A phone call has no patience for anything that is not real time. Pick up a receiver, say hello, and count how long a reply takes before the silence starts to feel wrong. It is not long. That unforgiving clock is the whole engineering problem behind real time voice AI, the technology a number of Gulf businesses are now putting on their phone lines in place of a human agent or an old-fashioned menu system.
xAIa, whose own site is headed “A PRO TECHnology Group Company,” operates out of Dubai. On its product page it says its “voice AI employees that answer inbound in about half a second and work outbound around the clock, in the caller’s own language, with a live hand-off to your people the moment a call needs a human,” and the same page states a specific number for that first answer: 400 milliseconds, set against what it labels a 1,240-millisecond industry average. Both figures are xAIa’s own and are reported here as the company’s claim about its own system, not as independently audited measurements.
Why real time voice AI lives or dies in milliseconds
xAIa’s own explainer on the subject frames the problem in human terms rather than engineering ones. On its blog, the company writes that in natural human back-and-forth, the gap between one person stopping and the next starting runs to roughly two hundred milliseconds, and that stretching that gap degrades the exchange fast: tolerable at half a second, noticeable at a full second, and liable to make a caller talk over the reply once it passes a second and a half. The company’s stated target for its own product is to begin a reply inside roughly a second, ideally well under — the point, it says, where a human brain stops hearing “delay” and starts hearing “conversation”.
The same post breaks down where that time goes: speech has to be converted to text, the system has to reason about what was said and check anything it needs to check, and the reply has to be converted back into spoken audio. Run those three stages one after another, xAIa says, and a system is well past a second before a caller hears a single word; getting under the threshold means overlapping them and keeping the processing close to the caller rather than routing it across a continent and back.
What a telecoms standards body says about the same pause
xAIa’s 400-millisecond figure is a claim about AI response latency, not a network transmission standard, and the two should not be read as the same measurement. But a comparable number has sat at the centre of telephony engineering for decades. The International Telecommunication Union’s standardisation sector, in Recommendation ITU-T G.114, approved in May 2003 and still listed as “In force” — with a later amendment added in November 2009 — states that it “provides guidance on the effect of end-to-end one-way delay (sometimes termed latency), and an upper bound one-way network delay,” and that, by its own wording, “a one-way delay of 400 ms should not be exceeded for general network planning…it is important to appreciate that highly interactive tasks (e.g., many voice calls, interactive data applications, video conferencing) can be affected by much lower delays.” The standard’s own wording, in other words, does not treat 400 ms as a comfortable target for a voice call — it flags that a call is exactly the kind of highly interactive task that can be bothered by delays well short of that ceiling. That ceiling is a rule about how long a signal should take to cross a telephone network, decided years before conversational AI existed — not a rule about how fast software should think. The coincidence is still a useful anchor: it shows that engineers who study nothing but delay on a phone line independently flagged the same few hundred milliseconds that xAIa cites as the point where a voice interaction starts to feel strained.
The behaviours that give a bot away
Speed alone does not make a call feel human, xAIa argues in a separate post on the same theme. “The tell is in the timing, not the timbre,” the company writes, pointing to interruption, acknowledgement sounds and recovery from a misheard word as the behaviours that separate a voice AI agent a caller tolerates from one they forget is software.
- Barge-in: the system stops talking the instant the caller starts, rather than finishing a scripted sentence while the caller is already speaking over it.
- Backchannel: small acknowledgement sounds during a call signal that the line is still live and the caller is being followed, the way a human listener’s “mm-hm” does.
- Recovery: when a word is misheard, the system targets the specific thing it missed instead of restarting the whole exchange from the top.
xAIa frames this as the difference between a “functional bot” and one a caller stops thinking of as a bot at all — its own description, not an independent assessment, but one that matches why a voice AI agent that answers fast and still talks over its caller can feel just as wrong as one that is simply slow.
The honest limits
xAIa does not publish how its 400-millisecond and 1,240-millisecond figures were measured, over what sample of calls, or by whom. Both numbers come from the company’s own product page, and neither has been audited by EgyptPulse or any third party cited in this article. The ITU-T standard, for its part, governs one-way network transmission delay generally — it was not written about conversational AI and does not certify any vendor’s product, including xAIa’s.
Frequently asked questions
What does real time voice AI actually need to do in under a second?
According to xAIa, it needs to transcribe the caller’s speech, work out what they meant, and generate a spoken reply, with that whole chain landing inside roughly a second — the window the company says is where a pause stops reading as “delay” and starts reading as normal conversation.
What is barge-in, and why does it matter on a call?
xAIa describes barge-in as a voice AI agent’s ability to stop talking the moment a caller starts speaking, rather than finishing its sentence regardless. The company says a system that cannot be interrupted forces a caller to sit through a full scripted line before it will listen again — a common source of frustration with older automated phone systems.
How fast does xAIa say its voice AI agents respond?
xAIa states a figure of about 400 milliseconds for its own agents, compared with what it calls a 1,240-millisecond industry average on the same page. Both numbers are the company’s own claim about its own product and are not independently verified here.
Who is behind xAIa?
xAIa says on its own site that it is the AI consultancy and implementation arm of PRO TECHnology Group, under a page headed “A PRO TECHnology Group Company.” PRO TECHnology describes itself as a Dubai-founded technology group operating across enterprise IT, distribution, digital printing, indoor air quality and enterprise AI since 1998.
Product details and capabilities in this article are as published by xAIa at the time of writing.
