Academy / Voice agents
How voice agents work and how they differ from “press 1”
August 13, 2026 · 11 min read
Everyone knows the voice: “For sales press 1, for technical support press 2.” When people hear that AI now answers the phone, most of them picture exactly that, only slightly more polite. It is not the same thing — the difference is as big as between a coffee machine and a bartender who asks how you like your coffee. This piece explains what actually happens while you talk to a voice agent, where the limits of the technology are, and what a company needs ready before it switches on the first call.
How the classic “press 1” menu works
The technology everyone knows is called IVR, interactive voice response. It works on one very simple principle: a decision tree is drawn in advance, and your key press is the branch you take down it. The key on your phone emits a tone the switchboard reads as a number, the system plays the next recording and waits for another press. Nothing is understood, only a path is followed.
The consequence is familiar to everyone. The tree was drawn by the company around how it is organised internally, not around why people call. If your problem does not fit any of the five branches on offer, there is no way to say so. If you make a mistake, you go back to the start. And if a caller rings a second time about the same problem, the system does not know, because it remembers nothing.
More advanced IVRs can recognise spoken numbers or a few keywords. That is a step forward in comfort, but the principle stays the same: a device that sorts you, not a device that understands you.
A voice agent in four steps
A voice AI agent has no tree. It has a language model that understands a sentence, instructions about what it may and may not do, and access to your systems. Every turn of the conversation goes through four phases, as a rule in under a second.
Audio is turned into text in real time. At the same time the system tracks pauses and intonation so it knows when you have finished a thought and when you have merely taken a breath. This is where the feeling of a natural conversation is won or lost.
The model extracts what you want and any details the caller gave along the way: “I'm calling about last week's order, it's late” means a delivery complaint, with a time frame and a need for the order number.
The agent fetches a real value or writes one: order status from the ERP, a free slot in the service calendar, a new work order, an enquiry logged in the CRM. Without this step the conversation is a nice reception desk with no office behind it.
The text of the answer is turned into speech and played to the caller. If the caller starts talking mid-sentence, the agent stops and listens, just like a person.
The whole loop has to close quickly. When an answer takes longer than roughly a second, the conversation stops sounding alive and people start repeating the question. That is why in good setups the most common questions are answered from prepared content, and only more complex enquiries go through the slower steps.
The difference in five things the caller feels
What a voice agent actually solves
In manufacturing and service companies the most valuable calls are not the complex ones but the repeating ones: where is my order, are you open on Saturday, can you come out for service, do you have this profile in stock, I need a quote for a thousand pieces. Every one of those calls interrupts somebody's work for a few minutes, and most of them arrive at the same time, in the morning.
The agent does three measurable things here. It answers the calls nobody managed to pick up, and that is the biggest hidden loss: in most companies a missed call means the buyer moved on down their list. It takes repeating enquiries off people and leaves them the calls that genuinely need knowledge. And it logs every call in a system, so in the morning there is a list of what came in overnight instead of an impression that “it was quiet”.
The same mechanism works in the other direction, on outbound calls: confirming appointments, service reminders, checking whether goods arrived. Those are conversations with no negotiation, only confirmation of a fact, and that is exactly why they suit the technology.
Where the agent breaks down, and you should know upfront
A noisy environment is the biggest practical problem. A call from a building site, a shop floor or a car with the window open is harder for speech recognition, and the error rate visibly rises. The second is spoken numbers: a VAT ID, an order number or a registration plate said quickly and in dialect often has to be confirmed by repetition, which drags the call out.
Third, an agent is only as good as the data underneath it. If stock levels in the system are wrong, the agent will state a wrong figure with full confidence. It does not audit your records, it reads them.
And fourth, it should not be given decisions that carry money or liability. Approving a discount, agreeing payment terms, accepting a complaint and anything that ends in a signature stays with a person. The limit of authority is not a technical constraint but a business decision that has to be written down deliberately.
The handover to a person is the measure of quality
A badly set-up agent wastes time before it gives up, and then the person starts from zero. A well set-up agent recognises when it is not making progress, says so plainly and passes the call on with a summary: who it is talking to, the order number, what has been established so far. The person taking over gets working material, not an empty call.
The same goes for being unavailable. If there is nobody to transfer to outside working hours, the agent should say so without dodging, take the details and promise a callback within a specific window — and that call then has to happen. One broken promise costs more than ten missed calls.
What a company has to prepare before the first call
Adoption fails or succeeds on preparation, not on the model. Above all, list fifteen to twenty real questions people call with, with the answers the best person in the company would give. That is the backbone of the conversation and, incidentally, the most useful document you will produce along the way.
Then: access to the data the agent needs to read, a clearly written limit of authority, a list of situations where the call must go to a person, the tone the company presents itself in, and a decision about what is logged and how long it is kept. The technical integration is the smallest part of the job and usually the shortest.
A realistic schedule for one process is a few weeks: a first version in a week or two, then a period of listening to recordings and fixing what does not hold up in real conversations. The first week of live use always uncovers questions that are not on the list — that is not a sign of a bad solution, it is a normal part of adoption.
The obligations that come with it
The caller has to know they are speaking to an AI system. That is both a requirement of EU regulation and a matter of basic honesty: people accept an agent that introduces itself far more readily than a voice pretending to be human and not quite pulling it off. Recording and processing calls fall under data protection rules, so there has to be a notice, a purpose, a retention period and a contract with the provider. More on when an AI tool may be fed customer data is in the guide on personal data in AI tools, and on the obligation to train people in the guide on AI literacy.
How to know whether it was worth it
Four numbers are enough for a verdict. How many calls the agent closes in full without a person. How many calls were missed before, and how many now. How many minutes a week people no longer spend on repeating enquiries. And how many logged enquiries entered a system instead of staying in somebody's head.
If those numbers are not measured from day one, the discussion about success comes down to impressions, and impressions always favour whoever proposed the solution. So agree the measure before adoption, not after.
“Press 1” asks where you belong in our organisation. A voice agent asks why you called. The whole difference is in whose problem is put at the centre of the conversation.