How AI receptionists work, from ring to action

Every call to an AI receptionist passes through six main stages, most of them invisible to the caller. The phone network delivers the call. Speech-to-text turns the caller's words into text. Conversation logic works out what the caller wants and what to say next, using only the facts and rules you supplied. Business tools then act: the calendar, the transfer line, the alert to your phone. Text-to-speech speaks the reply. Last, the system writes a record of what happened. Each of those stages can fail on its own. A pleasant voice proves that two stages work and says nothing about the other four. Those four decide whether the plumber gets the job and whether the law firm hears about a new client before a rival does. A federal regulator saw the same pattern in bank chatbots: useful for basic questions, weaker as problems got complex¹⁰.

Phone network

speech-to-text

conversation logic

business tools

text-to-speech

record

Your customers can keep dialing the number they already know. There are two common setups. The service gives you a new number to publish, or you forward your existing line to it, either all the time or only when nobody picks up. Forwarding is set at your phone carrier, not at the vendor, so ask who sets it up and who fixes it when it breaks. Then ask what happens when three calls arrive in the same minute. A system that answers one and gives the others a busy signal has moved the missed call, not removed it. Check the direction too. Some products, First AI Employee among them, answer inbound calls only².

Number provisioning

call forwarding

simultaneous calls

A caller talks over the greeting to say the water heater is leaking. The assistant has to stop, hear the interruption and judge when the caller has finished. Each step adds delay. Interruption handling needs its own test. FD-Bench, published at Interspeech 2025, evaluated three open-source dialogue systems using 293 simulated conversations and 1,200 interruptions. The systems still struggled with disruptions and noise²¹. That study does not rank today's receptionist services. Test what happens after you interrupt: does the assistant stop, hear the correction and continue from the right detail? Repeat the call with your customers' accents, from a truck, a busy kitchen and a weak signal.

Turn-taking

interruptions

latency

accents and noise

Good answers start with the facts you hand over. Short facts behave well: hours, service area, standard prices, which brands you repair. Long documents behave worse, because the system picks a passage and rewrites it, and a policy written for a careful reader can come out wrong. Answering from your own documents does not make it error-free. In one study, commercial legal research tools that answer from their own libraries still made things up 17% to 33% of the time, on questions far harder than a business FAQ¹³. So ask what it says when a caller raises something you never covered. Taking a message beats guessing an accountant's fee at midnight.

Structured facts

documents

unknown questions

Talking is what vendors demo. Acting is what changes your week. A booking counts only when an event appears on your calendar. First AI Employee, for example, books straight onto a connected Google, Outlook or Calendly calendar, or onto its own built-in calendar¹. Routing either connects the caller to a person or fails. Intake fills the fields you chose, so pick fewer and better ones. Alerts put the result in front of someone who can act. Take care if the system also calls customers back. The FCC confirmed that AI-generated voices count as an artificial voice under the TCPA⁹, so those calls need prior express consent unless an exemption applies²⁰. This is not legal advice.

Booking

routing

intake

notifications

Your first outage will land on a busy day, so decide now what callers hear. The calendar times out. The on-call phone rings out. The system half understands a caller, which is worse than not understanding. Someone asks a vet clinic about a dose or a law firm for advice. A good fallback tells the caller what happens next and puts the request in front of a named person. Two findings shape that design. Bank chatbots can trap people in loops with no way to reach a human¹⁰. And in text-chat studies, a chatbot with a human name and persona lowered satisfaction when customers arrived angry¹⁴. Send upset callers to a person fast.

Calendar down

transfer fails

low confidence

out-of-scope

On Monday a customer disputes the price they heard on Saturday. What the system kept decides if that costs five minutes or an afternoon. A summary is the system's reading of the call, good for sorting work but not proof. A transcript is closer but can hold words nobody said: a 2023 test of one engine, Whisper, partly on speakers with a speech disorder, found about 1% of transcripts held whole invented phrases¹². A recording is the best proof and the most regulated. Federal law lets one party consent to recording¹⁶. About 11 states require every party's consent¹⁷, and California is one of them¹⁵. Ask how long each record is kept and who can export it. This is not legal advice.

Summary

transcript

recording

retention

Follow one evening call to a property manager. The greeting says the caller is talking to an automated assistant. California bans online bots that hide what they are to make a sale, though its reach to phone calls is unsettled¹⁸. Utah requires a business's generative AI to say so when a person clearly asks¹⁹. Disclosure can cost sales: in one outbound experiment in China, it cut purchases by more than 79.7%⁸. Disclose anyway. This is not legal advice. The caller says her tenant has no heat. The system matches that to your urgency rules, checks the address, finds an evening slot and books it. You get the address, callback number, the words no heat, and a calendar event.

What the caller says

what the system checks

what the business receives

Ask these ten questions and get the answers in writing. Who sets up call forwarding? How many calls can it answer at once? Can I hear it handle an interruption and a heavy accent? What does it say when it does not know? Does a booking create a real calendar event? What does a caller hear when the calendar or a transfer fails? Who gets alerted, and how fast? How long are summaries, transcripts and recordings kept? What does the greeting say about being AI? When the software cannot finish a call, who picks up? Answers differ: Smith.ai offers live agents as a per-call add-on³, while Dialzara transfers to your own team⁵.