Snowflake adds low-latency Decision Fast model for real-time AI choices

What Snowflake is releasing

Two weeks before this announcement, Snowflake introduced Snowflake Decision in Cortex AI Functions, also in private preview. The model is designed for classification, scoring, filtering and routing, and handles high-volume tasks with bounded answers. Instead of open-ended text, it returns a typed answer, such as a specific category, a numerical score or a yes/no probability, that a workflow can act on directly.

Snowflake says customers asked for much higher speed for decisions inside interactive or latency-sensitive workflows. The company describes this as "arcade speed". The new release, again in private preview, includes three capabilities, all covered by the governance, access controls and audit policies customers already use:

• snowflake-decision-fast: a lower-latency variant of snowflake-decision for workloads where response time matters as much as decision quality.
• REST API support: Snowflake Decision models can be called through Cortex AI Gateway, so applications, services and agents outside of SQL can use structured decision-making.
• Cortex AI Functions: snowflake-decision-fast is natively available through AI_COMPLETE in SQL.

A live ticket-classification demo

To show the model in an interactive setting, Snowflake built a real-time text classification application. As a customer types a support ticket, for example a complaint about being charged twice, the app evaluates the live text and makes three decisions at once: which team should receive the ticket, whether it is urgent, and how the customer feels. On every keystroke, it evaluates the current text and asks the model bounded questions to determine the next workflow step.

Snowflake highlights three mechanics of its decision models:

• Construction bounds: the app offers only valid options, so the model cannot return an invalid action. According to Snowflake, this removes the need for parsers, retry loops or complex error handling.
• Probabilities as the product: the model returns the chosen value plus per-option probabilities rather than free-text reasoning. Snowflake gives the example of a 0.99 for a "Billing" route as a signal that can be thresholded in live workflows.
• Visible latency: the company says endpoint round trips consistently complete in under 300 ms, which keeps the experience seamless as the user types.

The blog post includes a Python example that calls the model through Cortex AI Gateway with three questions: a yes/no check on urgency, a choice among support teams, and a score for customer frustration on a scale from calm to very frustrated.

Applying the same logic to whole tables

The same capabilities are available through Cortex AI Functions, letting developers apply classification and routing logic to large data sets with SQL. Snowflake points to a global support platform that must triage and route millions of tickets or messages every day. In its view, using a general-purpose LLM for such a narrow classification task can add unnecessary complexity, cost and processing overhead.

With snowflake-decision-fast in AI_COMPLETE, teams can apply the same decision logic across entire tables without building separate inference pipelines or managing external API integrations, according to the company. Its sample query assigns each ticket a routing queue and a yes/no escalation flag, returning structured output ready for downstream processing.

Snowflake also says it is iterating quickly on both quality and speed. Its stated focus is to push decision accuracy on established benchmarks while reducing latency and compute costs, and it expects ongoing updates to model quality in the coming months. A chart in the post compares average accuracy and row processing time for snowflake-decision and snowflake-decision-fast, based on Snowflake's internal decision benchmark. The text does not give the underlying figures.

Availability and what it means

Both the fast model and the REST access are in private preview. Customers who want early access should contact their Snowflake account team, and a representative will follow up with next steps, including private preview documentation.

For teams building agents, routing steps or per-record scoring, the pitch is that decisions can sit inside the flow of work without a large latency or cost penalty. Snowflake frames the shift as a change from asking "Can we afford to put AI here?" to "Where else could a decision go?" The claims about speed and accuracy come from the company itself, and independent results are not part of the announcement.