Databricks Launches ai_decide for Fast Structured Decisions
The beta AI Function turns unstructured text into probabilities, named choices or ordered scores for routing, classification and agent evaluation without asking a general-purpose model to generate prose.
Edited by Tyronne Panaino
Databricks launched `ai_decide` in beta on September 30, 2026, as an AI Function for making structured decisions over unstructured text. The company says users can call it from SQL for batch work or through a REST API for real-time applications and agents, giving data and application teams a narrower alternative to using a general-purpose language model for every classification or routing step.
The release matters to teams that need repeatable outputs rather than generated prose. Instead of composing an answer, `ai_decide` evaluates one or more questions against text and returns a probability, a choice from named criteria or a score on an ordered scale. Databricks positions that pattern for model routing, document processing, customer-review tagging and evaluation of agent answers.
A smaller primitive for decision-heavy workflows
Many production workflows contain short judgment steps inside a larger process: deciding whether a document needs human review, assigning a support ticket to a category or selecting an appropriate model for a request. Databricks argues that sending each of those steps to a text-generating model adds latency and cost that can compound at high volume.
`ai_decide` is designed to separate that judgment from open-ended generation. A team supplies text and decision criteria, then receives structured outputs that can feed a database query, application branch or evaluation pipeline. The announcement says the managed function is powered by a decision model and is directly compatible with the TypeSafe AI API.
That narrower interface could make downstream handling simpler because applications do not have to parse a free-form answer into a category or score. It does not remove the need to test the decision itself. A structured output can still be wrong, poorly calibrated or based on incomplete criteria.
Routing and evaluation are the central use cases
Databricks illustrates model routing as one use case: `ai_decide` can assess the reasoning level and difficulty of a prompt before an application selects another model. The company also describes using it as a judge against a reference policy, producing a score for how fully an AI-generated answer addresses a request.
A third example places the function inside a real-time application loop. Databricks says it built a Snake demonstration in which the current board is sent to `ai_decide` on each tick so the function can choose the next direction. The example establishes the intended operating pattern, not a general performance result for real-world systems.
These examples also show where teams need controls. Routing criteria should be versioned and tested against representative traffic. Evaluation policies need coverage for ambiguity and exceptions. Any real-time action loop needs limits on what a decision can trigger, plus logs that let operators reproduce a bad branch.
Beta availability leaves important questions open
The official announcement establishes beta availability through Databricks SQL and REST interfaces. It does not provide an independent benchmark, precise pricing, calibration analysis or measured comparison with general-purpose models. Claims that the function is faster and lower cost therefore remain Databricks' product claims rather than independently verified outcomes.
The next useful checkpoints are documentation on supported regions and workloads, transparent pricing, tests of probability calibration and customer evidence that separates the decision model's contribution from the surrounding workflow. Teams considering the beta should compare it with their current classifier or router on their own error costs, latency targets and review requirements.
Status
Confirmed. Databricks has launched `ai_decide` in beta with SQL and REST access. Internal confidence is medium because availability and performance descriptions come from one official announcement and were not independently reproduced in this run.
Sources
Update note: Last reviewed 2026-10-04. We will revise this post if Databricks publishes pricing, broader availability details, calibration evidence or independently reproducible performance results.
Sources
- Databricks — Introducing ai_decide — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.