Less Reasoning. More Doing. Meet Koa Decisions.

Salesforce’s new decision model, Koa Decisions, processes single-step decisions in milliseconds, contrasting with larger AI models that take over a second. It builds on the company’s existing HyperClassifier, which routes customer inquiries to specialized agent components, such as shipping or FAQ handling. The model reads a request and predefined categories, then selects the appropriate option without requiring retraining or custom machine learning pipelines. This approach reduces latency and computational overhead, enabling immediate action in workflows.

Koa Decisions extends beyond routing to support flows, Apex, and prompt templates for Salesforce developers, alongside an API for external integration. It joins Salesforce’s Koa family of models, which includes a separate reasoning model built with NVIDIA. The model is designed for quick, targeted decisions, such as assigning a customer query to the correct team or selecting a knowledge article, rather than generating text. Its training leverages Google’s Gemma open-weight model with novel techniques and adds multimodal support for images and audio.

In tests within Salesforce’s Agentforce, Koa Decisions routed requests in approximately 26 milliseconds, compared to 1.4 seconds for a general-purpose LLM. The model returns a single token answer—a clear choice and confidence score—rather than generating text token by token. This efficiency allows other software to act on the decision immediately, streamlining processes like escalating a sales deal risk or assigning support tickets. The reduced latency and computational cost make it suitable for high-volume, real-time applications.

Customers can test Koa Decisions via a free Trailhead account, which includes a demo requiring no retraining to add new categories. Previously, enterprise teams spent months building custom predictive models for classification tasks, but Koa Decisions enables rapid deployment by leveraging pre-trained language understanding. This shift accelerates time-to-production from months to hours, allowing businesses to integrate decision-making into existing workflows without manual model selection.