Three words, three different products
Ask three vendors for a quote and you will get a callbot, a voicebot and an IVR described in almost the same language. Natural conversation, answers around the clock, integrated with your CRM. Underneath the language sit three different machines with three different failure modes, and the gap between them is wide enough that buying the wrong one shows up in the budget and on the line.
We ran live call centres for about ten years, up to 1,500 operators at peak, across France, Mexico and beyond, and in our own operations we spent those years on the buying side of this equipment. The projects that went badly rarely went badly because a product was weak. They went badly because somebody bought a routing menu for work that needed an open conversation, or paid for an open conversation on traffic a menu could have finished for a fraction of the money.
The three words are not interchangeable. Here is what each one is, and which job each one belongs on.
An IVR routes the call. That is the whole job.
IVR stands for interactive voice response, and in French the same box is a serveur vocal interactif. It plays a prompt, collects a keypress or a single spoken keyword, and sends the call somewhere: a queue, a department, an agent group. The better ones read one field out of a database, say it aloud and end the call. Balance, order status, opening hours. That is the complete inventory.
An IVR does not know what the caller wants. It holds a tree that somebody drew, and the caller's task is to locate themselves inside it. When the caller's problem matches a branch, an IVR is the fastest and cheapest thing on this list by a wide margin. When it does not match, the tree has nothing left to offer.
What an IVR is for
An IVR does not answer the question. It decides who will.
The economics explain why IVRs are still standing on so many business lines. Gartner put the cost of a self-service contact at $1.84 in 2025 against $13.50 for a live assisted one. A caller who wants a tracking number and nothing else is worth serving with a machine, and that was true long before anyone said AI in a sales meeting.
In our own experience, complaints filed against IVRs are usually complaints about scope. Somebody installed a routing device in front of traffic that needed an answer, and the device did exactly what it was built to do.
A callbot or a voicebot is a script with a microphone
Callbot and voicebot describe the same product. French buyers tend to say callbot, Spanish-speaking buyers tend to say voicebot, and vendors follow the local habit rather than any technical distinction. What both words name is a voice interface over a fixed set of intents.
The build looks like this. Someone lists the reasons people call: book, cancel, reschedule, check an order, reset a password. Each reason gets a dialogue branch with the questions it has to ask and the fields it has to fill. Speech recognition goes on the front, a synthetic voice on the back, and the whole thing connects to the booking system or the CRM. Inside its own list a well-built voicebot is genuinely good. It takes free speech instead of a keypress, it confirms what it heard, it writes the record, and the caller never touches a keypad.
The limit is the list. A voicebot cannot handle a request that is not on it, and there are only three things it can do when one arrives: ask the caller to rephrase, force the request into the nearest branch, or transfer. It behaves like a capable new hire who was handed four pages of instructions and no authority to depart from them. This is not weak engineering. It is what a fixed intent set is.
Which is why a voicebot demo tells you very little. Vendors demo the happy path, because the happy path is where the product is strong. The useful question in that meeting is what the last ten real calls did when the caller said something off-list.
An AI voice agent is built for the call nobody scripted
An AI voice agent starts from the opposite end. In place of an intent list there is a brief: what the business sells, what the agent may promise, what it must never say, which systems it may read and write, and the conditions under which it hands the call to a person. The conversation itself is open. The caller can phrase the request sideways, change their mind halfway through, stack two questions into one sentence, or arrive with a situation nobody imagined when the agent was written.
What that buys is exception handling. A neuro-agent can take a request it has never seen, work out what is actually being asked, pull the record, and either close the matter or hand a colleague a call that is already halfway done. It also buys a harder operational job. An open conversation has to be constrained, tested against the awkward cases and watched in production, because a system that can say anything can say something wrong. Skip that work and you have bought a liability with a pleasant voice.
The second difference shows up over time. Adding a service to a voicebot means a new branch, new phrasing, new test coverage and a release. Adding it to a neuro-agent mostly means telling the agent about the service and checking that it holds. In a business that changes its offer every quarter, that gap compounds.
- IVR: a keypress or one keyword goes in, a route or a single database field comes out. No model of intent at all.
- Callbot or voicebot: free speech goes in, a fixed intent set governs what comes out. It handles what it was taught and transfers the rest.
- AI voice agent: an open conversation, live access to your systems, and an answer for the request that was never on anyone's list.
And where the word neuro-agent fits
A fourth word circulates in this market and it does not name a fourth product: neuro-agent. It is the term we use for our own agents across this site, so it is worth saying plainly what it means. A neuro-agent is an AI voice agent, the category described just above, the one built for the call nobody scripted. The word points at how the thing is built, on neural speech and language models rather than on a decision tree, not at a separate class of product sitting alongside the other three.
The distinction matters when you are reading quotes, because none of this vocabulary is standardised. Nothing stops a vendor calling a scripted callbot a neuro-agent, or calling a genuinely unscripted agent a voicebot, and the words on the proposal will not tell you which one you are being sold. Only the behaviour will: whether it can be interrupted, whether it can handle a question outside its script, and whether it passes the call to a person when asked. Buy the answers, not the label.
Which one belongs on which line
Match the tool to the traffic. Before any vendor call, write down the reasons people phone you and the volume behind each one. That list decides the purchase, and it usually decides it in under an hour.
- An IVR is right when the traffic is one or two questions with a database answer and the caller already knows which question they have. Balance, tracking, opening hours, a split between two departments. Here the cheap option is the correct option, and replacing it with something conversational buys nothing.
- A callbot or voicebot is right when the requests repeat and you can write the list down honestly. Appointment booking, confirmations and reminders, a fixed order flow, a delivery query. If a supervisor can name the ten reasons people call and the tenth is already rare, a voicebot covers the line.
- An AI voice agent is right when the eleventh reason is where the money is. A first call from a stranger, a sales conversation, a complaint, a caller who cannot name the thing they need. Anything where the value of the call sits in the exception rather than the pattern.
- None of the three is right for the call your best operator exists to take: the negotiation, the furious high-value customer, the decision that needs authority. Get that call to a person quickly, and use whichever machine you have to make sure it lands on the right person with the context attached.
The fourth pattern is the one we keep coming back to, and it is not a compromise. An IVR at the front for the two questions that never need a human. An AI voice agent behind it for everything the menu cannot finish. A warm transfer into the live team at the point where the agent reaches its limit, with the caller's details already on the screen. Nobody has to choose a single product for a whole phone number.
A voice channel wins by resolving, not by sounding human
The trust numbers are the reason this decision matters, and they do not say what the vendors of the very humanlike voice would like them to say.
Five9's 2024 consumer survey found 56% of consumers often frustrated by service chatbots and 75% still preferring to talk to a real human. SurveyMonkey put the share who prefer AI over a human at 8% in 2025. In France, the 2025 Observatoire des Services Clients by Ipsos BVA scored chatbot satisfaction at 50%, the lowest-rated channel it measured. Read together, they are a verdict on machines that finish nothing.
Over the same years the phone got more important, not less. ContactBabel tracked phone preference for urgent, complex issues rising from 28% to 47% between 2018 and 2024, and McKinsey found in 2024 that 71% of Gen Z pick up the phone when self-service fails. People come to voice at the point where the cheap channels have run out. They arrive already irritated, on the second attempt at something.
A caller forgives a machine for sounding like a machine. They do not forgive it for wasting the call.
What to ask the vendor
Ask for the share of calls that ended without a human and what happened to the rest, not for another voice demo.
Gartner forecasts that AI agents will autonomously resolve 80% of common service issues by 2029, cutting costs by around 30%. That forecast rests on resolution. It says nothing about how convincing the voice was. In a pilot, the number that matters is the share of calls that closed without a person, and, for the ones that did not, whether the caller reached a colleague who already had the details in front of them.
So buy the cheapest of the three that can actually finish the job in front of it. Measure what it fails to finish. Then buy the next one up, for that traffic only. That order keeps the budget honest and keeps the caller out of the middle of your architecture.