{"slug": "snowflake-decision-a-decision-model-in-cortex-ai-functions", "title": "Snowflake Decision: A Decision Model in Cortex AI Functions", "summary": "Snowflake introduced Snowflake Decision in private preview, a decision model available through the AI_COMPLETE function in Snowflake Cortex AI Functions for high-volume classification, scoring, filtering and routing tasks. Snowflake said the model scored 57.63 on 29 Jev Decision Index 0.2.1 benchmarks spanning 88,837 test requests, the highest quality score among 72 models evaluated on the same benchmarks. The model returns typed answers such as a category, score or yes/no probability, and runs under existing Snowflake governance, access controls and audit policies.", "body_md": "Many enterprise AI workloads are decision problems: determining which category a record belongs to, how risky a transaction is and whether a ticket should be processed further. Many of these tasks have bounded answers, and they run at high volume across large tables.\n\nGeneral-purpose LLMs can handle these tasks, but they may carry higher cost and latency than the work requires. Decision models are purpose-built for simple, high-volume tasks that do not require complex reasoning, including classification, scoring, filtering and routing. They return a typed answer, such as a category, a score or a yes/no probability, that your workflow can act on directly.\n\nCustomers already use [Cortex AI Functions](https://docs.snowflake.com/en/user-guide/snowflake-cortex/aisql) to classify, score and route data at scale. Now, we’re introducing Snowflake Decision (private preview), a decision model available through [AI_COMPLETE,](https://docs.snowflake.com/en/sql-reference/functions/ai_complete) to help teams scale simple, high-volume tasks even more efficiently. The call runs in Snowflake Cortex AI Functions, so the governance, access controls and audit policies you already have in place apply to every decision.\n\n## A model designed for efficient, accurate structured decisions\n\nConsider a support workflow processing thousands to millions of tickets. Before each ticket reaches an agent, it needs to be assigned to the right team. The same message may also need to be scored for customer frustration and checked for expressed urgency. Rather than handle these as separate model calls, a decision model can evaluate all three questions together, returning structured answers your workflow can use directly.\n\nWhen choosing a model for these workflows, decision quality matters alongside speed, cost and response format. This is why we are excited to put Snowflake Decision in the hands of customers. In our evaluation, Snowflake Decision achieved the highest quality score among 72 models compared on [Jev Decision Index 0.2.1](https://huggingface.co/spaces/multimodalart/jev-decision-index) benchmarks. For customers, higher-quality decisions can mean more reliable classification, scoring and routing across high-volume workflows.\n\nNote: We evaluated Snowflake Decision on 29 [Jev Decision Index 0.2.1 benchmarks](https://huggingface.co/spaces/multimodalart/jev-decision-index) that run end to end within our 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; current interface limits, such as a maximum of 32 options, prevent us from running the full benchmark. Snowflake Decision scored 57.63, the highest quality score among 72 models evaluated on the same 29 benchmarks. To ensure a like-for-like comparison, we calculated every model’s score using the same benchmark results and the index’s published, chance-adjusted methodology.\n\n## Triage support tickets with Cortex AI Functions\n\nCortex AI Functions bring model calls into SQL. You call the decision model through AI_COMPLETE by passing your request as JSON in the prompt argument. You provide the ticket text and clearly defined questions, using three supported question types:\n\n- **Choice:** Select an option from a defined list, such as billing, technical support or sales.\n- **Score:** Rate the input against an ordered rubric, such as calm, frustrated but civil, or very angry.\n- **Yes or no (noul):** Return a probability from 0 to 1 for a statement such as “The message conveys urgency.”\n\nIn this example, each row in the **support_tickets** table contains a customer message in the **ticket_text** column. A single call to Snowflake Decision routes the ticket, scores frustration and assesses urgency:\n\n``` js\nSELECT AI_COMPLETE(\n    model => 'snowflake-decision',\n    prompt => TO_JSON({\n        'state': ticket_text,\n        'questions': {\n            'department': {\n                'type': 'choice',\n                'instructions': 'Which team should handle this',\n                'criteria': {\n                    'billing': 'Payment or subscription issues',\n                    'technical': 'Bugs or integration problems',\n                    'sales': 'Pricing or account questions'\n                }\n            },\n            'frustration': {\n                'type': 'score',\n                'instructions': 'How frustrated the customer appears',\n                'criteria': ['Calm', 'Frustrated but civil', 'Very angry']\n            },\n            'is_urgent': { 'type': 'noul', 'instructions': 'The message conveys urgency' }\n        }\n    })\n) FROM support_tickets;\n```\n\nThe result comes back as an answers map, keyed by your question names, with the chosen value, per-option probabilities and, for choice and score questions, a confidence value.\n\n```\nResponse \n\n{\n  \"answers\": {\n    \"department\": {\n      \"choice\": \"billing\",\n      \"confidence\": 0.9305973965815073,\n      \"probabilities\": {\n        \"billing\": 0.9537315977210048,\n        \"sales\": 0.017468203540633418,\n        \"technical\": 0.02880019873836161\n      },\n      \"type\": \"choice\"\n    },\n    \"frustration\": {\n      \"confidence\": 0.7082686180607838,\n      \"legend\": {\n        \"0\": \"Calm\",\n        \"1\": \"Frustrated but civil\",\n        \"2\": \"Very angry\"\n      },\n      \"probabilities\": {\n        \"0\": 0.014753474459327466,\n        \"1\": 0.1797341135001498,\n        \"2\": 0.8055124120405226\n      },\n      \"score\": 1.790758937581195,\n      \"type\": \"score\"\n    },\n    \"is_urgent\": {\n      \"noul\": 0.8807970779778823,\n      \"type\": \"noul\"\n    }\n  }\n}\n```\n\n## What you can do with a decision model on Snowflake\n\nCortex AI Functions today offer teams a range of model options for different tasks. Our decision model expands that selection by adding another option alongside the general-purpose models already available. It is fit for high-volume, bounded decisions run across many rows: classification, routing, scoring, filtering and triage.\n\nWhen to reach for a decision model:\n\n- The answer is one of a known set of options, a score on a rubric or a yes/no\n- You are running the decision across thousands of rows\n- Speed and cost matter\n- You want a probability you can threshold on\n- You want to use a decision model for routing and categorization as the first step of a multistep LLM pipeline\n\nWhen a general-purpose model is the better fit:\n\n- You need summaries, open-ended reasoning or free-form extraction\n- The answer space is unbounded\n- The task requires multistep reasoning or creative generation\n\nTeams are looking at decision models for work such as:\n\n- **Support operations:** A support team needs every incoming ticket assigned to a queue before an agent picks it up. One call returns the queue, a frustration score and an urgency flag for each ticket in the table.\n- **Customer reviews analytics:** A product team wants to know which issues come up most often across a large review corpus. One call can return a review category and a severity score, so the team can easily aggregate sentiment and prioritize features.\n- **Answer quality checks:** A team running an AI assistant wants to know whether its answers follow a stated policy. The decision model judges each answer against that policy and scores how completely it addresses the customer's request, so reviewers only read the ones that fall short.\n- **AI data pipeline building:** Some pipelines send every request to a large model today. The decision model judges how much reasoning each request needs, and customers can then build a pipeline that sends the hard cases to the larger model and keeps the rest on a more cost-effective path.\n\n## Getting started\n\nTo request access, reach out to your Snowflake account team. A Snowflake team member will follow up with next steps, including private preview documentation.", "url": "https://wpnews.pro/news/snowflake-decision-a-decision-model-in-cortex-ai-functions", "canonical_source": "https://www.snowflake.com/content/snowflake-site/global/en/blog/snowflake-decision-cortex-ai-functions", "published_at": "2026-10-08 19:08:27+00:00", "updated_at": "2026-10-08 19:50:00.995776+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products", "ai-tools", "large-language-models", "ai-infrastructure"], "entities": ["Snowflake", "Snowflake Decision", "Cortex AI Functions", "AI_COMPLETE", "Jev Decision Index 0.2.1", "Hugging Face"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/snowflake-decision-a-decision-model-in-cortex-ai-functions", "markdown": "https://wpnews.pro/news/snowflake-decision-a-decision-model-in-cortex-ai-functions.md", "text": "https://wpnews.pro/news/snowflake-decision-a-decision-model-in-cortex-ai-functions.txt", "jsonld": "https://wpnews.pro/news/snowflake-decision-a-decision-model-in-cortex-ai-functions.jsonld"}}