OpenAI revises its ChatGPT sales meeting prep guide to clarify who connects what

A correction, not a new feature

OpenAI has revised its guide to building a sales meeting prep agent in ChatGPT, as note reported. The update appeared on November 6 (US time) in the OpenAI Developers Cookbook, a collection of how-to articles. The guide was first published on April 22, 2026.

The revision adds no new capabilities. It corrects how the guide explains connections, sharing and scheduled runs so that the text matches how workspace agents behave now. A workspace agent, in OpenAI's usage, is an AI that uses connected apps to repeat a fixed task and can be shared across a team.

That makes it a useful read for anyone who has been unsure what a shared agent can see and do. The details about permissions are the substance of the change.

What the example agent does

The guide's example looks at Google Calendar for customer meetings on the next business day. It skips internal meetings and any meeting with no external attendees. On a Friday, that means it prepares for Monday.

For each meeting it pulls recent customer notes and feedback from SharePoint, Microsoft's document storage service. It also searches the web for company news from the past 30 days. It then writes a preparation document of 2-3 pages, saves it to a team folder, and emails the user a summary with a link.

There are prerequisites. You need a ChatGPT Business, Enterprise or Edu workspace where workspace agents are available. Administrators must also allow reading calendars, searching and creating documents, and sending through Gmail. If writing is restricted, the guide explains how to have the document returned inside ChatGPT instead.

Whose account does what

The key design choice is that two kinds of connection are mixed. Google Calendar and Gmail use each user's own connection. Even if one agent is shared with colleagues, every person has to connect their own calendar and email.

SharePoint works the other way. It runs through a service account for the sales team, meaning one shared account used to reach customer documents and the save location. Anyone who can run the agent can use whatever SharePoint actions that account allows.

So before sharing, the guide says to compare two things: what the shared account can read, and which actions the agent is allowed to take with it. OpenAI advises turning on only the actions needed to search, read and create documents. Sharing, deletion and unrelated writing should stay off. It also tells you to check that the people receiving a file can actually open it.

Slack changes the picture. Deploying an agent there requires shared authentication for every app connection. Because this example relies on personal calendars and email, the guide keeps it running inside ChatGPT.

Permissions and approvals are separate

The guide treats operation permissions and approvals as two different settings. By default, workspace agents ask for confirmation before any write operation. OpenAI asks users to test how file creation and email sending behave before relying on recurring deliveries.

The suggested test happens in Preview, run from the creation screen with your own account. You check three things: the file exists, the email arrives, and the link opens for the recipient. The guide also suggests comparing the output with prep documents people made for past meetings, or loading existing team templates so the sections match what your team expects.

Recurring runs use the published version of the agent, and they act as the person who created the schedule. That has a practical consequence. For colleagues to receive their own prep documents each week, each of them must connect their own Google Calendar and Gmail and add their own schedule. Sharing the agent and setting up each person's recurring run are separate steps.

What the update does not prove

The revision notes say this was a documentation review. A full run of checks with real Google or Microsoft accounts was not performed. An updated guide is therefore not evidence that saving and sending will work smoothly in your company's setup.

The sensible approach is to start small. Have the first documents returned inside ChatGPT and check the content. Once you test saving and sending, limit the agent to the folders and recipients it needs, and have a person review the early results.

One more local check: make sure the agent's idea of a business day matches your calendar. The note write-up specifically mentions Japanese public holidays and company non-working days as things to verify before relying on the schedule.