BigQuery’s new AI.PREDICT turns agent answers into zero-shot forecasts

Google Cloud has put a zero-shot tabular foundation model behind two native BigQuery SQL functions, extending what a conversational data agent can do beyond retrieval and aggregation. In preview now, AI.PREDICT performs regression or classification from historical examples, while AI.EVALUATE scores the result against a test set. Google explicitly positions TabFM alongside the BigQuery MCP server for agentic applications.

The execution path gets shorter

A conventional predictive workflow typically separates feature engineering, model training, tuning, deployment and inference. TabFM instead receives labeled historical rows and target rows as in-context examples, then produces predictions in one forward pass. In BigQuery, the caller supplies a training table or query, a prediction table or query and the label column; numeric labels trigger regression, while string or Boolean labels trigger classification.

That matters for analytics agents because a question such as “which accounts are most likely to churn?” can now resolve to a governed SQL function rather than an external model endpoint. The output stays tabular: regression returns a predicted value, and classification returns both a predicted label and class probabilities. Google says distributed BigQuery inference can process prediction tables with millions of rows in minutes.

Zero training does not mean zero evaluation

The operational shortcut creates a new responsibility for the agent layer. A generated AI.PREDICT call can be syntactically valid and still be the wrong modeling choice. Google’s own guidance recommends TabFM for small-to-medium historical datasets and rapidly changing tasks, while retaining traditional models such as XGBoost when teams need more feature capacity, hyperparameter control or feature-importance explainability.

The preview also has hard boundaries: inputs are limited to 20 feature columns and classification to 10 categories. The paired AI.EVALUATE function should therefore be treated as a release gate, not an optional follow-up. An agent can generate precision, recall and F1 for classification or regression metrics such as mean absolute error before a prediction is exposed to users or automation.

Watch the second meter

Pricing is another reason to keep prediction distinct from ordinary text-to-SQL. During preview, Enterprise editions bill through slots and on-demand users pay by bytes processed. Google’s documentation says token-based TabFM pricing begins October 30, 2026, alongside the existing slot or byte cost for the rest of the query.

For practitioners, the useful pattern is a two-stage agent: first resolve the governed population and features, then invoke prediction only when the request genuinely asks for a forecast or classification. Log the generated SQL, evaluation result and inference spend separately. TabFM collapses the modeling path; it does not collapse the need for validation, explainability or cost controls.