Databricks puts a decision model inside SQL—but production teams still own the threshold

Databricks has launched ai_decide in beta, exposing a decision model through SQL and a REST API for classification, scoring and routing tasks that do not require free-form generation (Databricks announcement). The practical shift is simple: instead of asking a chat model to generate and then parse an answer, teams can define a closed set of questions and receive structured probabilities directly.

What the function returns

A single call evaluates one shared input—called state—against one or more questions. The supported question types are noul, which estimates the probability that a condition is true; choice, which selects among named labels; and score, which places the input on an ordered rubric (function reference). Results arrive as a variant value containing answers, metadata and an error field, so downstream SQL can filter, aggregate or route on the output (Databricks docs).

Databricks positions the beta for support-ticket routing, document review, model routing and AI-answer evaluation. The company says the TypeSafe Jev-powered function can close a decision loop in a fraction of a second and costs less than using a general-purpose LLM for comparable structured tasks, but the launch post does not publish a reproducible benchmark or price comparison (launch post). Treat those performance claims as vendor-reported until workload-level evidence is available.

The output is not the policy

The function does not remove the hardest production decision: what should happen at each probability or score. A needs_review probability of 0.62 is not itself an approval rule. Teams still need thresholds, an abstention band, a human queue and a record of which function version and rubric produced the decision.

That boundary matters because Databricks says generated answers can vary between calls. For choice, the response includes every label’s probability plus a confidence value; for score, it returns a probability-weighted average and the full distribution across rubric levels (return schema). Those distributions make uncertainty visible, but application code must decide whether uncertainty triggers fallback, review or rejection.

Deployment constraints are part of the design

The beta is available only in some regions, must be enabled through preview management, does not run on Databricks SQL Classic and requires Databricks Runtime 15.4 LTS or later; Databricks recommends 18.2 or later (requirements). Databricks says input parameters stay within its security perimeter and are not stored, although runtime metadata is retained. The underlying model is currently covered by Apache 2.0 terms, and the vendor reserves the right to change the model if another performs better in its internal benchmarks (security and model terms).

For practitioners, the right rollout is therefore not “replace every classifier.” Start in shadow mode, store the probability distribution beside the existing decision, compare errors by class, and define where humans take over. ai_decide makes structured judgment easier to call from the warehouse. It does not make the resulting business rule deterministic.