Voice AI Hasn't Hit Its ChatGPT Moment Yet, Enterprise Builders Say
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, CTO of enterprise voice AI platform PolyAI, made the remarks alongside Otter CMO Alex Gay during a panel on voice interfaces. The discussion was reported on Oct. 11, 2026.
The gating problem, in Wen's account, is reasoning speed. The next challenge is making reasoning very fast so models fetch answers quickly and conversation feels natural. For telephony and customer-support workloads, that means tight coupling of retrieval, reasoning and synthesis under strict turn-taking constraints. Latency is not cosmetic. It determines whether barge-in, backchannels and interruptions work.
Wen also pointed to the front end of the stack. He said automatic speech recognition (ASR) models often miss important keywords, creating an issue in capturing whole context. In production pipelines, that error propagates. A missed entity or product name in the transcript becomes a missed slot for the dialogue manager and the downstream LLM.
Accuracy compounds, trust erodes
Gay described the same failure mode from the meeting side. Gay is CMO of meeting notetaker Otter, which is working on digital twins to represent people in meetings.
Otter's dependency on the base transcript is direct. Gay said if original transcription lacks needed accuracy, all follow-up actions become flawed and users lose trust in the platform. Summaries, action items, retrieval and twin behavior all inherit WER and diarization errors. Accuracy compounds.
That explains Otter's current focus on capture and consent mechanics. The company plans to try notifying all participants in chat that a meeting is being recorded even when its bot is not present. The proposal is framed as a fallback disclosure path for meetings where a joinable bot cannot attend or is not admitted.
Disclosure was a shared theme. Wen said it is important to establish that people are talking to AI in enterprise calls. In regulated support flows, explicit identification affects recording consent, data handling and escalation expectations, as well as basic conversational grounding.
Why the interface bet persists
The caution contrasts with longer-running predictions about voice as a primary modality. 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
Platform work has moved in parallel. OpenAI piloted Custom Voices, a tool that lets developers clone a voice with 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 a split between demo readiness and deployment readiness. Full-duplex transport and interruptible output solve floor control. They do not solve answer quality under time pressure. A voice agent must hear correctly, retrieve correctly and speak promptly, in that order, every turn.
Looking at what this means for builders, the HumanX comments point to two distinct latency budgets. One is acoustic and streaming: endpointing, partial ASR hypotheses, incremental TTS. The other is cognitive: tool calls, knowledge lookup and reasoning traces compressed into a few hundred milliseconds without sounding evasive. Text chat tolerates a spinning indicator. Voice does not.
In my view, the "ChatGPT moment" analogy is useful but imprecise. Chat-based assistants broke through when a single model capability made new behavior obvious to non-experts. Voice has no single capability to unlock. It requires joint optimization across ASR, LLM reasoning and TTS, plus product work on identity, consent and auditability that text largely deferred. Progress will read as reliability gains rather than a launch event.
That reliability framing clarifies the Otter and PolyAI priorities. Keyword retention, transcript fidelity and explicit AI identification are not edge concerns. They are preconditions for delegation. Users will let a twin attend or an agent resolve a billing call only if the record is correct and the counterpart is known.
Worth flagging for enterprise teams, the upside of solving those preconditions is large. Hands-free operation, shorter handling times and persistent meeting memory change how support centers and knowledge workers allocate attention. The technology arc still points toward voice as an ordinary control plane, not a novelty. It simply has not cleared the bar where ordinary users assume it will work.
