OpenAI has added voice conversations to ChatGPT Work on iOS, Android and the web, letting eligible users ask aloud for documents, email drafts or summaries of connected Slack messages. The September 23 update puts spoken requests in reach of Work, the part of ChatGPT designed to carry out longer tasks across files, apps and a browser.
According to OpenAI’s release notes, unfinished Work can continue in text after the voice call ends. That makes a phone conversation a way to hand off a task and return to its results later. There is still a reason to keep the screen within reach: whenever an action requires approval, users must tap the on-screen controls. Saying yes aloud does not count.
The update also lets Live, ChatGPT’s conversational voice interface, use plugins available to the account. Free and Go users get that capability in ordinary Chat; mobile Work requires an eligible higher-tier plan. OpenAI’s current Work page and Voice documentation establish this split:
| Plan | Voice access on web and mobile | Conditions |
|---|---|---|
| Free and Go | Voice in Chat with supported plugins; no mobile Work. | Only plugins and capabilities included with the plan. |
| Plus and Pro | Voice in Chat and Work, including supported app, browser and document tasks. | Account availability, connected-app permissions and usage limits apply. |
| Business, Enterprise and Edu | Work is listed for web and mobile; Voice in Work requires access to both features. | Administrators can restrict Work, apps and actions. Enterprise and Edu can also restrict access by role. |
| Pro Lite, Healthcare, Clinicians and K-12 Teachers | Public documentation does not provide one reconciled mobile Work eligibility list. | An omission does not establish that an account lacks access. |
Availability also depends on region, app version and the particular plugin. An eligible subscriber starts by choosing Work, selecting its Voice control and allowing microphone access if prompted. OpenAI says written responses and task results appear in the chat. These are documented capabilities, not an independent test of how reliably each app completes a spoken request.
The practical boundary lies between preparing work and carrying out an action with a connected account. A request to draft an email does not establish permission to send it. Nor does the voice update supply a universal list of actions that always require confirmation: that depends on the connection’s access, enabled actions and approval settings.
Connecting an account authorizes access to specified information and capabilities in that service; workspace controls can further limit which actions are available. ChatGPT’s approval setting determines when it asks before using that access. Any required approval during mobile or web Voice must happen on screen. Changing the approval setting does not revoke the connection’s underlying access.
OpenAI’s permission options range from “Always ask,” which prompts before reads or changes, to “Allow read actions,” which permits reading without a prompt but asks before changes. “Allow low-risk actions” can run those actions automatically, while higher-risk actions may require approval or be denied. Eligible connections may offer “Allow all actions,” allowing supported actions without further prompts, subject to remaining controls and safety protections. Every option is not available for every app or workspace.
What users authorize can extend well beyond retrieving a single message. Slack’s documented permissions cover search, channel and direct-message content, files and message writing, with additional permissions for enabled features such as reminders or uploads. Outlook’s documentation lists permissions to read and modify mail, send messages, and access calendars and contacts.
Those permissions do not expand what the connected user can access in the source service. The exact authorization request depends on enabled services, actions and prior administrator approval. The documentation establishes the possible scope; it does not establish one identical consent screen for everyone.
Users can disconnect an account to stop future access through that connection. Other connections may remain active, and disconnection does not automatically delete existing chats, saved files or memories. Provider-side unlinking or administrator action may also be needed.
We see the useful change as being able to talk a task into motion, then inspect the work when it is ready. The decision worth making before that first request is how much account access the task needs, and which actions should still wait for a tap.