Databricks launched ai_decide on 30 September 2026, adding a beta AI function that takes unstructured text and returns a structured decision. The company says it can answer questions against that text in a fraction of a second, at lower latency and cost than a large language model used for a similar task.
Key points
- For each question, ai_decide can give a probability, select one of several named criteria or assign a score on an ordered scale.
- The beta function can be called from SQL for decisions across data or through a REST API for real-time applications.
- Databricks says its decision model avoids generating text for tasks such as classifying reviews and routing prompts.
ai_decide turns questions into structured answers
The distinction is in the answer the function is asked to produce. A language model might write a response to a support ticket, even when the immediate task is only to choose its category. Databricks says ai_decide instead evaluates one or more questions against the ticket’s text and returns a decision for each. Depending on the question, that answer can be a probability, a selection from named options or a score on an ordered scale.
It is closer to sorting letters into labelled trays than drafting a reply to each letter. The labels still have to fit the question: Databricks gives the examples of assigning a support ticket to a category, deciding whether a document needs human review and choosing which model should handle a prompt. The company says its decision model avoids the text-generation work of a large language model when the required output is one of those narrower answers.
Sorting customer reviews could work in the same way. Words describing a problem could become a tag for the main issue and an indication of whether the customer is looking to return the product. Those tags could then be grouped by product and month to find recurring complaints alongside mentions of returns. That use follows Databricks’ example of applying the function to reviews, rather than asking it to compose a response to each one.
Databricks says teams often put language models to work on classification and routing, although those tasks do not necessarily call for complex reasoning or generated prose. It argues that the extra latency and cost accumulate when such decisions are made at high volume, including in document-processing workflows. Its speed and cost comparison for ai_decide is a company claim about similar tasks.
SQL and REST offer two routes to ai_decide
ai_decide is available in beta as a native Databricks AI Function. SQL calls are intended for decisions across governed data at batch scale, while a REST API lets an application request a decision as it runs. That gives developers two ways to use the same kind of structured output: applying questions across stored text, or asking them when a new input arrives.
The SQL example starts with customer reviews. Databricks says a single query could classify each review by the problem it describes and flag requests for a return, then group the results by product and month to trace recurring complaints. The function’s role in that workflow is to turn the wording of individual reviews into fields that can be counted and compared within the data.
Another proposed use is choosing among models for an AI assistant. Databricks describes assessing a prompt’s difficulty and reasoning level before sending it to a corresponding model. It also proposes using ai_decide to judge generated customer-support answers against a refund policy, checking whether each answer follows the policy and scoring how completely it addresses the request. Those are examples offered by the company, rather than reported deployments.
Databricks says ai_decide is directly compatible with the TypeSafe AI API. Its announcement identifies TypeSafe AI’s Jev as an example of the decision-model approach: given text and questions, such a model produces decisions and probabilities instead of generating prose. For teams already building against that API, compatibility makes the new function an option within the Databricks platform.
Databricks puts ai_decide behind Snake
For a real-time demonstration, Databricks built a live game of Snake hosted on Databricks Apps. At each tick, the application sends the current board to ai_decide and asks which direction the snake should move next. Databricks says the decision loop completes in a fraction of a second, allowing the next move to depend on the board as it stands at that moment.
The company proposes the same pattern for an application choosing its next action, or an agent deciding which tool or branch to use. In each case, the requested result is a choice that can be acted on immediately, rather than an explanation written for a person. The Snake game demonstrates that pattern with a changing board and a direction to choose on every tick.
Databricks also says developers can serve open-weight decision models and run them directly in SQL on its platform, beyond using the managed ai_decide function. The company points to an open-Jev walkthrough for that route.
