DATA SENSITIVITY

Treat sensitivity as a living control

A label matters only when it changes how information is handled.

“Confidential” is often painted onto documents after the real decisions have already been made. A useful classification program drives concrete behavior across storage, retrieval, collaboration, development, support, and deletion. It also recognizes context: a single attribute may be harmless while a combination of attributes can identify or expose an individual.

Keep the model understandable

Use a small number of levels with clear examples—such as public, internal, confidential, and restricted. Add handling rules for regulated or especially sensitive domains when needed. For every level, define approved locations, access criteria, encryption expectations, sharing channels, logging, retention, and disposal.

Enforce close to the data

Carry sensitivity metadata into indexes and retrieval. Filter before content reaches a model, not after an answer is generated. Prevent restricted information from entering unapproved analytics, test environments, consumer AI services, or verbose logs. Require step-up approval for bulk export and unusual cross-domain queries.

Design for mistakes

People mislabel files and copy data into new contexts. Combine training with detection: discovery scans, data-loss prevention, anomalous-access alerts, periodic entitlement reviews, and easy correction workflows. Measure unlabeled sensitive data, stale permissions, exceptions, and time to remediate—not simply training completion.

Classification should become more accurate over time. Owners need a way to challenge labels, tune automated detection, and reassess data whose purpose or risk has changed.

Reference: FTC guidance on protecting personal information.