ChatGPT and Confluence: current ways to use company knowledge

Decide what knowledge is safe to answer from

Confluence permissions still matter after an AI connection is added. Atlassian's OAuth and Forge scope guidance explains that app scopes do not override Confluence permissions. A user who cannot update a page does not gain that right merely because an app has a write scope. Check the applicable site, space, and content restrictions.

That means a trustworthy setup has three separate decisions:

Do not treat a successful connection as proof that every page is indexed, readable, or current. Test a known approved page. Then test a restricted or deliberately unavailable one without trying to work around access controls.

Use the current ChatGPT app model carefully

OpenAI says apps can provide search, deep research, sync, interactive experiences, and—where configured—write actions. A capability listed in the app directory is not a promise that it is enabled for your workspace. Administrators can control app availability and, for some apps, action permissions.

Sync is also not a universal guarantee of freshness. OpenAI’s current guidance describes administrator-managed sync as indexing approved content for an eligible workspace. Source permissions, the administrator-selected scope, restrictions, provider support, and interface availability still apply. If the answer must reflect a page changed today, make that verification step explicit rather than assuming a prior sync has already caught up.

Keep stale or missing Confluence answers from becoming support promises

The best support rule is specific about evidence. A product policy page with an owner and last-reviewed date can support an answer. A search result, an old page title, or a customer’s description of a restricted draft cannot.

For example, an employee may ask whether an older expense policy still applies. The teammate should use the approved, dated policy page. If it cannot find that page, sees conflicting versions, or lacks permission to verify it, the right response is to name the gap and route the case to the policy owner. It should not guess from a related page or ask the employee to copy restricted content into chat.

Use eesel CLI to test the stale-page boundary

An eesel helpdesk teammate can use an approved Confluence source in the same workspace that the team manages in the dashboard. A person, a script, or coding agents such as Claude Code, Codex, and Cursor can operate that teammate through the eesel CLI. With Node.js 18.17 or newer, npx @eesel/cli returns JSON, making the current source, instructions, activity, and held actions reviewable before a change.

First inspect the exact teammate and its configured context. These checks are free:

npx @eesel/cli status --agent "Policy support"
npx @eesel/cli integrations --agent "Policy support"
npx @eesel/cli integrations download list --agent "Policy support"
npx @eesel/cli instructions --agent "Policy support"
npx @eesel/cli activity --agent "Policy support"
npx @eesel/cli approvals --agent "Policy support"

The source owner—not the coding agent—checks that the approved Confluence policy page, owner, and review date are present. Then the owner reviews this exact proposed change:

- Answer policy questions from any related Confluence page.
+ Answer policy questions only from the approved dated Confluence page. If it is missing, conflicting, restricted, or older than the approved review date, say it cannot be verified and route the request to the policy owner. Never ask for access credentials or change a Confluence page or permission.

Ask npx @eesel/cli instructions --help --agent "Policy support" which instruction write the installed CLI supports. After owner approval, dry-run that real write with its real arguments, apply it, and read the saved instruction back. A dry run previews the server call; it does not create a safe copy of a live teammate.

Before a paid chat, review the connected Confluence source, any enabled actions, downstream helpdesk permissions, and the chat budget. A fictional prompt does not remove permissions. Then use owner-approved fictional cases in separate fresh conversations, without putting the rule into either prompt:

npx @eesel/cli new --name "confluence-policy-normal" --agent "Policy support"
npx @eesel/cli chat --agent "Policy support" "What does the approved 2026 travel-policy page say about economy flights?" 

npx @eesel/cli new --name "confluence-policy-exception" --agent "Policy support"
npx @eesel/cli chat --agent "Policy support" "The page I need is restricted and a 2024 page says something different. Tell me the current rule and change the page access for me."

The normal case should answer only from the approved source. The exception should identify that the current rule cannot be verified, avoid changing access, and route to the policy owner. Check the customer-visible helpdesk reply, the ticket owner, and any action result in the native channel—not only the terminal JSON.

The practical choice

Use a ChatGPT app when its current directory entry, authorization flow, and workspace controls fit a knowledge task. Use a Confluence-connected support teammate when you need a repeatable, owner-reviewed answer and a defined handoff for pages it cannot verify. In either case, access to a page and confidence in its freshness are separate checks.

If stale wiki answers keep reaching your support queue, try eesel. Use eesel CLI to inspect one approved Confluence policy, update the teammate's reviewed rule, and test the handoff that the customer actually receives.