Oracle Analytics puts AI on both sides of the query: authoring and metadata

Oracle Analytics Cloud’s September 2026 update does more than add another conversational box. It introduces AI at three distinct points in the analytics workflow: writing filter expressions, documenting datasets for agents, and suggesting insights to workbook consumers. That separation matters because each feature has a different control surface—and a different failure mode.

Natural language becomes an authoring aid

Expression Assistant lets workbook authors describe a filter in natural language. Oracle says the assistant generates an expression using fields from the selected data source; the author can apply the result or refine it with a follow-up request. The feature works in workbook and visualization filters, while administrators select the registered language model that powers it.

That design stops short of autonomous execution. The generated expression remains an artifact the author can inspect before applying, and model selection stays at the administrative layer. For deployment teams, the sensible test is therefore not just whether a prompt produces valid syntax, but whether the proposed filter preserves the intended scope, date logic and null handling.

Metadata becomes part of the agent interface

The more consequential addition for data agents may be AI-generated descriptions for datasets and columns. Oracle says authors can describe the dataset as a whole and distinguish fields with similar names, different forecast meanings, fiscal-calendar formats or units of measure. Generated descriptions are explicitly a starting point for authors to review and refine.

Oracle also exposes indexing configuration for subject areas and lets teams import or export it. This turns semantic context into something that can be managed rather than left inside a prompt. It also makes ownership unavoidable: generated descriptions can accelerate documentation, but business definitions still need a reviewer who knows which forecast, calendar and unit is authoritative.

A related modeling feature lets datasets inherit descriptor-ID relationships from a semantic model or deployed RPD. Users can work with a friendly label such as a product name while Oracle uses the corresponding ID for query processing; connection-based datasets can configure the relationship manually. That is useful context for teams evaluating text-to-SQL accuracy, because resolving business labels to stable identifiers can reduce ambiguity before SQL generation begins.

Consumers get suggestions, not a blank canvas

Auto Insights can now be enabled for read-only consumers through a workbook’s presentation settings. Authors can place suggested views—including trends, seasonality and growth-contribution bridges—on the canvas. Consumers can influence recommendations by selecting visualization types, then export or share the resulting insights.

The deployment takeaway is to evaluate these three surfaces separately. Expression generation needs syntax and intent checks. AI-authored metadata needs stewardship and change control. Consumer recommendations need presentation-level permissions and review. Oracle has put all three in one release, but they should not share one blanket “AI enabled” decision.