Someone calls at 17:12. The office has just closed, the receptionist has gone home, and the voicemail asks them to leave a message after the tone.
They hang up. Another lead has likely been lost.
That is the kind of call an AI receptionist should catch. Answer the phone, find out what the caller needs, and give somebody enough information to pick it up in the morning. The same job appears at 10.30am when the receptionist is already on another call. It appears again in a small company or a trade that cannot afford a receptionist at all.
Either way, the job is to catch a call that a person cannot take and prepare a useful handover. Running the entire front desk, on the other hand, demands much more.
Current voice software cannot do that reliably.
Give it a job it can actually do
Give it the evening calls, the overflow, and the first few questions before a person takes over.
It can ask what the call concerns, spot an urgent request, answer a small set of questions with agreed answers, take an appointment request, and send a summary to the right person. For a company that cannot yet justify a full-time receptionist, that covers a real gap. For a busy receptionist, it removes calls that never needed much judgement.
That is plenty already. The perfect employee who never sleeps, never makes a mistake and never takes annual leave can stay in the AI sales deck.
A good human receptionist notices hesitation. They recognise the customer who says everything is fine while plainly being furious. They untangle a confused explanation, remember what happened last time, and break the normal process when the situation asks for it.
Today’s voice software handles routine calls. Complaints, confused callers and anything requiring judgement still need a human. Push those through the software and you get a frustrated caller.
Judge an AI receptionist by the calls it saves and the handovers it improves.
A phone call is a terrible form
In ordinary conversation, one wrong word rarely matters. In a CRM, one wrong character sends the email to the wrong person, or nowhere at all.
“John Smith” is the easy case. An unfamiliar surname, a strong regional accent, a bad connection, or a new company name gives the transcription system fewer clues to work with.
Email addresses are worse. They behave more like serial numbers spoken as sentences: dots, dashes, initials, repeated letters, unusual surnames and domains invented by marketing departments. One wrong character gives you a bounce or, occasionally, the wrong person.
Speech-recognition systems choose the most likely transcription from the audio they receive. That is what they are designed to do. Providers offer vocabulary adaptation because uncommon names and specialist terms are harder to recognise without extra context. This helps with a known product list or staff directory. New callers bring names and domains the system has never heard. Google’s Speech-to-Text documentation explains how phrase adaptation influences recognition.
Reading the address back catches some errors. It also creates a nasty failure of its own. The agent confidently repeats the wrong transcription, the caller hears something close enough, and both sides agree on an address that does not exist.
Once the speech engine records the wrong character, post-call extraction copies the mistake into the CRM. Voice platforms can turn a completed conversation into neat fields, but those fields still come from the transcript. Vapi’s structured-output documentation describes this post-call extraction process.
Let the call and the form do different work
We would split the job in two. The call gathers the reason, urgency and preferred time to speak. A short form gathers the characters that must be exact.
An inbound call will usually arrive with the caller’s number in its metadata. Twilio, for example, passes the originating number straight through to the application handling the call. Its call-handling guide documents the incoming From value.
The AI receptionist can then ask a simple question: “Can I text this number with a secure link for your contact details?”
If the caller agrees, send the link while the call is still active. They type their name and email address, and the form catches missing symbols and obvious mistakes. Meanwhile, the reason for the call is already saved against their phone number.
The handover is simple:
- Answer the call and find out what the caller needs.
- Save the caller number supplied with the call.
- Ask permission to send a short follow-up link.
- Save the call summary immediately under that number.
- Add the caller’s typed details when the form arrives.
If the caller withholds their number, calls from a landline, or wants to use another number, ask for a mobile number. Digits are easier to repeat and check than arbitrary letters. Send the message during the call and ask whether it arrived.
SMS fits this job well. It reaches an ordinary mobile without another app.
Speed beats a polished voice
A lovely voice can still make a dreadful phone call.
Before every reply, the system must decide that the caller has finished, transcribe what they said, generate an answer, turn it into speech and send the audio back. Each step adds delay.
More natural-sounding voice models can take longer to generate speech. Larger language models may handle complicated requests better, but they can also take longer to answer. Network distance, long instructions and slow calendar or CRM lookups add more time. ElevenLabs calls the gap between requesting speech and the first audio playing “time to first audio”. The caller waits longer than that because transcription and response generation happen before the speech request. Its latency guide breaks down the full path.
We treat one second as the warning point. When the silence lasts long enough to feel unexplained, callers repeat themselves, speak over the reply or assume the call has dropped.
Stream the audio, keep the instructions short, use fast models for routine questions and acknowledge the caller before a slow calendar lookup leaves dead air.
Test a batch of calls and listen to the slowest ones. A good average can hide individual calls with awkward two-second pauses.
Give callers an easy exit
Decide in advance when the software should hand over to a human.
Let the agent give opening hours and take simple appointment requests. Complaints, payment disputes, emergencies and callers who keep saying “No, that is not what I mean” need a person.
Set up these escape routes:
- Put callers through to a person as soon as they ask, within agreed hours.
- Outside those hours, save the transcript, the summary and the caller’s own message.
- Admit when it has no answer and pass the call on.
- Take callback details when the calendar or CRM is down.
What to decide before installing one
Write the job description before choosing the voice.
Which calls are currently missed? What do people usually want? Which questions have agreed answers? What information must be exact? What should happen when the system is unsure? Who receives the call or summary when it gives up?
Most of these answers already exist somewhere in the business, usually in somebody’s head. Our short diagnostic asks the same kind of questions across the whole operation, not only the phone.
Then handle the edge cases:
- an unusual name and email address;
- a withheld number;
- a caller in a noisy car;
- somebody who changes the subject halfway through;
- an interruption while the agent is speaking;
- a request the knowledge source does not answer;
- a failed CRM or calendar connection;
- a caller who wants a person immediately.
If the design works only when the caller speaks clearly, follows the expected script, and has a Gmail address, it does not work yet.
Common questions
Does an AI receptionist replace a receptionist? Not in any business we would recommend it for. It covers the calls a person cannot take: after hours, during overflow, and the first few questions before somebody picks the call up. A receptionist who is already busy gets a shorter queue, not a redundancy notice.
What does it cost to run? Voice platforms charge by the minute of call time, on top of a small monthly rental for the phone number. That means the bill tracks how much you actually use it, so the honest answer depends on your call volume rather than a headline price.
Can it book appointments? Yes, if it is connected to the calendar you already use. Bear in mind that a slow calendar lookup is one of the more common causes of dead air on the line, so the connection needs to be quick as well as correct.
Should callers be told they are speaking to software? We think so. Callers usually work it out within a sentence or two, and a system that pretends otherwise loses trust exactly when it needs to hand over to a person.
Build something you would use
An AI receptionist is doing its job when it answers calls that would otherwise be missed, handles routine enquiries and gives staff a useful summary when a person needs to take over. That eases the queue for a busy receptionist, covers the evening hours, and gives a small company somebody to answer before it can afford a full-time hire.
That is how we design voice systems. We do not release something we would not use ourselves.
The same approach applies to the rest of our work. SortaSystems builds custom tools, automations and integrations around how a business operates. We are based in Colchester and work with small businesses in Essex and across the UK.