ChatGPT Work takes voice orders from your phone now

A typed prompt gets reread before you hit enter. A spoken one does not. On September 23, OpenAI brought Voice into ChatGPT Work on web and mobile and added plugin support to the Live voice experience across web, iOS and Android. TechCrunch reported the practical upshot: you can now start a Work task by talking to your phone. The capability is not new. The trigger surface is, and that is the part with consequences.

What actually happened

Two changes shipped together.

Live voice gained plugins and connected apps, so a spoken conversation can now reach the integrations tied to your account. Written responses still appear in the chat alongside the spoken ones.

Voice arrived in the Work tab. Plus and Pro subscribers get the full set by voice — drafting documents, summarizing email and Slack, building sites and presentations, the cloud browser, the finance surfaces. Free and Go accounts get voice in chat with whatever plugins their plan supports, not the Work tab.

Two limits are worth writing down. Live does not support video or screen sharing. And it cannot pull files out of your ChatGPT Library on its own — a supported file can still be attached to the chat by hand, depending on the account.

This builds on the GPT-Live voice model and the desktop Work integration that landed in July.

Why voice-triggered AI agents matter for your business

We have built enough of these to know where they break, and it is not the transcription.

Voice removes the review step you did not know you were relying on. When you type "summarize the Q3 supplier thread and draft a reply," you see the sentence before it runs. You catch "reply" when you meant "draft." Spoken, that sentence goes straight to an agent with connected-app access, usually while you are doing something else — which is the entire point of the feature and also the entire problem. The fix is not to avoid voice. It is to make the agent read back what it is about to do and wait, specifically on anything that sends, pays, posts or deletes.

The tier split is a standardization decision, not a feature note. Work-by-voice sits behind Plus and Pro. If you build a process around it, you are committing to a paid seat for every person who runs that process. Price that before you train anyone on it.

Connected apps are the real surface area. Plugins in Live voice means the blast radius of a misheard instruction is now every system you authorized. Go audit what those connections can actually reach — most people connected them once, for one task, and granted more than that task needed. Scope the tokens down to the job.

The pattern we keep landing on with clients: voice is excellent for capture and retrieval, and it needs a hard confirmation gate before any write. Dictating the note is fine. Sending the email is a decision.

Sources: TechCrunch, ChatGPT release notes.