What happens between the ring and the first sentence
The call arrives over the phone network and enters a chain of three processes that hand off to each other without pause. Speech recognition turns voice into text as the sentence unfolds, without waiting for it to finish. The language model reads that text, matches it against the instructions you wrote and decides what to answer or do. Speech synthesis turns that answer back into a voice. The whole thing starts again at every turn.
What makes the exchange feel natural or painful is the gap between the end of your sentence and the start of its own. Under a second, the conversation feels like a conversation. Over two, the caller thinks the line has dropped and starts repeating, which triggers a fresh analysis and stretches the silence further. A supplier who cannot tell you that delay has probably not measured it.
The fourth element is invisible and decides the rest: end-of-speech detection. It judges whether you have finished your sentence or are drawing breath. Set too short, the machine cuts you off. Set too long, it leaves an awkward gap after every sentence. It is the setting that most clearly separates two products built on the same parts.
1
Pick-up
The call comes in over the phone network.
2
Speech recognition
The voice becomes text, as the sentence unfolds.
3
Language model
The text is matched against your instructions, an action is chosen.
4
Speech synthesis
The answer becomes a voice again.
| Delay | What they perceive | What they do |
|---|---|---|
| Under a second | Nothing out of the ordinary | They carry on |
| One to two seconds | A noticeable gap | They hesitate, sometimes they repeat |
| Over two seconds | A silence that feels like a dropped call | They repeat, or they hang up |
Repeating restarts the analysis from scratch. Past two seconds, the delay feeds itself.
- Streaming synthesis
- ≈ 0.5 s
- Batch synthesis
- ≈ 4 s
Measured on our own agent configuration, not on a market panel. That gap is what ruled batch mode out: it is unusable on the phone. A cascaded chain, where three parts hand off, adds another one to two seconds compared with an end-to-end audio model.
AI phone system, callbot, voicebot, IVR: what these words mean
An interactive voice response system is the oldest of the four and the only one that understands nothing. It reads a menu, waits for a keypress or a word from a closed list, and routes. Its strength is predictability, its weakness is that it asks the caller to file themselves into categories they are discovering at the moment they ring.
Callbot and voicebot mean the same thing in common usage, a program that holds a conversation. The first term comes from customer service, the second from voice, and nothing really separates them. Some vendors say voice agent instead, which is still the same category of product. These words describe a technique, not a use.
AI phone system describes the use rather than the technique: what is being sold is the reception of a company’s calls, with its hours, its services, its calendar and its transfer rules. The difference is not cosmetic. A generic voicebot can converse; a phone system knows who you are, what you sell and who to put the call through to when it cannot answer.
| Term | Understands a free sentence | Acts in a tool | At the end of the chain |
|---|---|---|---|
| Interactive voice response | No, a keypress or a word from a closed list | No | Routes, then voicemail |
| Callbot, voicebot | Yes | Varies by product | Answers, then takes a message |
| Voice agent | Yes | Varies by product | Same category, different label |
| AI phone system | Yes | Yes: calendar, hours, company rules | Books the appointment, or puts the call through |
Callbot and voicebot mean the same thing: the first term comes from customer service, the second from voice.
What an AI phone system still gets wrong
Noise is still the first obstacle. A call made from a building site, a roadside or a restaurant floor during service degrades speech recognition before the language model even comes into play, and no well-written instruction rescues a word that was not heard. Regional and foreign accents pose the same problem, with a real inequality between languages.
Proper nouns come next, and a reception is full of them: a surname, a street, a company name. A word outside everyday vocabulary is transcribed approximately, and the appointment ends up in the name of someone who does not exist. Good products make people spell and read back aloud; it is ugly, it is slow, and it is the only method that holds.
Then there are requests outside the scope. A complaint, a difficult personal situation, a question whose answer was never written: the machine does not guess, and it must not try. The expected behaviour is that it says so and applies your rule. An invented answer on a sensitive subject costs more than an unhandled call.
Four families of failure, and the counter to each.
- Ambient noise, on a site, at the roadside or on the floor during service. Recognition breaks down before the language model even starts. Counter: none in software, the caller has to ring from somewhere quieter.
- Regional and foreign accents, with quality that genuinely varies between languages. Counter: a recognition model trained on telephone audio, not on studio audio.
- Proper nouns, a surname, a street or a company. A word outside everyday vocabulary comes out wrong and the appointment carries the name of somebody who does not exist. Counter: make them spell it, then read it back aloud.
- Requests outside the scope, a complaint or a question whose answer is written nowhere. Counter: say so and apply the transfer rule. Never invent.
When a human takes over, and how
Three situations trigger a hand-over to a human, and they are settled in advance. The caller asks, explicitly or by insisting. The request falls outside the written scope. Or the machine has failed twice in a row to understand, which is the threshold past which a caller hangs up of their own accord.
There are two ways of handing over, and they do not cost you the same. A direct transfer rings a number you name while the caller waits; it assumes somebody picks up, otherwise the call falls back. Taking a message records the subject, the number and the degree of urgency, and you ring back. The first is better when there are people around, the second when there are not.
The question to ask a supplier is about the fallback: what happens if the transfer fails. A chain that ends in voicemail reproduces exactly the problem you set out to solve.
Three situations trigger a hand-over to a human. They are settled before go-live, not after the first call goes wrong.
- 1 The caller asks, explicitly or by insisting.
- 2 The request falls outside the scope you wrote.
- 3 Understanding has failed twice in a row, the threshold past which a caller hangs up of their own accord.
| Method | What it assumes | When to prefer it |
|---|---|---|
| Direct transfer | That somebody picks up at the other end, otherwise the call falls back | When there are people around |
| Taking a message | That you ring back, with the subject and number already recorded | When there is nobody |
The two combine: transfer first, message if nobody picks up after a number of rings you set.
Do callers accept it
The question comes up less than it used to, and comes up differently. What annoys people is not talking to a machine, it is waiting, repeating, navigating a menu and ending on an answerphone. An automated reception that picks up straight away and books an appointment in forty seconds is judged on that result, not on its nature.
Transparency, on the other hand, is not negotiable, and it is also in your interest. A reception that lets people think they are talking to a member of your team exposes you, not the vendor who wrote the sentence. The announcement goes at pick-up, in one short sentence, and experience shows it does not make people hang up: what makes them hang up is finding out mid-exchange.
Acceptance does still vary by trade and by reason for calling. A booking, an opening time, a call-back request go through without friction. An intimate subject, a dispute or difficult news call for a human voice, which is one more reason for the transfer rule to be written before go-live rather than after the first call goes wrong.
| Reason for the call | Goes through without friction | Why |
|---|---|---|
| Booking, appointment | Yes | The result is immediate and the caller can check it |
| Hours, address, availability | Yes | The answer is factual, nobody expects anything else |
| Call-back request | Yes | The caller keeps control of what happens next |
| Complaint, dispute | No | The caller wants to be heard, not processed |
| Intimate subject, difficult news | No | The human voice is part of the answer |
The last two rows are exactly the ones the transfer rule has to cover.
From the answerphone to the AI phone system: what changed
For a long time, an unanswered call had two outcomes: voicemail or nothing. In practice the two are equivalent, because a caller looking for an available professional does not leave a message, they dial the next number. IVR moved the problem without solving it, by adding a wait before the same silence.
What changed is not the synthetic voice, which already existed. It is the understanding of a free sentence and the ability to act in a tool, usually a calendar. A reception that genuinely books the appointment during the call removes the call-back, and the call-back was where half the requests were being lost.
One continuity is worth keeping in mind: these products remain machines that apply instructions. Their quality rests on what you write into them, what you read back and what you correct after listening to a few calls. A phone reception installed and never reviewed decays the way a notice pinned to a wall decays when nobody updates it.
Common questions
- How is it different from an interactive voice response system?
- IVR routes without understanding: it reads a menu and waits for a keypress or a word from a closed list. An AI phone system takes a free sentence, works out the request and acts on it, by booking an appointment or putting the call through.
- What happens if two people ring at the same time?
- Both are answered. It is the most concrete difference from a human reception, where the second caller waits and then dials the next number. How many run in parallel depends on how your line is configured, which is worth checking with the supplier.
- Do I have to change my phone number?
- Not in most cases. A forward from your current line is enough, and you decide its condition: straight away, after a number of rings, or only outside your hours. Porting the number remains possible, but it is rarely necessary to begin with.
- How long does it take to get it running?
- The technical part takes a day. What takes time is writing the answers to the questions you actually get, then listening to the first calls to correct them. Allow two or three review passes over the first few weeks before the reception sounds like yours.
