Databricks announced Predictive AI Functions in beta on its official blog on October 14, 2026: four new SQL-native functions in one launch, plus an upgraded ai_forecast powered by a stronger time-series foundation model. The target problems are concrete ones, namely what sales look like next quarter, why a metric moved, and which days in a series look wrong. Those questions used to queue up for a data science team; now analysts can ask them directly in the SQL editor and in dashboards, and business users can ask in plain language through Genie, which calls the appropriate function for them.
Five functions, one per type of question
The split follows the business question. ai_forecast projects a time series into the future. ai_top_drivers identifies which segments most drove a change in a metric. ai_detect_anomalies flags outlier rows in temporal data. ai_predict_value and ai_predict_class fill in continuous values and categorical outcomes respectively, such as estimating a number or assigning a category. All of them are native SQL functions that can run across hundreds of millions of rows and feed straight into existing queries and pipelines. On availability, ai_top_drivers is usable in Genie today, while the other four are listed as coming soon there and are used mainly from SQL and dashboards for now.
Why the pitch is no tuning
The upgraded ai_forecast is powered by a transformer-based time series foundation model pretrained on a large, diverse collection of time series. The idea is that it learns patterns such as seasonality and trend the way a language model learns text, then produces forecasts on a company's own data with no tuning. The new version also handles variables businesses actually care about: built-in holiday support accounts for effects like Christmas dips and back-to-school spikes, and external covariates let teams feed in drivers such as pricing changes or marketing spend. On the fev-bench benchmark, Databricks says the new ai_forecast beats a leading cloud warehouse's comparable function and its own previous version on task win rate; that is a vendor-reported comparison, and any serious evaluation should backtest on the team's own history. The same week, Databricks also took its document processing GA, see Databricks Document Intelligence APIs going GA; together, both lines turn work that used to queue for specialists into default platform functions.
Keep the beta label in view
The caveats are equally clear. These functions are still in beta, and the models are maintained centrally by Databricks. That buys consistent results across dashboards, SQL and Genie and removes infrastructure work, at the cost of no model choice and no self-tuning, and limited ability to dissect a bad forecast the way a team could with its own model. The sensible role is exploration and routine operational analysis: use it to decide quickly what deserves a deeper look, then build dedicated models for the metrics that truly matter. For teams already on Databricks, the trial cost is genuinely low, a single SQL statement. For teams whose data lives elsewhere, this release alone is not a reason to migrate.
