Voice AI has not had its ChatGPT moment yet. The people building it blame reasoning speed, not speech.

Two executives, one verdict: not yet

Full-duplex voice models, the kind that can speak and listen to a person at the same time instead of waiting for a pause, are real and already shipping. That has not been enough to give voice AI its own version of the moment ChatGPT created for text. Shawn Wen, co-founder and chief technology officer of the enterprise voice AI company PolyAI, said as much on stage at HumanX Amsterdam, the AI industry conference held 22 to 24 September 2026, in remarks reported by TechCrunch on 11 October. "We have reached the milestone of developing full-duplex models," Wen said, but he does not count that as the breakthrough. The constraint he actually points to is what happens after a model hears you: how fast it can think.

What full-duplex actually fixed

Before full-duplex models, most voice assistants worked in a strict back and forth: you spoke, the system waited for silence, then it processed and replied. That turn-taking lag is a big reason early voice bots felt robotic even when the underlying language model was capable. Full-duplex architectures remove the wait by letting a model listen continuously while it is also speaking, closer to how two people actually talk over each other, interrupt, and recover. PolyAI builds enterprise voice agents for call centers and customer service lines, with named clients including NatWest, UniCredit, Foot Locker, PG&E, and Metro Bank, so Wen is describing a problem his own company's product has to solve commercially, not a lab demo. PolyAI's own site describes its stack as Dialog Studio for building an agent, Wren for designing and testing one, and two in-house dialog models it calls Dialog-RSN-1 and Dialog-Raven-3.5, all aimed at the same goal: resolving a caller's problem, in the company's words, "without escalation" to a human. That is a vendor's framing of its own product, not an independent benchmark of it, but it shows what the industry is actually selling against: not a chatbot demo, but a replacement for a phone tree a caller might otherwise hang up on.

The bottleneck nobody solved: reasoning speed

Removing the turn-taking lag exposed a different one. Wen told the Amsterdam audience that "the next challenge is to make reasoning very fast, so that the models can fetch answers quickly," because a voice agent that can listen and speak at once still breaks the illusion the moment it has to pause and think. In a text chat, a few seconds of latency barely registers. On a live phone call, the same delay reads as the system stalling, and callers notice immediately. Wen also argued that a customer-service voice agent has to sound natural enough that callers actually trust it, and that companies deploying one should make clear upfront that the caller is talking to AI rather than a person, not only for compliance but because the interaction works better when expectations are set honestly from the start.

Otter's bet: knowledge, not just notes

The same "not yet" verdict showed up from a different angle at Otter.ai, the AI meeting-notes company that named Alex Gay as its first chief marketing officer on 11 August 2026. In announcing the hire, Otter said Gay would lead its push beyond transcription toward what it calls an enterprise conversational knowledge platform, built on top of more than 40 million professionals the company says already use it. "Voice is one of the richest, most underutilized sources of knowledge inside every organization," Gay said in the announcement, and Otter co-founder and chief executive Sam Liang added that Gay "brings a rare combination of brand craft and product marketing rigour," the kind of hire a company makes when it is repositioning, not just filling a marketing seat. That framing treats a meeting recording as raw material for something closer to an AI agent that can act on what was said, not just a transcript to search later. Otter is reportedly developing "digital twins" that could represent people in meetings they cannot attend, an idea that leans on the same reasoning speed and accuracy Wen was describing from the call-center side.

The accuracy problem underneath it all

Per TechCrunch's reporting on both executives, speaker identification and intent capture matter as much as raw transcription, because automatic speech recognition still misses keywords often enough to break the context a system needs downstream. Gay was blunt about where that leaves the category: "transcription was never the end point," he said, meaning accurate text is the floor, not the product. Both executives raised a quieter issue too: an error in a transcript does not stay contained, it propagates into every automated action a system takes based on that text, from a summary to a scheduled follow-up. Otter has said it plans to notify meeting participants whenever a session is being recorded, even when no visible bot joins the call, a disclosure step aimed at the same trust problem Wen raised about customer-service calls.

Our take

Two executives with competing products and no obvious reason to flatter each other's industry are describing the same gap: the speaking part of voice AI got solved before the thinking part did. That is a checkable claim, since full-duplex models genuinely exist now and the companies selling them are not shy about saying so, and it is a useful corrective to marketing around "natural" voice assistants that still make callers wait. Investors have put billions of dollars into voice AI startups betting the ChatGPT moment is close. Based on what the people actually building the products say, close is not here yet, and the bottleneck worth watching is how fast the reasoning behind the voice can go, not how human the voice itself sounds.

That gap is also where to be skeptical of a sales pitch. A vendor demo recorded in a quiet room, with a scripted question and a fast scripted answer, hides exactly the latency and mishearing problems both executives are describing. Before trusting a voice AI product with a real customer call or a real meeting, a buyer should ask how it performs on an unscripted interruption, an accent the model was not tuned on, or a question that needs more than a cached answer, because that is where the reasoning delay and the transcription errors actually show up.