Voice conversations can now reach the apps you have already connected, and Voice runs inside ChatGPT Work. An unfinished task continues in text when the call ends.
The gap this closes is the one between meetings. A rep who ends a call at 10:58 with a fifteen-minute drive to the next one has had two options: type the follow-up on a phone keyboard, or let it slide to the end of the day, by which point four calls have blurred into one. Voice that can reach the apps you have already connected turns that dead time into the follow-up itself. The thing worth having is not faster typing — it is recency. A summary dictated eleven minutes after a call carries the detail that the same rep cannot recover at six in the evening, and detail is what makes a follow-up read as though someone was listening.
A few terms in the note are doing quiet work. A plugin is OpenAI's current container for reusable instructions plus a connection to an outside service — your mail, your calendar, a CRM connector — installed once and called by name inside a conversation. Live is the real-time spoken mode, the one where you talk and it answers while you are still talking. Work is the workspace surface, separate from Chat, with its own models and its own admin settings.
"Voice in Work requires both Voice and Work access" is two permissions, not one. Your administrator can have granted you Work and not Voice, or the reverse, and the symptom in both cases is a feature that simply is not there. "Existing app connections, permissions, and usage limits apply" is the sentence that stops this being bigger than it is: speaking to the assistant does not connect anything new, and it does not raise a cap you were already hitting.
The short version: you can now talk to ChatGPT and have it use the apps you already connected, and in Work you can ask out loud for a document, a deck or a spreadsheet and pick the job back up in text once the call ends.
OpenAI added plugin support to ChatGPT Voice and extended Voice into ChatGPT Work in a release note dated September 23, 2026. The entry reads: "Live now supports plugins on web, iOS, and Android. You can use the plugins and connected apps available to your account during a Voice conversation and follow written responses in the chat."
Announced: September 23, 2026, in the ChatGPT release notes.
Availability: Voice with plugins on web, iOS and Android. "Voice is also available in Work on web and mobile." Free and Go users "can use Voice in Chat with the plugins their plan supports." "Voice in Work requires both Voice and Work access."
Price: Not stated.
Regions: Not stated.
What the Work surface adds, in OpenAI's words: "Ask it to create documents, presentations, and spreadsheets, use connected apps, or work in a browser. When you end a Voice call in Work, an unfinished task can continue in text." One more clause carries weight: "Existing app connections, permissions, and usage limits apply."
What it doesn't do: The entry names no supported plugins, sets no rollout date, gives no price and no regions, and states no limit on how long a Voice call can run. It also does not say whether the assistant can act inside a connected app while you are speaking or only read from it, and that distinction is the difference between hands-free work and dictation with lookup. "Existing app connections, permissions, and usage limits apply" means nothing new is granted: a plugin you have not connected will not appear because you asked for it out loud.
SDR — Your version of this is the gap between dials, not the gap between meetings. After each connect or no-answer, say the outcome and the next step out loud instead of batching notes at five o'clock. Before you build the habit, test one thing: whether a Voice session can actually reach your mail or calendar plugin, because the note is explicit that only your existing connections apply and nothing new is granted by speaking.
AE — Work is the part that matters to you, and the continuation clause is the feature: an unfinished task carries on in text after the call ends. Use it to start a recap or a mutual action plan in the car and finish it at a keyboard. Do not try to produce a buyer-ready artifact hands-free; a first draft with the right facts in it is worth more than a polished one written three hours later from memory.
Manager — This changes when coaching evidence exists, not how much of it there is. Notes dictated minutes after a call are the only ones that still contain the offhand remark the buyer made on the way out. Put one rep on it for a week and read their CRM notes against the team's. If the difference is visible, you have a habit worth spreading; if it is not, you have saved everyone a rollout and learned something about your notes problem.
RevOps — Two things to settle before anyone announces this internally. First, permissions: Voice in Work needs both Voice and Work access, so a mixed-seat team will contain reps who can do this and reps who cannot, and the ones who cannot will assume the feature is broken rather than ungranted. Publish who has what. Second, the note does not say whether a plugin can be actioned mid-call or only read from. Until that is answered, do not write a policy that assumes a spoken instruction can write into a connected system.
Dictating well is a prompt problem, not a microphone problem — a spoken instruction still has to carry the structure a typed one does, and rambling into a Voice session produces exactly what rambling into a text box produces. Start from prompts built for AI tool operations and modern sales workflows.
Two minutes, once a week. What changed in AI, and what to run because of it.