Atlassian's full AI opt-out costs an Enterprise upgrade: three of four tiers can't switch metadata off

The part that stays on is the derived layer, and it repays reading item by item. According to Atlassian's data-contribution documentation, as summarised by dev.to, metadata contribution covers readability scores, task classifications (that a ticket is "sales work", for instance), story points, sprint end dates, SLA values, and semantic-similarity measures pulled from the Teamwork Graph ^[7]. None of that is anyone's prose. It is a working model of how an organisation estimates and where its estimates fail, and it is the layer an admin below Enterprise cannot switch off ^[5].

What the content setting covers is easy to picture: Confluence page titles and body text, Jira work-item titles, descriptions and comments, and custom status and workflow names ^[6]. That is material an admin can imagine reviewing. The metadata list is harder to picture and harder to audit, and it is the one that is locked ^[3].

So the escalation path for an admin who wants the derived signals to stop does not exist inside the admin console. Free, Standard and Premium are three of the four named plans, and none can reach a complete off state by configuration ^[13]. A Premium admin who flips every switch on offer is still contributing, and the remedy sits on the price list rather than in the settings page ^[15]. That moves a governance requirement into procurement, where the approver is different and the clock is a renewal cycle.

Credit where it is owed. The opt-out is real and documented, and the path is Atlassian Administration, then Security, then Data contribution ^[8]. Atlassian has said the collection feeds product improvement, and Rovo competes in a market where an assistant that knows your workspace beats a generic one ^[12]. Retention is bounded on paper too, with in-app data leaving the improvement datasets about 30 days after an opt-out and content attributes at about 90, retraining to follow ^[9].

The bound has a tail. Aggregated, de-identified data may persist for up to seven years ^[9], roughly 28 times the 90-day window that governs content attributes ^[14]. The short clocks apply to material Atlassian can still tie to a customer. The long clock applies to what it says it cannot.

The distribution of all this is the part worth arguing about. Non-admin users get no switch, and their data's fate is settled a level up by someone who first has to notice the change happened ^[10]. Set that against the tier matrix and the customer contributing the most by default is the one paying Atlassian the least ^[11]. That is not an artefact of engineering, per dev.to's reading; it is a decision about where privacy sits on the price list ^[11].