Snowflake opens public preview of AI model for classifying and scoring data
Snowflake Decision returns structured answers and confidence values for automated workflows. The company quotes low processing costs, but supplies no accuracy benchmark in the release note.
Teams can now use Snowflake Decision in public-preview workflows, but the change is availability rather than a new decision method: its private-preview documentation already covered category selection, rubric scoring, and true-or-false assessments. Through Cortex AI_COMPLETE, a request can bundle distinct judgments and return probabilities or confidence values that applications can use to route uncertain cases for review. Snowflake quotes $0.042 per million tokens and about 1.58 milliseconds per row; it has not published accuracy benchmarks or workload details to validate those performance claims.
01
Choice questions return the selected option, probabilities for every option, and confidence; rubric scores return probabilities for each scoring level.
02
Snowflake’s support-ticket example combines department selection, frustration scoring, and urgency assessment in one AI_COMPLETE request.
03
Users can apply Decision to Snowflake tables through SQL for tasks including document classification, feedback analysis, and ticket prioritization.
Snowflake put its Decision model into public preview on November 20, 2026, letting customers classify and score data through Cortex AI_COMPLETE. The model evaluates information against predefined criteria and returns structured answers with probabilities and confidence values, giving workflows a way to act on results or send uncertain cases for review.
The release moves Snowflake Decision beyond the private preview documented on October 7. That earlier version already supported classification, scoring against a supplied scale and true-or-false assessments. November’s announcement changes the availability stage; the basic approach of asking defined questions and receiving structured results was already part of the model.
Define the decision before running the model
Snowflake describes Decision as purpose-built for high-volume classification, scoring and decision-making across large datasets. Users supply the criteria against which information should be evaluated. The documented question types give applications several ways to turn that evaluation into an answer, depending on whether the task needs a category, a score or a probability.
Choose a category: A choice question supplies a fixed set of options. The model returns the selected option’s key, a probability for every option and a confidence value. That gives downstream software both a selection and the distribution across alternatives.
Apply a scoring scale: A score question evaluates content against an ordered rubric—a defined set of scoring criteria. It returns a probability-weighted score, the probability assigned to each level and confidence, rather than only a single selected level.
Assess a statement: A true-or-false question returns the probability that a specified statement is true. This supports checks framed around a condition, rather than a choice among several named categories.
Multiple independent assessments can share one AI_COMPLETE request. Snowflake’s support-ticket example combines choosing the responsible department, scoring customer frustration and assessing urgency in one call. Those are separate judgments about the same content; grouping them reduces the need to send a separate model request for each criterion.
Confidence becomes a routing input
Applications can use the returned probabilities and confidence values to choose between automated actions, human review and fallback processing. Snowflake presents those values as inputs to workflow logic: an uncertain result can take a different route from one accepted for automation. The model’s answer and the application’s next action are therefore separate parts of the process.
The scale comes through SQL, the language used to query tables. Users can apply classification and scoring across Snowflake tables, and the earlier documentation describes evaluating a table or query result in one SQL statement. Listed uses include document classification, customer-feedback analysis, support-ticket prioritization and routing work to agents.
The cost and speed figures are Snowflake’s claims. Its public-preview release note does not include an accuracy benchmark or describe the workload behind the per-row timing. Those figures describe the company’s processing pitch, but do not establish how reliably the model will apply a customer’s particular classification rules or scoring criteria.
Decision can also be combined with Cortex functions including AI_TRANSCRIBE, AI_PARSE_DOCUMENT, AI_EXTRACT and AI_AGG. That puts classification and scoring within the same Snowflake workflow as transcription, document parsing, extraction and aggregation, rather than leaving the structured decision as an isolated output.
Sources
docs.snowflake.comOct 7, 2026: Snowflake Decision with AI_COMPLETE (Private Preview)
docs.snowflake.comNov 20, 2026: AI_COMPLETE adds support for Snowflake Decision (*Preview*) - Snowflake Documentation
Reader comments
Newest comments first. Replies stay oldest first.