AI receptionist: what it does on a call, the rules to check and how to launch one safely

What an AI receptionist does when the phone rings

An ai receptionist is software that answers a business phone line in a synthetic voice, holds a spoken conversation and acts on what it hears. For a dental practice, a salon, a law office or a plumbing firm, the jobs are those of a front desk: greet the caller, find out what they need, answer routine questions, book a slot, take a message and pass the call to a human when it should. It picks up at eleven at night and while three other lines ring, but it knows nothing you have not written down for it.

Take a call to a dental office: someone asks to come in on Thursday with a broken filling. The agent asks about pain or swelling, checks the calendar, offers two free times, confirms the name and number, books one and sends a text confirmation. If the caller mentions a swollen face and a fever, the rules send the call to the person on duty instead. That branch, not the voice, decides whether the system helps or harms.

How an AI phone receptionist works under the hood

Five parts work on every call. A carrier supplies the line, routed over SIP, the protocol of internet telephony. Speech recognition, which Wikipedia describes as methods that translate spoken language into text or other interpretable forms, turns the voice into text. A language model reads that text with your instructions and decides what to say and do, a speech synthesiser reads the reply aloud, and tool calls reach your calendar or CRM.

In OpenAI's telephony documentation, a carrier's SIP trunk is pointed at the provider, a webhook tells your server a call is coming in, and the server accepts or rejects it with the instructions and voice to use. Separate commands transfer the call to another phone address or hang up. The same page limits ringing to 3 minutes and a connected call to 2 hours; these are one provider's rules, and they change.

Quality depends on the written instruction more than on the voice: opening hours, the prices the agent may quote, its policies and the topics it must refuse. Tools should be narrow: find slots, book, leave a message, transfer. A tool that can edit the whole calendar can also empty it after one misheard sentence.

Where a voice agent on the phone still goes wrong

Hearing comes first. Wikipedia's speech recognition article notes that speech is distorted by background noise, echoes and recording characteristics, that accuracy varies with accent and pronunciation, and that telephone speech is typically captured at 8,000 samples per second. Errors are counted as the word error rate: substitutions, deletions and insertions divided by the words actually said. One wrong digit in a phone number loses a customer, so the agent should read numbers back and confirm by text.

Judgement comes second. Asked about something missing from its instructions, a model can give a fluent, wrong answer: a price, a guarantee, a medical opinion. The rule must say: when the answer is not in the knowledge, admit it and arrange a callback. Angry callers, complaints and legal threats go to a person, and emergencies are never triaged: the agent tells the caller to dial the local emergency number.

Telling callers they are speaking to an AI voice

In the EU, the European Commission's FAQ on Article 50 of the AI Act says providers of AI systems that interact directly with people must design them so that people are informed they are dealing with an AI system, unless this is obvious. People must be told from the start of the first interaction, the obvious-case exception is read restrictively, and AI agents are covered. The Commission states that Article 50 applies from 2 August 2026, and in practice it lands in your greeting.

In the United States, the Federal Communications Commission's Declaratory Ruling FCC 24-17, adopted on 2 February 2024, confirms that the Telephone Consumer Protection Act's restrictions on artificial or prerecorded voice cover AI technologies that generate human voices, so such calls need the called party's prior express consent unless an emergency purpose or exemption applies. The messages must identify who is calling and, for telemarketing, offer an opt-out. The ruling concerns AI that initiates outbound calls, not the answering of incoming ones, but callbacks and reminder calls by the same system fall under it. The UK, the Gulf and other regions set their own rules; check your case with a specialist.

Call recording consent and health data on the line

Most systems record and transcribe calls. The US federal rule, published by Cornell's Legal Information Institute as 18 U.S.C. 2511(2)(d), lets a private person intercept a call they are a party to, or where one party has given prior consent, unless the purpose is criminal or tortious. Wikipedia's survey lists California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania and Washington among the states that require the consent of all parties. It adds that in Kearney v. Salomon Smith Barney, California's Supreme Court held a one-party-state caller recording someone in California to the Californian rule.

In the UK, the Information Commissioner's Office says in its monitoring guidance that recording all calls is not usually proportionate, while evidence of transactions or training and quality control can justify it. It says you must tell callers that you are recording and why, that a recorded message is good practice, and that the rest of the privacy information can follow by email or a website link. The ICO marks this guidance as under review after the Data (Use and Access) Act. In Germany, Wikipedia notes, recording a telephone call without the consent of the parties is a criminal offence.

For US clinics, under 45 CFR 160.103, also published by Cornell's Legal Information Institute, a business associate is anyone who, on behalf of a covered entity, creates, receives, maintains or transmits protected health information for a regulated function. A voice service that hears symptoms and stores transcripts for a dental practice fits that on a plain reading, so ask the provider about a business associate agreement first. Have a specialist confirm how HIPAA applies to your setup.

What an AI receptionist costs and how the bill is built

The bill has three layers: setup, telephony (the number and the carrier's minutes) and model fees for hearing, thinking and speaking. A packaged service folds them into a monthly plan with included minutes, which suits steady volume, while paying per minute suits a seasonal business. A custom build pays each provider directly. Count the calls you miss today and what a booking is worth; that is the figure to beat.

OpenAI's guide to managing realtime costs shows why length matters: its voice sessions are billed per second for as long as the session is open, so idle calls should be closed. Its speech-to-speech models are billed by tokens instead: caller audio at one token per 100 milliseconds, the assistant's audio at one per 50 milliseconds. The guide says the entire conversation is sent to the model for each response, so later turns in a long call cost more than early ones, and a long consultation belongs with a person.

Launching an AI receptionist with 30 test calls

Start with the script: a greeting that says the caller is speaking to an AI assistant and that the call is recorded. Then come the ten questions callers ask most, taken from last month's calls, with exact answers. List what the agent may do (book, move, cancel, take a message) and what it must never do: quote unlisted prices, give medical or legal advice, promise refunds. The escalation rules name who receives a transfer, during which hours, and what happens when nobody picks up.

Then run at least 30 test calls before any customer hears it: a noisy street, a strong accent, an angry customer, an emergency, an unusual surname spelled aloud. Score each on outcome, booking and transfer, then fix and repeat the failures. After launch, compare the missed-call rate and the share of transfers with the month before, and read transcripts weekly for the first month. Keep a human backup anyone can switch on in one step, such as forwarding the line to a mobile.

Off-the-shelf service or a custom voice agent: what to prepare

A packaged ai virtual receptionist starts fast, but it is limited to its integrations, its vendor's rules and the place it keeps your recordings. A custom agent runs on a model provider with your own instructions, tools and data path, and needs maintenance. The deciding questions: is your booking system supported, how unusual are your rules, and does health data pass through the calls.

Either way, prepare common questions and answers, quotable prices, calendar access, escalation contacts, a greeting with both notices, your callers' countries and 30 test scenarios. VITON13 Studio builds this as a fixed-scope project: a written job description of the agent, up to three actions in tools you approve, 30 test cases it must pass, and escalation rules that hand anything uncertain to a person. The service page lists it from $73 with AI assistance or $93 human-led, in 5–8 working days, with two revision rounds; model and hosting fees are paid to the provider. Scope, revisions and dates go into the estimate, so the cost of an ai receptionist is known before work starts.