Predictive AI arrives in BigQuery

Google now offers a pre-trained foundation model for tabular data in BigQuery, enabling enterprise teams to generate predictions from structured data without extensive machine learning training. This integration means businesses can leverage predictive AI directly within their data warehouse environment. The new feature significantly streamlines the process of extracting insights from large datasets.

The TabFM model, developed by Google Research, is currently available in preview. It performs classification and regression, which are two fundamental types of predictive machine learning. Classification predicts categorical outcomes, such as whether a customer will churn or if a transaction is fraudulent. Regression, conversely, forecasts numerical values, like a customer’s expected lifetime value or the amount of a future insurance claim.

Traditionally, enterprise teams following a typical machine learning workflow needed to train and deploy a distinct ML model for each specific use case. TabFM, however, uses historical data as in-context examples to generate predictions. This approach significantly simplifies the process compared to building and maintaining custom models for every scenario.

Teams can provide labeled historical data, where outcomes are already known, alongside new data requiring predictions. BigQuery’s AI.PREDICT function then utilizes this information to produce forecasts. The function learns patterns from the historical data and applies them to the new data, eliminating the need for separate model training and deployment.

TabFM automatically handles data expressed in categories, such as customer type or product category, and datasets with missing values. This capability reduces the preparation work typically required before generating predictions. Google also provides an AI.EVALUATE function to assess the model’s performance. Teams can compare TabFM’s predictions with known outcomes to measure accuracy, determining the usefulness of the predictions for specific applications.

Efficiency and Cost Benefits for Enterprises

The integration of TabFM into BigQuery offers substantial benefits for enterprises, particularly in terms of accessibility, efficiency, and cost. It democratizes predictive analytics, making it available to teams that may not possess extensive machine learning expertise. This reduces the engineering and infrastructure overhead associated with developing and maintaining traditional ML models.

Amit Kumar Jena, AI development head at IT consulting firm Kanerika, emphasizes that individuals proficient in SQL can now directly generate classification or regression predictions against tables already residing in BigQuery. This process uses the two new functions without requiring prior model training, tuning, or deployment. Crucially, it avoids moving data to a separate ML platform, which simplifies governance and lowers the infrastructure costs of running parallel systems.

Manoj Chandra Jha, principal analyst at Nord-IQ Research, highlights that TabFM consolidates the traditional multi-persona workflow. This workflow typically involves an analyst, a data scientist, and an ML engineer. With TabFM, this complex process collapses into a single SQL call, potentially reducing headcount requirements and increasing overall operational efficiency. The ability to perform predictive tasks directly within BigQuery streamlines operations and empowers a broader range of personnel to leverage advanced analytics.

This consolidation of tasks means that businesses can accelerate their decision-making processes. They can quickly iterate on different predictive models and scenarios without significant delays or resource allocation. The reduced need for specialized ML engineers also allows existing teams to expand their capabilities, fostering a more data-driven culture across the organization. By simplifying the pipeline, Google aims to make advanced predictive capabilities an intrinsic part of everyday business intelligence.

Considerations for Implementation

Despite the clear advantages, enterprises should understand the limitations of the new TabFM model before completely replacing existing machine learning workflows. Google’s own documentation suggests that traditional models, such as XGBoost, remain appropriate for specific scenarios. These include very large datasets, datasets with more features than TabFM’s current limit of 20 feature columns, or when complete control over custom hyperparameter tuning is necessary.

Google advises using traditional models like XGBoost when very large historical datasets are involved, or when there is a need for feature-importance explainability, which identifies which input features contributed most to a prediction. These scenarios often demand a level of granularity and control that a pre-trained foundation model might not fully offer. The company’s guidance underscores that TabFM is a powerful tool but not a universal solution for all predictive analytics challenges.

Analysts also point out that in large datasets and high-frequency workloads, the cost advantages of TabFM can diminish. While TabFM reduces the initial cost of setting up a model-training pipeline, Google plans to introduce token-based pricing for TabFM starting October 30, 2026. This new pricing structure will be in addition to standard BigQuery charges. This change could lead to higher overall costs for workloads that generate predictions at scale when compared with a previously trained and cached model, as Jena explains.

Jha advises enterprises to pilot TabFM on their specific workloads and carefully analyze the costs based on actual query volumes. Comparing these costs with their existing ML infrastructure is crucial before allocating significant ML budgets to the service. This cautious approach ensures that businesses make informed decisions about where and how to best deploy TabFM. Jha concludes that TabFM is more of a workload-specific tool than a wholesale replacement for traditional ML models. It is better suited for tasks such as ad hoc analysis, experimentation, and lower-frequency business use cases where rapid insights are prioritized over deep customization or extreme cost optimization at massive scale.