Databricks: ai_decide

A Databricks function built to be compatible. The entry copy describes ai_decide as a native SQL and REST function in beta that returns a probability, choice or score per question, and says teams already on the vendor API can use it directly. Compatibility is a claim in a product post.

No. The entry copy's caveat is explicit that the announcement gives no latency or accuracy number for ai_decide against hosted Jev, so the fast and low-cost framing in the post is comparative marketing without a published baseline.

Against governed data, which is the actual pitch. You call it from SQL over data already in the platform, or from REST for applications and agents. The tradeoff is that a beta inside a workspace means workspace permissions and region constraints come with it.

It is labeled beta in the entry copy, so expect interface churn. The post positions it for model routing, document processing and agent evaluation, which are the cases to prototype first rather than the ones to bet a critical path on.

The post names three: routing prompts to the right model, tagging customer reviews with metadata, and evaluating agent quality, plus analyzing customer reviews for recurring product issues. The framing is that a lot of teams run an LLM for questions like which category a support ticket belongs to or whether a document needs human review, and pay generation latency and cost for what amounts to a single bit.

For each question it returns either a probability, a choice from named criteria, or a score on an ordered scale, evaluated over unstructured text in a fraction of a second. That is the same three-way shape as the System One primitives, which is the mechanism behind the compatibility claim rather than a published mapping table.

It is a Databricks platform and products announcement published September 30, 2026 by Jane Zhang, Matthew Ding and Archika Dogra, and it presents ai_decide as a native AI Function inside the workspace rather than a separate service you call. That framing is what makes the governed-data pitch work, and it is also why workspace permissions and region constraints come with it.