Two things on this website will talk to you. One is a box in the corner of the screen that answers typed questions. The other is a phone number that, when you ring it, is answered by an AI. People reasonably assume these are the same product wearing two costumes. They aren't — and the gap between them explains most of what is genuinely hard about conversational AI, and most of what it is actually good for.
What "conversational AI" actually means
The term covers more ground than the marketing implies. Strip it back and it means any system that holds a back-and-forth with a person in ordinary language, instead of making them navigate a menu someone else designed. It needs three parts: something that takes the input, something that decides what to say, and something that delivers the answer.
That's it. The old phone tree that asked you to press 1 for accounts was conversational AI's dim ancestor — it just required you to already know which of five options your problem belonged to. What changed is that the "decide" part got good enough to handle a question phrased any way at all. What did not change is that the medium you're working in — text on a screen, or sound down a phone line — dictates almost everything about how the system has to be built.
One turn of a phone call
Here is what happens in the space between you finishing a sentence and the AI starting one.
Two details matter more than the boxes. The first is that "decide" usually involves a lookup: the system checks what the business has actually published about opening hours or pricing, rather than improvising. The second is that this loop runs for every utterance. A four-minute call is not one impressive answer — it is fifty small ones, and any stage that is slow is slow fifty times.
The widget is a different animal
Now the same picture for the box in the corner of the screen, side by side.
The widget has no microphone and no connection to the phone network. It cannot put you through to the voice agent, and on this site it doesn't pretend otherwise — when it reaches the edge of what it knows, it hands over a phone number and an email address. That limitation is worth naming, because "seamless chat-to-voice handoff" is a common slide in vendor decks and a genuinely difficult thing to build.
The two also tend to get their knowledge differently. When the body of material is small and stable — a website's own pricing, services and policies — the simplest approach is to hand the system all of it with every question. When a conversation could go anywhere and the material runs to hundreds of pages, that stops being practical, and the system instead looks up the handful of relevant passages on demand. Same source of truth, different retrieval. It matters because the failure modes differ: the first goes stale quietly, the second can retrieve the wrong passage confidently.
The clock is the whole game
If you only take one thing from this: text is patient and speech is not.
That research is the reason voice is engineered so differently. Humans hand conversational turns back and forth with gaps of about 200 milliseconds, remarkably consistently across wildly different languages. We are exquisitely tuned to that rhythm, which is why a delay of a second on a phone call registers as something is wrong — a bad line, a distracted person, a machine — long before you consciously work out what. In a chat window, the same delay is nothing at all, because typing was never that fast to begin with.
Everything expensive about voice AI follows from that one fact. It's covered in more depth in the half-second pause.
Where it actually pays for a business
The useful question isn't "is conversational AI good" but "which conversations, on which channel."
The pattern is that a widget serves people who have already found you, and the phone reaches people who haven't. That distinction decides more purchases than any feature comparison. A business drowning in out-of-hours enquiries from its own website has a widget problem. A business that needs a thousand tenants surveyed to a defensible standard has a voice problem, because most of those tenants are never going to visit a website to fill in a form.
That second case has a hard economic edge to it, and the national statistics bear it out. The Office for National Statistics recovered response levels on the Labour Force Survey from a low of 44,238 in mid-2023 to 80,078 in the first quarter of 2026 — but it did so, in its own words, "because of interventions, including increases to survey incentives and interviewer numbers." Even then, the figure sits 3,984 below the pre-pandemic quarter. Reaching people is not getting easier or cheaper. That, rather than any claim about robots being clever, is the actual argument for automating the repetitive tier of outbound conversation.
The honest trade-offs
A chat widget deflects questions; it does not create demand. If nobody is visiting your site, a better box in the corner changes nothing.
Voice AI is genuinely strong on conversations that are high in volume, tightly defined and repetitive — and genuinely weak on ambiguity, judgement and emotion. Both channels need to disclose that they're AI, and both need a visible route to a human; the transparency rules under the EU AI Act started applying on 2 August 2026, and the UK is building its own statutory guidance through a code of practice the Information Commissioner has been formally directed to produce.
And be sceptical of return-on-investment figures that come from the companies selling the software. The defensible way to evaluate any of this is narrow and boring: pick one conversation your business has hundreds of times, measure what it costs you today, and test whether the machine handles it without embarrassing you. Ask it something awkward that isn't in the material and watch whether it admits ignorance or invents an answer. That five-minute test tells you more than any demo.
The interesting thing about conversational AI in 2026 isn't that machines can talk. It's that the constraints are now well enough understood to say, fairly precisely, which conversations they should be having — and which ones they shouldn't.
Related: Speech-to-speech: why AI calls sound human · The half-second pause · Do you have to tell people it's an AI?