Snowflake Launches Decision Model for High-Volume Tasks in Cortex AI Functions

A model for bounded decisions at scale

According to Snowflake, many enterprise AI workloads are decision problems: which category a record belongs to, how risky a transaction is, or whether a ticket should be processed further. These tasks often have bounded answers and run across large tables. General-purpose LLMs can handle them, the company says, but may carry higher cost and latency than the work requires.

Decision models are described as purpose-built for simple, high-volume tasks that do not require complex reasoning. They return a typed answer, such as a category, a score or a yes/no probability. Because the call runs in Snowflake Cortex AI Functions, the governance, access controls and audit policies a customer already has in place apply to every decision, according to the company.

Benchmark claim and its limits

Snowflake says Snowflake Decision achieved the highest quality score among 72 models compared on Jev Decision Index 0.2.1 benchmarks. It scored 57.63. The evaluation covered 29 benchmarks that run end to end within the current interface, spanning 88,837 test requests across reasoning, language, retrieval, tool use and creative judgment.

These 29 benchmarks are a subset of the broader Hugging Face benchmark. Snowflake notes that current interface limits, such as a maximum of 32 options, prevent it from running the full benchmark. For a like-for-like comparison, the company says it calculated every model's score from the same benchmark results using the index's published, chance-adjusted methodology.

How the support ticket example works

Snowflake illustrates the model with a support workflow processing thousands to millions of tickets. Instead of separate model calls, one call can route a ticket to a team, score customer frustration and check for urgency. The request is passed as JSON in the prompt argument of AI_COMPLETE, with the model set to 'snowflake-decision'.

Three question types are supported:

• Choice: select an option from a defined list, such as billing, technical support or sales.
• Score: rate the input against an ordered rubric, such as calm, frustrated but civil, or very angry.
• Yes or no (noul): return a probability from 0 to 1 for a statement such as "The message conveys urgency."

The result comes back as an answers map keyed by the question names. It contains the chosen value and per-option probabilities, plus a confidence value for choice and score questions.

When to use it, and when not to

Snowflake recommends a decision model when the answer is one of a known set of options, a score on a rubric or a yes/no, when the decision runs across thousands of rows, when speed and cost matter, and when a probability to threshold on is useful. It can also serve as the first routing or categorization step in a multistep LLM pipeline.

A general-purpose model remains the better fit for summaries, open-ended reasoning, free-form extraction, unbounded answer spaces, multistep reasoning or creative generation.

The company lists several use cases: support operations, customer review analytics, answer quality checks against a stated policy, and AI data pipeline building. In the last case, the decision model judges how much reasoning each request needs, so hard cases go to a larger model and the rest stay on a more cost-effective path.

Availability

Snowflake Decision is in private preview. Customers who want access should contact their Snowflake account team, which will follow up with next steps, including private preview documentation. The source gives no pricing or general availability date.