Google announced at 21:51 UTC on 28 August that Gemini-based data classification in Google Drive has entered open beta. Files are read by a model and tagged with sensitivity labels, and those labels feed the controls that decide how the file may be shared, how long it is kept, and what an agent may do with it.
Who gets it
Enterprise Plus plus two adjacent top-tier plans. Not Business Standard, not Business Plus, not Enterprise Standard. Rollout runs to 30 September, so an administrator reading the announcement will in many cases find nothing in their console for weeks.
The line that matters
The announcement states that the applied label can be changed by the file's owner or an editor. That is a deliberate design choice with an obvious rationale — classifiers misfire, and a person who cannot correct a wrong label will route around the system entirely. But it means the label is advisory at the point of creation and authoritative everywhere downstream. The control that blocks external sharing, sets retention, or gates an agent's access is anchored to a value that the person most motivated to move the file can edit.
What the common framing gets wrong
"AI now classifies your Drive files" is being read as automated enforcement. It is automated suggestion with manual override, which is a materially weaker guarantee and is the difference between a compliance control and a compliance aid. Two further trims are getting lost: "available" means open beta, not general availability, and "available" means available to the customers on the highest tiers, which is a minority of Workspace seats.
How to read it correctly
As a labelling accelerator, this is a real advance — manual classification at scale has never worked, and a model that proposes labels across a whole corpus removes the reason most programmes stall. As the foundation for an enforcement policy, it needs the override path audited and, for the categories that actually matter, locked.