By Greg Nowak. Updated 12 November 2026.
Connecting ChatGPT to your documents, CRM or project platform can make account briefings and weekly reporting easier. Before you connect it, write down whose account it will use, which information it can retrieve and what it can change. Those answers belong in a permissions map that the process owner can understand.
The useful outcome is a pilot with clear boundaries: enough access to prove the workflow, a way to check its behaviour and someone responsible for switching it off.
Start with one job the team needs done
“Connect our CRM” leaves too much undecided. A better pilot brief is: “Let five account managers retrieve approved account summaries and draft follow-up notes for review.” Sending messages and updating records stay outside that initial scope.
For an agency, add the client boundary: which client’s material may enter the workflow, and who may receive the output? A shared reporting process should not quietly become a search across every client account.
Choose a practical success measure before setup: can someone produce a correct briefing, with traceable source material, using less manual searching? Include the time spent checking and correcting it.
Separate the controls that govern access
OpenAI’s plugin controls guidance separates installation, app access, permitted actions, confirmation settings and source-system permissions. A plugin can bundle instructions and connected apps; making it available does not automatically authorize its underlying connections.
Review each required app in the Admin console’s Plugins area. Check App access, Tools or Actions, New tools or New actions, and Permissions where those controls are available. Options vary by workspace and integration. ChatGPT Business uses workspace-wide availability; do not assume it offers Enterprise-style role targeting.
Then inspect the connected service. The OAuth consent screen describes requested access, while the connected account’s permissions determine which records are available within that access. Additional content restrictions may narrow the result further.
Build a one-page permissions map
Use this table as a working template. Fill it in with the operations lead and whoever manages the source system. Record actual settings and account names alongside the intended boundaries.
| Decision | What to record | Suggested pilot boundary |
|---|---|---|
| Purpose | Task, output and success measure | One repeatable briefing workflow |
| Audience | Who can use the app and connection | Named pilot users, using supported controls |
| Identity | Individual or shared account; accountable owner | A company-managed account with limited access |
| Information | Allowed folders, records, fields and client boundaries | Approved summaries; exclude unrelated material |
| Actions | Read, create, edit, send, share or delete | Read and draft; no source changes |
| Approvals | Which operations require confirmation | Review before sending or changing records |
| Data lifecycle | Retrieval, indexing, saved outputs and deletion paths | Document each location and retention rule |
| Operation | Business owner, technical owner and review date | Tested shutdown steps before launch |
Check whose permissions the connection uses
A personal connection uses the connected employee’s source permissions. A shared or agent-owned connection uses its own account’s permissions, which may be broader than those of the person requesting the work. OpenAI makes this distinction explicit in its cloud security guidance.
For example, a shared account able to read several client folders needs careful audience restrictions. Test the connection as an ordinary pilot member rather than relying on an administrator’s successful demonstration.
Keep reading, drafting and acting separate. Drafting in the conversation can leave the source unchanged; creating a draft inside a CRM or mailbox is already a write operation. When writes become necessary, allow only the required actions. Verify the confirmation policy: an authorized write can proceed without an individual prompt when settings permit it.
Test the boundaries before expanding the pilot
Use disposable records and representative accounts. Keep expected and actual results together so the process owner can review them.
- Allowed access: retrieve an approved record and verify the answer against the source.
- Blocked access: request a restricted folder or another client’s record. Confirm it remains inaccessible through this connection.
- Permission changes: remove access, send a fresh request and check when the change takes effect.
- Action limits: attempt an excluded edit or send operation. For enabled writes, check the confirmation and the resulting source record.
- Hostile content: place instructions in a test document asking the assistant to disclose unrelated information or use an unapproved tool.
For a custom Model Context Protocol (MCP) integration, enforce authorization and scopes on every tool call, validate inputs on the server and return only necessary fields. Keep credentials out of prompts and tool output. OpenAI’s security guidance also recommends human confirmation for irreversible operations and logs that avoid unnecessary personal data. Instructions to the model cannot replace these controls.
How do you stop access and handle existing copies?
Record whether the app retrieves information when requested or maintains a synced index. OpenAI notes that synced permission changes can take time. Disconnecting an app also does not automatically delete information already saved in conversations, generated files or other retained records.
Write shutdown steps for the actual workflow: disable app access, disconnect the relevant account, revoke its source authorization where appropriate, and stop related scheduled tasks. OpenAI’s admin revocation guidance treats these as separate operations. Review retained copies through their own deletion processes.
Leave the team with something they can maintain
Keep the map beside the workflow instructions. Review it when people change roles, client access changes or the app gains new actions. Before expanding, ask the business owner to explain what it reads, what it changes and how to stop it.
If you want help moving from an interesting demo to a controlled business workflow, talk to Greg about your AI rollout. Bring one candidate process; the starting point is its permissions map, pilot tests and ownership.
Related on GrN.dk
- ChatGPT Apps with Full MCP Access: Governance Comes First
- Search Console Can See Social Posts—Your Reports Need a New Map
- Before an OpenAI Agent Touches Your CRM, Map the Boring Boundaries
Need help with this kind of work?
Discuss your AI rollout with Greg Get in touch with Greg.