Permissions, data and limits

What the app can access, what it stores, what goes into a prompt, and troubleshooting.

Permissions the app requests

Scope	Why it is needed
`read:servicedesk-request`	List service desks and request types for the configuration screens, and read the request being triaged and its participants.
`read:jira-work`, `write:jira-work`	Read issue data, comments and change history; assign the ticket, add a label, add a watcher, set the major-incident field, and link a cluster together. No issue is ever created.
`read:jira-user`	Resolve agents, teams and escalation users.
`storage:app`	Store configuration, decisions and statistics. The audit trail itself is Forge SQL, one row per event.
`report:personal-data`	Atlassian's Personal Data Reporting cycle. Once a day the app reports the account IDs it holds — across its own storage and the SQL audit table — and erases what belongs to an account Atlassian reports as closed.

The only outside address is api.atlassian.com: your own site's APIs and, for team routing, the Atlassian Teams API.

What goes into a prompt

The summary, description, comments and selected field values of the ticket being handled.
The list of candidate teams, people, request types and priorities, with the prompts you wrote.
Your dispatch, escalation or clustering instructions.

Prompts are processed by Forge LLM inside Atlassian. Nothing goes to an outside AI provider and nothing is used to train a model. Full detail in the privacy policy.

What the app stores

Configuration for all three agents, per service desk and per request type, plus the global Teams connection: email, organisation ID and API token.
The decision made for each ticket, stored against the issue key.
Skill descriptions, stored against the person's Atlassian account ID.
Daily statistics, audit entries, and the time of the last incident scan per desk.

What happens without an active licence

Triage stops. Everything else stays open. All three agents run from Jira event triggers and a scheduled trigger, and every one of those declares appIsLicensed: true in the manifest — so Forge does not invoke them at all. There is no half-run and no error on a ticket: new tickets simply arrive and sit there, exactly as they did before the app was installed.

Nothing in the admin screens is gated, and that is deliberate. Every configuration screen, the statistics, the audit log, the Connection tab, the issue panel and every decision already recorded stay fully readable and editable. A banner at the top of the app says the licence has lapsed and that triage is paused, because the failure mode this protects against is an administrator hunting a broken routing prompt for an hour. Nothing is deleted; renewing resumes triage on the next ticket, with no reconfiguration.

The silence is the symptom

Because the platform declines to invoke the triggers, an unlicensed install looks identical to one where every agent was switched off — no failures in the audit log, because nothing ran to fail. If routing stopped and nothing else changed, check the banner first.

Uninstalling

Uninstalling erases what the app holds. A preUninstall handler empties the audit table first, then sweeps the key-value store — configuration, the Teams API token, per-issue decisions, skills and statistics — repeating until a pass finds nothing left, to a 45-second budget.

Assignments, labels, watchers, incident links and major-incident flags the agents set stay in Jira, because they are ordinary Jira data. See Where your data goes.

Troubleshooting

Nothing is being routed: Check in order: there is an active licence — without one the triggers are never invoked and every agent goes quiet at once, which is the banner at the top of the app; Enable Dispatcher Agent is on for that project; at least one team is selected in Step 1 — Teams; each selected team has a routing prompt; the Only process tickets matching this JQL filter is not excluding the tickets; and the global Connection tab shows as connected. Also remember the Dispatcher only acts on tickets created after it was enabled.
All three agents stopped at the same moment: That is the shape of a lapsed licence rather than a configuration mistake. One agent breaking is a prompt or a setting; all three going silent together, with nothing in the audit log, is the platform declining to invoke the triggers.
Tickets go to the wrong team: Open the AI Triage panel on the ticket: it names the team, the reason and any warnings. Then fix the team prompts, focusing on the boundary between the two teams being confused.
No teams appear in Step 1: The Teams connection is missing or wrong. Go to Jira → Apps → AI Triage → Connection. The token must belong to an account with organisation access; if the organisation ID was discovered incorrectly, override it with the value from admin.atlassian.com.
Escalation never triggers: Common causes: Enable Smart Escalation is off for that project, the threshold is set to critical only, no action is ticked under Escalation Actions, the JQL filter excludes the ticket, or the comment was internal rather than from the customer. If the label appears but nothing is reassigned, the desk has no escalation user — the reassign and watcher actions are skipped without one.
Escalation triggers too often: Raise the threshold to critical, and add custom instructions describing what normal tone looks like for your customers.
Incidents are created for unrelated tickets: Raise the Cluster Threshold, shorten the Time Window, and add clustering instructions naming what must never be grouped.
Incident Detection will not save: Two fields are required once it is enabled: Link Type, and Which issue types are incidents. The screen names whichever is missing.
A slow-building wave was caught late: The recurring scan runs hourly at its fastest, so a trickle of tickets that never trips the per-ticket check waits for the next scan. Widen the Time Window rather than expecting a shorter interval — an hour is the floor.
The agent tabs are missing from the global page: They are not there by design. Jira → Apps → AI Triage has only Statistics, Connection and Audit Log. The agents are configured in Project settings → AI Triage, per project.

Include the issue key and the service desk when you open a ticket. The stored decision for that issue is the first thing we read.