Google Drive Gemini data classification review: faster DLP labels, narrower control surface

Google's August 28 open beta for Gemini-based data classification in Drive is not a consumer AI feature. It is a Workspace admin tool for a specific, high-friction problem: getting useful sensitivity labels onto large numbers of Drive files before data loss prevention, retention, audit, and agent access policies can do their work. On paper, the pitch is strong. Instead of building a custom model from manually labeled examples, admins can define plain-language instructions and let Gemini inspect Drive content and apply classification labels.

That makes this release most interesting as a comparison with Google's existing custom-model AI classification path. The older approach gives organizations a model trained on their own examples, but Google's documentation says it requires designated labelers and at least 100 training files per label option before a model can be trained. Training can take 4 to 6 hours, and Google recommends waiting for stronger scores before broad auto-apply. For a three-option label such as Public, Internal, and Confidential, that means hundreds of correctly labeled files before the system is useful.

The Gemini beta removes much of that startup burden. Admins choose a label, write instructions for each eligible label option, pick which options Gemini may auto-apply, and scope the files by owner or shared drive. Google says editors and owners with the right permissions can review, accept, or modify Gemini-applied labels, and audit logs record labeling activity. That human review loop is important: this is not a magic compliance switch, and it should not be treated as one.

What Works

• Lower setup cost: The biggest improvement is replacing training-file collection with administrator-defined instructions.
• Useful policy fit: Labels can feed DLP, retention, audit investigation, and access controls already used in Workspace environments.
• Admin control remains central: Google says admins select labels, instructions, eligible options, and the user or shared-drive scope.
• Auditability is built in: Label application and user changes are exposed through Drive events and related admin metrics.

Where It Falls Short

The limits matter. The Gemini method is in open beta, not general availability. It is also restricted to supported paid editions listed by Google: Enterprise Plus, Google AI Pro for Education, Frontline Plus, and certain add-ons described in the help documentation. Organizations outside those tiers should not plan around it yet.

The feature also has structural constraints. Google's help page says organizations can create a total of five custom models and Gemini instructions, including up to one Gemini instruction. Files must be in shared drives or owned by users with supported licenses, and inactive file scanning with Gemini instructions must be enabled manually. Conflict handling is rule based: data protection rules override AI classification, user-applied labels take priority over AI labels, and label-option order can decide conflicts between AI sources. Those are sensible controls, but they add setup details that security teams must test carefully.

Verdict

As a review score, Gemini-based Drive classification earns a 7.7 out of 10. It solves a real bottleneck for Workspace admins who already depend on classification labels, especially teams trying to reduce sensitive-data exposure before agentic workflows spread across company documents. Compared with the custom-model route, it looks faster to start and easier to tune in language that policy owners understand. The caveat is that beta status, edition limits, one-instruction capacity, and the need for careful label governance keep it from being a universal recommendation.

Choose it if your organization already uses Google Workspace security tooling and needs quicker label coverage across Drive. Skip it for now if you require mature model performance reporting, broad multi-policy experimentation, or support outside Google's eligible Workspace tiers.

Sources

• workspaceupdates.googleblog.com
• knowledge.workspace.google.com

Cover photo by Zulfugar Karimov on Pexels, used under the Pexels License.