "HIPAA compliant" is not a certification, so ask each vendor for its BAA and security documentation.
Can an AI receptionist transfer calls to a real person?
Yes. A warm transfer passes the caller and a short summary to a staff member, with a callback promise and an owner alert if nobody answers.
What is the difference between an AI receptionist and an answering service?
An answering service uses human agents who take messages. An AI receptionist answers instantly, around the clock, and books directly into your systems, but handles emotional or unusual calls less well.
An AI receptionist for small business pays off when you treat it as round-the-clock coverage with clean booking and honest disclosure, not as a voice that pretends to be human. Start with your missed-call numbers, check your booking software's API, write the greeting and the emergency script, then choose a packaged service or a custom build based on how unusual your systems are.
