Why Phone Bots Still Sound Clunky

Enterprise voice AI has not reached its "ChatGPT moment," PolyAI CTO Shawn Wen said on stage at the HumanX conference last month, despite progress on full-duplex models. TechCrunch
Wen leads technology at PolyAI, which builds voice systems for businesses. He spoke alongside Alex Gay, marketing chief at Otter, during a panel on voice interfaces. The discussion was reported on Oct. 11, 2026.
Full-duplex means the AI can listen and speak at once, like a normal phone call. Wen said the main blocker is reasoning speed. Models need to think and fetch answers much faster so talk feels natural.
For phone support, the system must do three things in tight order: find facts, reason about them, and turn the answer into speech. Delay is not a small polish issue. It decides whether a caller can interrupt, murmur agreement, or cut in.
Wen also pointed to the first step. Speech recognition, which turns voice into text, often misses key words and loses full context. That early miss spreads. If a name or product is missed in the text, the dialogue manager that tracks the conversation and the language model that follows miss it too.
Accuracy compounds, trust erodes
Gay described the same problem from the meetings side. Gay is marketing chief at Otter, a meeting notetaker working on digital twins that can sit in for people in meetings.
Otter depends directly on the written transcript. Gay said if that transcript is not accurate enough, later steps go wrong and users lose trust in the platform. Summaries, to-do lists, search and twin behavior all copy two kinds of mistakes: wrong words, and wrong labels for who spoke when.
That is why Otter is focused on recording and permission. It plans to test notifying everyone in chat that a meeting is being recorded even when its bot is not there. This is a backup warning for meetings a bot cannot join or is not allowed into.
Disclosure came up for both speakers. Wen said people should know when they are talking to AI on business calls. Clear labeling affects permission to record, how data is handled, when to bring in a human, and simple turn-taking.
Why the interface bet persists
The caution sits next to bold predictions that voice will become a main way to use AI. ElevenLabs co-founder and CEO Mati Staniszewski says voice is becoming the next major interface for AI. TechCrunch He has separately argued that AI audio models will be commoditized over time. TechCrunch
Other building blocks have moved forward. OpenAI tested Custom Voices, a tool that lets developers copy a voice from a 15-second sample. TechCrunch It later started rolling out advanced voice mode to some ChatGPT Plus users, a mode that allows users to speak to ChatGPT and receive real-time responses without delay and to interrupt ChatGPT during responses. Reuters
The broader context here is the gap between a good demo and a system ready for daily use. Being able to interrupt solves who speaks when. It does not solve giving a good answer fast. Like a game of telephone, a voice agent must hear right, find facts right, and speak quickly, in that order, every turn.
Looking at what this means for builders, there are two time budgets. One is for sound: telling when you finished, guessing words early, and speaking in small pieces. The other is for thinking: checking tools and records in a fraction of a second without stalling. Text chat can show a loading spinner. Voice cannot.
In my view, the "ChatGPT moment" comparison is useful but not exact. Chat bots broke through when one clear ability was obvious to everyone. Voice needs many parts to improve together: hearing, reasoning, and speaking, plus clear rules for identity, permission, and records. Progress will look like steady reliability, not one big launch.
Worth flagging for business teams, the payoff is still large if those basics are fixed. Hands-free help, shorter calls, and a reliable memory of meetings would change how support teams and office workers spend attention. Voice is still headed toward becoming normal, not a novelty. It just does not work well enough yet for people to take it for granted.


