Databricks Releases ai_decide, a Native Structured-Decision AI Function
A category of AI model purpose-built for structured decisions, rather than prose generation, is gaining tooling support from major infrastructure providers. Databricks releasing a native ai_decide function signals that structured-decision inference is being treated as a distinct workload, not just a use pattern atop existing LLMs.
Databricks released ai_decide, an AI Function built for fast, structured decisions over governed data, offering lower latency and cost than LLMs on classification, scoring, and routing tasks. Stated use cases include model routing, document processing, agent evaluations, and real-time app logic. This follows Databricks' earlier demonstration of running open-weight Jev-style models via ai_query, and sits within a broader pattern established by TypeSafe AI's Jev model. Jev accepts unstructured text or semi-structured data and returns numeric confidence scores for classification-style questions rather than generating prose. TypeSafe describes the model as answering in 70 to 500 milliseconds at roughly $42 per billion tokens of equivalent workload, and frames the approach as suited for decisions where every valid answer could be written on a sticky note. JevBench, TypeSafe's composite benchmark, scores models on Intelligence, Calibration, Speed, and Cost using a geometric mean so that strength on one axis cannot mask weakness on others. Jev 1.13.0 leads the JevBench composite despite GPT-5.6 Luna having substantially higher hard-case accuracy, because Jev outperforms on latency, calibration, and cost. TypeSafe states the result is 'evidence about a narrow typed-decision workload, not evidence that Jev is generally more capable than GPT-5.6 Luna.' The open-weight SemIf-OpenJev model can be served on Databricks serverless GPUs and called from SQL or Lakeflow jobs via ai_query, with the workflow packaged as a single importable notebook. Together AI released together/Tev1-4B-experimental, a similar classifier built on Qwen3.5 4B, and published a guide showing fine-tuning costs approximately $17.
On October 1, Databricks released ai_decide, a native AI Function for fast structured decisions over governed data, covering classification, scoring, and routing tasks.
The release moves Databricks from demonstrating third-party open-weight model integration via ai_query to shipping its own first-party structured-decision primitive, with stated use cases including model routing, document processing, agent evaluations, and real-time app logic.
Coverage on September 29 added framing that positions Jev, a model that scores predefined choices and returns confidence values rather than generating prose, as suited for decisions with a fixed, enumerable answer set.
The recommended pattern, per the coverage, is to escalate only uncertain or open-ended cases to frontier generative models.
On September 25, Databricks demonstrated serving SemIf-OpenJev, an open-weight Jev variant, on governed enterprise data using serverless GPUs, AI Runtime, and Model Serving, with results callable from SQL or Lakeflow jobs via the ai_query function, packaged as a single importable notebook.
Jev, released September 15 by TypeSafe AI, accepts unstructured text and returns typed probability scores for classification questions rather than generating prose.
Together AI released together/Tev1-4B-experimental on September 23, a classifier built on Qwen3.5 4B that works like TypeSafe AI's Jev, a decision model that returns typed probability scores rather than prose.
Together hosted it on their serverless platform and published a guide showing a similar model can be fine-tuned for approximately $17.
TypeSafe AI launched Jev on September 15, a model that takes unstructured text and returns numeric confidence scores for decisions instead of prose.
On September 19, TypeSafe released JevBench, a composite benchmark of intelligence, calibration, speed, and cost where Jev 1.13.0 leads GPT-5.6 Luna despite its lower hard-case accuracy. Simon Willison flagged bias risk on September 21, citing a ranking experiment where Jev scored Cupertino top and East Palo Alto bottom with no inspectable rationale.
The Neuron Daily covers Jev and TypeSafe founder Diogo Almeida's background
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free