Snowflake introduced Snowflake Decision in private preview on October 8, 2026, as a model for classification, scoring, filtering and routing across high-volume enterprise data. Available through AI_COMPLETE in Snowflake Cortex AI Functions, it returns typed answers such as categories, scores and yes-or-no probabilities. Snowflake says existing governance, access controls and audit policies apply to each decision.

The model is designed for bounded tasks with known answer formats, where general-purpose large language models may cost more or respond more slowly than needed. Snowflake gives support-ticket triage as an example: one call can assign a ticket to a team, score customer frustration and assess urgency, returning structured answers for a workflow to use.

In Snowflake’s evaluation, the model scored 57.63 and had the highest quality score among 72 models compared using 29 benchmarks from Jev Decision Index 0.2.1. Those benchmarks covered reasoning, language, retrieval, tool use and creative judgment, with 88,837 test requests in total. They were a subset of the broader Hugging Face benchmark because the current interface has limits, including a maximum of 32 options. Snowflake said it calculated all model scores from the same benchmark results using the index’s published, chance-adjusted method.

Through AI_COMPLETE, users provide a request as JSON and can ask three question types: choice from a defined list, a score against an ordered rubric, or yes/no, which returns a probability from 0 to 1. The response is an answers map keyed by question names; it includes the selected value and per-option probabilities, plus a confidence value for choice and score questions.

Snowflake recommends decision models when answers come from known options, a scoring rubric or a yes/no assessment, especially across thousands of rows when speed, cost or thresholdable probabilities matter. They can also make an initial routing or categorization step in a multistep LLM pipeline. General-purpose models are a better fit for summaries, open-ended reasoning, free-form extraction, unbounded answers, multistep reasoning or creative generation.

Suggested uses include routing tickets and assessing frustration and urgency; categorizing customer reviews and scoring severity; checking whether AI assistant answers follow a policy and meet a customer’s request; and directing requests that need more reasoning to a larger model while handling simpler ones more cost-effectively. Customers can request preview access through their Snowflake account team, which will provide next steps and private-preview documentation.

Snowflake introduced Snowflake Decision in private preview on October 8, 2026, as a model for classification, scoring, filtering and routing across high-volume enterprise data. Available through AI_COMPLETE in Snowflake Cortex AI Functions, it returns typed answers such as categories, scores and yes-or-no probabilities. Snowflake says existing governance, access controls and audit policies apply to each decision.

The model is designed for bounded tasks with known answer formats, where general-purpose large language models may cost more or respond more slowly than needed. Snowflake gives support-ticket triage as an example: one call can assign a ticket to a team, score customer frustration and assess urgency, returning structured answers for a workflow to use.

In Snowflake’s evaluation, the model scored 57.63 and had the highest quality score among 72 models compared using 29 benchmarks from Jev Decision Index 0.2.1. Those benchmarks covered reasoning, language, retrieval, tool use and creative judgment, with 88,837 test requests in total. They were a subset of the broader Hugging Face benchmark because the current interface has limits, including a maximum of 32 options. Snowflake said it calculated all model scores from the same benchmark results using the index’s published, chance-adjusted method.

Through AI_COMPLETE, users provide a request as JSON and can ask three question types: choice from a defined list, a score against an ordered rubric, or yes/no, which returns a probability from 0 to 1. The response is an answers map keyed by question names; it includes the selected value and per-option probabilities, plus a confidence value for choice and score questions.

Snowflake recommends decision models when answers come from known options, a scoring rubric or a yes/no assessment, especially across thousands of rows when speed, cost or thresholdable probabilities matter. They can also make an initial routing or categorization step in a multistep LLM pipeline. General-purpose models are a better fit for summaries, open-ended reasoning, free-form extraction, unbounded answers, multistep reasoning or creative generation.

Suggested uses include routing tickets and assessing frustration and urgency; categorizing customer reviews and scoring severity; checking whether AI assistant answers follow a policy and meet a customer’s request; and directing requests that need more reasoning to a larger model while handling simpler ones more cost-effectively. Customers can request preview access through their Snowflake account team, which will provide next steps and private-preview documentation.