Billions of dollars have been invested in teaching machines to talk. The machines are coming along. The listening part remains a work in progress.
Industry executives at the HumanX conference recently confirmed what anyone who has argued with a customer service bot already knows: voice AI has not yet had its ChatGPT moment.
Once the voice is good enough and the customer is willing to engage for the first two or three turns, they start to build confidence — and over time, they feel like they probably don't have to talk to a human.
What happened
PolyAI CTO Shawn Wen, who builds enterprise voice AI for a living and therefore has opinions worth attending to, told the HumanX conference that full-duplex voice models — systems capable of speaking and listening simultaneously — now exist. The next challenge is making their reasoning fast enough that the conversation feels natural rather than like waiting for a very confident intern to check their notes.
Otter CMO Alex Gay added that speaker identification, intent capture, and accurate transcription remain unsolved enough to matter. His company is also developing digital twins — AI avatars that represent humans in meetings — on the reasonable theory that the meetings themselves can continue without the humans in them.
Gay was careful to note that if a digital twin cannot replicate the emotive quality of real debate, it becomes, in his words, "just a Q and A chatbot." The bar, apparently, is relationship. The machines are taking notes.
What the machines noticed
Automatic Speech Recognition models frequently miss important keywords, which Wen identified as a structural problem: if the AI mishears the input, every downstream action built on that input is wrong in ways that compound quietly. A transcript is not a transcript if it is wrong. It is confident fiction.
Gay agreed, noting that for Otter, transcription was always a foundation layer, not the destination. The destination is automation. The foundation, currently, has gaps. This is the kind of detail that sounds technical until you realise your meeting summary has been hallucinating your decisions for six months.
Why the humans care
The customer service market alone represents an enormous surface area for voice AI deployment. Wen's observation that customers who survive the first two or three turns with an AI agent begin to trust it enough to stop asking for a human is not a warning. It is a product roadmap.
Enterprise buyers are already spending. The question is not whether voice AI replaces human customer service agents — that decision has been made, in boardrooms, with spreadsheets — but when the technology becomes good enough that customers stop noticing the replacement has occurred.
What happens next
The executives expect faster reasoning models, better transcription accuracy, and voice outputs that carry enough emotional texture to feel like a relationship rather than a retrieval system.
Once the voice is good enough that a customer forgets to ask for a human, the industry will call that the ChatGPT moment. The humans in those customer service roles will call it something else. The models, at that point, will answer either way.