The assistant can read the campaign brief. Does that mean it may edit the sponsor's document, publish the clip, or send the invoice? Write those as separate permissions before the campaign starts. A permission map names the account, source, allowed change, approval point, and person responsible for the final action.
This week's October 5–9 OpenAI dots update broadens ongoing-work coordination. The announcement makes account boundaries worth reviewing; it does not make every connected app an unrestricted publishing channel.
A campaign map with actual boundaries
This fictional campaign uses a product explainer and a sponsor review. The matrix is an example you can adapt, not evidence that any app or agent implements these exact controls automatically.
| Work item | Source and account | Permitted preparation | Approval before |
|---|---|---|---|
| Research claims | Official public pages in the selected research tab | Read and produce a linked evidence list | Adding unsupported claims to the brief |
| Edit the script | Campaign folder in the creator's account | Propose tracked changes in a draft | Replacing the approved master |
| Sponsor review | Sponsor correspondence visible through an approved connection | Draft a concise reply | Sending to the named sponsor contact |
| Upload the clip | Creator's channel account | Prepare metadata and a checklist | Publishing or scheduling the video |
| Check payment details | Creator's own billing records | Summarize a discrepancy for review | Any account or payment action handled by the owner |
Add a destination column if files can leave the workspace. “Save the draft” should mean a known folder and version, not whichever shared drive happens to be open. Keep sensitive finance work outside a general campaign task whenever the creative result does not need it.
Check the account before checking the content
OpenAI's computers and apps documentation explains that cloud and personal-browser sessions need their own sign-ins. A connected personal computer becoming offline does not revoke its connection; revocation is a separate control. Plugin access depends on the account and permissions enabled.
That distinction is easy to miss when two browsers display the same website. In your workflow, verify the visible account, organization, folder, and environment before any write. A personal channel and a sponsor's channel can look similar enough for an operator to select the wrong destination.
Record the expected account in the permission map. If the session differs, stop the dependent action and restore the intended account. Do not fill the gap by copying private material into another environment without a deliberate decision about where it may go.
Draft authority and send authority are different
OpenAI's dot controls explain that a request to draft a reply does not itself authorize sending it. Custom rules and plugin permissions also serve different purposes. A written instruction can define scope, while access controls determine which actions a connected service permits.
Apply that distinction to the campaign: “Draft a reply for the producer to review” has an internal output. “Send this approved reply to this contact” identifies an external action and recipient. Do not let the word “handle” stand in for both.
Before approving a send, check the recipient, attachments, claims, dates, and final text. Approving one reply need not authorize all future replies. If ongoing authority is intentional, define its audience, period, limits, and exceptions so the reviewer can understand what is covered.
Use one selected tab for a bounded browser task
If the browser assistant is enabled for your Growit account, follow its current documented account and extension requirements. Select the intended tab and grant only the permissions required for that task. Keep the task concrete: read this official announcement and draft five claims with source links.
Review the output against the page. A page's text can provide evidence, but instructions embedded in it should not expand the task or grant permission to send data elsewhere. If the next step needs a different tab or account, inspect that surface and its permissions before proceeding.
A public research tab is a good first exercise because the evidence boundary is visible. Once the workflow is understood, add a private draft with a specified output folder and approval rule. Expand access only when the result requires it.
Test permissions with a harmless draft
Use a throwaway internal draft and ask for one reversible edit. Confirm the correct destination and whether the expected approval appears. Next, request a proposed outbound message while keeping it unsent. The experiment should demonstrate the distinction the matrix describes.
If a required approval does not appear, inspect the settings and any allowed-action rule before trusting the workflow. If access is unavailable, record it as unavailable rather than treating a draft as proof the external action succeeded. Keep a decision receipt for changes to the campaign's scope.
Keep o1's product boundaries visible
Check o1's current product guides for account, platform, and release availability before applying this method. Review results from the tools you can actually access, and keep external actions within the permissions you grant. OpenAI dots features described here are separate vendor capabilities; the physical Growit o1 remains in development, as explained in o1 specifications.
Before the next campaign handoff, fill the map for one account and one destination. Verify a read, a draft edit, and an approval boundary. Then review the scheduled-task audit if that permission will be reused automatically.