Back to Blog
    AI & AutomationNov 21, 2026

    Snowflake Decision: AI That Classifies, Scores and Routes in SQL

    Amir

    Amir Farid

    7 min read

    Snowflake Decision: AI That Classifies, Scores and Routes in SQL

    Snowflake Decision is now in public preview inside Cortex AI_COMPLETE, Snowflake announced in its release notes on 20 November 2026. It is a purpose-built model for classification, scoring and decision-making across large datasets — not a general chat model. You call it from SQL, it reads a row, judges it against criteria you define, and returns a structured answer with a probability and a confidence value. Your data never leaves the warehouse, and your software can act on the answer or send uncertain cases to a person.

    What Snowflake Decision actually does

    Most AI models answer in free text. Snowflake Decision answers in bounded, structured form. It judges unstructured data — a support ticket, an email, an invoice note — against criteria you write, and returns an answer your software can use directly. Three question types:

    • Choice. Pick from up to 32 options, with a probability for each. Which department should this support ticket go to — billing, technical, delivery, or returns?
    • Score. Rate something against a rubric or scale, with probabilities across the levels. How frustrated is this customer, on a scale you define?
    • True/false probability. A likelihood from 0 to 1 — Snowflake calls this "noul". Is this message urgent? The model returns a number, not a yes or no, so you set your own threshold.

    One AI_COMPLETE request can bundle several judgements at once. Snowflake's example: choose a ticket's department, score the customer's frustration, and assess urgency — in a single call. One call replaces three.

    How it works inside SQL

    A support-ticket triage dashboard with priority lanes and one urgent ticket highlighted

    Snowflake Decision is called through Cortex AI_COMPLETE, the SQL function for running models over tables. You write a query, point it at a table, and every row gets judged. No separate application to build, no data pipeline, no export step. REST access is also available through Cortex AI Gateway.

    Your data stays in the warehouse

    The model runs inside Snowflake, so the data does not leave the warehouse. The work falls under Snowflake's existing governance and audit controls. For regulated businesses, that is the point: a ticket classified by an outside API is a data-governance conversation; the same ticket classified inside your own warehouse is a query.

    From private preview to public preview

    Snowflake Decision was in private preview since 7 October 2026. The 20 November release note moves it to public preview. A lower-latency variant, snowflake-decision-fast, arrived in private preview on 19 October 2026.

    What it costs — and the token trap

    Snowflake quotes two figures: $0.042 per million tokens, and about 1.58 milliseconds per row. Both are Snowflake-quoted figures, not independent measurements.

    The billing nuance matters before you run it over a million rows. AI Functions bill per million tokens, input plus output. Your criteria and output schema are sent as JSON with every row, so verbose criteria cost tokens on every record. A 500-token rubric over 100,000 rows is 50 million tokens before the model has read a word of data.

    • Keep criteria tight. Short, precise rubrics cost less on every row.
    • Estimate first. Use AI_COUNT_TOKENS to measure your criteria's token cost before running a full table.
    • Bundle judgements. One call answering three questions beats three calls — criteria and row text are sent once.

    What Snowflake says about accuracy — and what it doesn't

    Here is the honest part: Snowflake published no accuracy benchmark in the 20 November release note. The announcement describes what the model does; it does not say how well it does it.

    The numbers that do exist come from the October private-preview coverage, and they are Snowflake's own figures. On the Jev Decision Index 0.2.1, Snowflake Decision scored 57.63 across 29 benchmarks and 88,837 test runs, the top score among 72 compared models. One caveat: the model supports up to 32 options per choice, so only a subset of those benchmarks could run end-to-end. A top score is encouraging, not a guarantee for your data. Run it on a sample of your own data, score the answers by hand, and decide from that.

    Where it beats a general LLM — and where it doesn't

    A general model can classify too — the question is whether it is the right tool at scale. Snowflake Decision wins on four things:

    • Cost. At Snowflake's quoted $0.042 per million tokens, a million short rows costs a fraction of a general model's price.
    • Speed. Around 1.58 milliseconds per row, quoted by Snowflake, is built for bulk work — a whole table in one query.
    • Structure. Probabilities and confidence values come back as data, not sentences to parse. Your software can branch on them reliably.
    • Auditability. Criteria, input and output all live in the warehouse, under your governance controls.

    Where it does not win is open-ended work. Drafting a reply, writing a summary, reasoning through an unusual case, holding a conversation — that still needs a general model. The sensible split: a decision model sorts and scores, a general model writes and talks. Our guide to AI chatbots versus AI agents explains the same division of labour, and our AI agent development service builds systems that combine both.

    Practical uses, with guardrails

    For most businesses, the value is automation inside the data warehouse. Ticket triage, lead scoring, invoice routing and complaint prioritisation become a SQL query:

    • Support triage. Every ticket gets a department, a frustration score and an urgency probability in one pass. Urgent, frustrated customers jump the queue. Our chatbot development team and clinics using AI chatbots already follow this pattern.
    • Lead scoring. Score inbound enquiries against your own rubric and route the hot ones to sales first. See our AI agents for real estate lead follow-up.
    • Invoice routing. Classify incoming invoices and send them to the right approver without a human reading each one.
    • Complaint prioritisation. Score complaints by severity and route the serious ones to a senior team member the same day.

    Guardrails are not optional. Set a confidence threshold and send low-confidence rows to a human. Review a sample of decisions every week at first. Our AI automation service and custom software development team can wire decision models into your existing systems, and our CRM development service can put the scores where your team already works.

    Decision models are a new category

    Snowflake is not alone. Google launched Gemini Decision 1, which we covered on 13 November, and Baseten announced bison-1 and groundhog-1 on 20 November. Sarvam's Prism, covered here, does the same job in a different market. The pattern is the same: bounded answers instead of free text, probabilities instead of paragraphs. If your business runs on decisions at volume — tickets, leads, invoices, claims — test this category now while it is cheap. Our guide to AI agents for small business automation, AI automation for logistics companies, AI agents for ecommerce and n8n versus Make versus Zapier comparison show where this kind of automation pays off first, and our AI agent development cost guide for the UK covers build costs.

    Frequently asked questions

    What is Snowflake Decision?

    A purpose-built model inside Snowflake's Cortex AI, now in public preview. It classifies, scores and makes decisions over large datasets from SQL, returning structured answers with probabilities and confidence values — not free text.

    How much does Snowflake Decision cost?

    Snowflake quotes $0.042 per million tokens and about 1.58 milliseconds per row. AI Functions bill per million tokens, input plus output, and criteria JSON is sent with every row, so keep criteria tight.

    Do I need to move my data out of Snowflake to use it?

    No. It runs inside the warehouse through the AI_COMPLETE SQL function, so data stays under Snowflake's existing governance and audit controls.

    How accurate is Snowflake Decision?

    Snowflake published no accuracy benchmark in the 20 November release note. Its own figures from the October private preview put it top among 72 compared models on the Jev Decision Index, scoring 57.63 — but the 32-option ceiling limited which benchmarks could run end-to-end. Test it on your own data first.

    When should I use a decision model instead of a general LLM?

    Use a decision model for high-volume, bounded, auditable work: sorting, scoring and routing. Use a general model for open-ended work: drafting, conversation and judgement calls.

    Want to test whether decision AI fits your workflow? Book a free AI consultation with Coheser. The full release-note entry is on Snowflake's release notes page.

    #Snowflake Decision#decision AI#Cortex AI#workflow automation