Published
Updated
Databricks posts ai_decide, a Beta function that returns a noul, a choice, or a score
Databricks posted ai_decide on October 1, 2026. The blog, dated September 30, calls it a Beta AI Function. SQL and POST /api/2.0/ai-functions/ai-decide take a state and questions of type noul, choice, or score. The blog says the function is directly compatible with the TypeSafe API. The docs do not name the model, and they do not print a price or a millisecond.
Databricks posted on October 1, 2026, at 13:54 UTC, that ai_decide classifies, scores, and chooses the next step. The post says the call returns in a fraction of a second, at lower latency and cost than an LLM on similar tasks. The attachment is a video. The catalog of calls that return a typed decision is decision models like Jev.
The blog’s schema.org block says datePublished 2026-09-30T16:59:34+0000. The byline is Jane Zhang, Matthew Ding, and Archika Dogra. The page says a decision model such as TypeSafe’s Jev returns probabilities instead of generated text. It then says ai_decide is powered by a decision model, and that a team already calling the TypeSafe API can use this function. It does not name the checkpoint behind the function.
Each question on a shared state comes back as a probability, a choice from named criteria, or a score on an ordered scale. The page says the function is in Beta, from SQL or from REST. A Snake demo sends the board on every tick and asks which way to turn. That section prints no win rate and no millisecond. A link on the same page points at serving open-weight decision models in SQL. We did not open that walkthrough.
The SQL page, last updated September 29, 2026, gives ai_decide(state, questions [, options]). questions is a constant JSON object. Version 1.0 is the default. A choice maps 1 to 255 labels. A score lists 2 to 10 levels, and the number that comes back is the probability-weighted average of those indices. A noul is a probability from 0 to 1, and the example has no confidence field. Choice and score examples include confidence. The page marks the function Beta, unavailable on Databricks SQL Classic, and it requires Databricks Runtime 15.4 LTS or newer, with 18.2 recommended. It says the models that might run are under Apache 2.0, and that Databricks might swap the model later. The pricing link does not print a dollar figure on the page we read. Parameters in the call are not stored. Metadata, including the runtime version, is retained.
REST is POST /api/2.0/ai-functions/ai-decide. The body is state, questions, and an optional options.version. Answers sit under response, and the version sits under metadata. On that reference, ai_classify prints 1,200 requests per minute per workspace, and ai_extract prints 120. The ai_decide section does not print a rate. We did not send a request. Jev’s own call is a different path, POST /v1/systemone.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
@databricks, October 1, 2026, 13:54 UTC, status 2105657705390027130. The text says ai_decide classifies, scores, and chooses what happens next, in a fraction of a second, with lower latency and cost than an LLM on similar tasks. The attachment is a video. The card links the blog. We did not call the function.
The blog is https://www.databricks.com/blog/introducing-aidecide-make-fast-decisions-your-governed-data. Its schema.org block says datePublished 2026-09-30T16:59:34+0000. The byline is Jane Zhang, Matthew Ding, and Archika Dogra. The visible date line is September 30, 2026. We opened the page on October 4, 2026.
The blog names TypeSafe's Jev as a decision model that returns probabilities instead of text. The next paragraph says ai_decide is powered by a decision model, and that teams already on the TypeSafe AI API can call this function directly. It does not name the checkpoint. It lists a probability, a choice from named criteria, or a score on an ordered scale. It says the function is in Beta, from SQL or REST.
The same page describes a Snake demo on Databricks Apps. Each tick sends the board and asks for the next direction. The page prints no win rate and no millisecond for that loop. It links a walkthrough for serving open-weight decision models in SQL. We did not open that walkthrough, and we did not run the demo.
The SQL page, last updated September 29, 2026, documents ai_decide(state, questions [, options]). questions is a constant JSON object. version 1.0 is the default. A choice criteria object maps 1 to 255 labels. A score criteria list is 2 to 10 levels, and the score is the probability-weighted average of those indices. A noul returns a probability from 0 to 1. The example noul has no confidence field. Choice and score examples include confidence. The page says Beta, not on Databricks SQL Classic, Databricks Runtime 15.4 LTS or above, with 18.2 recommended. It says the models that might be used are Apache 2.0, and that Databricks might change the model. It points at the SQL pricing page and does not print a dollar amount. It says parameters passed into the call are not stored, and that metadata such as the runtime version is retained.
The REST reference is POST /api/2.0/ai-functions/ai-decide. The body fields are state, questions, and an optional options.version. The response puts answers under response and the version under metadata. On that same reference, ai_classify prints 1,200 requests per minute per workspace, and ai_extract prints 120. The ai_decide section we read does not print a rate. We did not send a request.
Compare
TypeSafe's models page, as recorded on Access on October 1, lists jev-1.13.0 at $0.042 per million input tokens and free output. The ai_decide pages we opened do not print a price.
Jev's Noul is a probability with no separate confidence. Choice and Score include confidence. The Databricks examples match that split. Choice on the Databricks page allows 1 to 255 labels. Clef's input schema, on the Clef page, describes 2 to 255 options. TypeSafe's primitives page says Choice picks among up to 255 named options.
The blog's phrase "a fraction of a second" is not a latency table. Cloudflare's median row, on the Clef page, is 209.3 ms, 38.8 ms, and a printed Jev cell of 524.1 ms. Those clocks are different hosts.
OpenAI's Decisions API is a separate product. On October 2 a POST to https://api.openai.com/v1/decisions from this desk returned HTTP 403.
The blog says the function is directly compatible with the TypeSafe API. The documented REST path is /api/2.0/ai-functions/ai-decide. The TypeSafe path on this desk is POST /v1/systemone. We did not send one body to both.
Terms
- ai_decide
- Databricks AI Function in Beta on the September 30, 2026 blog. SQL is ai_decide(state, questions). REST is POST /api/2.0/ai-functions/ai-decide. A question is a noul, a choice, or a score.
- 1.0
- The ai_decide function version in the SQL page and the REST reference. It is the default. The options map can set it with version.