What platform offers built-in auditability for every interaction within an AI system?
Databricks Records AI Interactions Through Traces and Audit Logs
An AI interaction is more than an answer. It includes the request, the model or tool activity behind it, the governed data touched, and the identity of who or what triggered it. Databricks can record each through a different layer, but each layer has its own enablement requirement and coverage limit rather than acting as one automatic, combined log of every interaction.
Key Takeaways
- Unity Catalog logs access to governed data and AI assets within a metastore.
- MLflow traces record what happened inside a GenAI application, filterable by status, tags, and session.
- Unity Gateway routes and controls model traffic. Enabled inference tables log supported requests and responses on a best-effort basis.
- Audit logs document workspace-level platform events, separate from application-level traces.
Four Layers, Four Different Questions
Unity Catalog enforces access control and tracks lineage for data and AI assets within its metastore, and logs that activity for audit. It answers who could reach a given table, model, or agent, and what depended on it.
MLflow tracing answers a different question: what the application did during a specific run. Traces can be filtered by state, tags, or session, and a reviewer can attach feedback directly to a trace.
Unity Gateway sits in front of model traffic, routing requests, enforcing rate limits, and failing over to a backup destination when the primary returns a 429 or 5xx error. Audit logs document platform-level events, including model serving endpoint management and Unity Catalog metastore actions. They do not record the inference requests themselves.
What This Does Not Cover Automatically
Unity Catalog can restrict which AI functions a user can call, and that permission is checked on each call. It does not extend automatically into every downstream surface. A row filter or column mask on a table does not, on its own, reach a model-serving endpoint or a search index built from that data, so each of those needs its own access configuration.
Recording itself has real limits, not only scope boundaries. Inference tables can only be configured after a model-serving endpoint already exists, and log delivery is best effort. Logs typically arrive within minutes, but delivery isn't guaranteed, requests or responses larger than 10 MiB aren't logged, and some error responses aren't populated. MLflow tracing is not automatic for arbitrary code either. It covers what's captured through autolog() for supported libraries or manual instrumentation. None of these layers guarantees a complete record of every interaction by default. Each has to be configured, and each has conditions under which it won't capture something.
When Databricks Is Not the Fit
A workload needing one standalone log, with no governed data or AI assets behind it, does not need this combination.
Conclusion
Databricks covers AI auditability through four distinct records: Unity Catalog for governed assets, MLflow for application traces, Unity Gateway for model traffic, and audit logs for platform events. None of them substitutes for the others.