Databricks says LogSentinel turns LLM-based data classification into policy enforcement

What Databricks said on X

On March 27, 2026, Databricks described an internal system called LogSentinel that uses LLMs to classify columns, apply hierarchical and residency-aware labels, and continuously detect drift as schemas change. The company said the system creates tickets when it finds violations, and reported results of up to 92% precision and 95% recall for PII on 2,258 samples.

The wording of the post also matters. Databricks did not frame LogSentinel as a standalone public product launch. Instead, it said the work is informing Data Classification in ways that improve policy enforcement and compliance workflows. That suggests the company is connecting internal evaluation and operational tooling to productized governance features in Unity Catalog.

What the Databricks docs add

Current Databricks documentation says Unity Catalog Data Classification uses an AI agent and an LLM to automatically classify and tag sensitive data in catalog tables. The docs say the system can scan incrementally, surface results in a system table, and feed governance controls such as attribute-based access control (ABAC). In other words, the platform is not only labeling data; it is designed to let those labels influence downstream access and policy decisions.

Databricks' governed tags documentation adds the enforcement layer. Governed tags are account-level tags with predefined rules, allowed values, and permission controls. Databricks says they can be applied across Unity Catalog objects and then used for consistent classification, compliance, operational automation, and ABAC. The docs also note that tag data is stored as plain text and may be replicated globally, so administrators should not put sensitive information directly into the tag values themselves.

Why this matters

The higher-level signal is that enterprise data governance is moving away from static, manual tagging toward continuously updated classification tied to policy execution. Schema drift has always made metadata governance decay over time. If the labels stop reflecting reality, access controls and compliance monitoring eventually stop reflecting reality too.

Databricks is effectively arguing that LLM-assisted classification can close part of that gap, especially when paired with governed tags and ABAC-style controls. Inference from the X post and documentation together: the company's direction is to connect detection, labeling, drift monitoring, and enforcement into one operational loop. That is more ambitious than simply adding AI to metadata management, and it targets a real pain point for teams dealing with sensitive data at catalog scale.

Sources: Databricks X post · Databricks Data Classification docs · Databricks governed tags docs