Salesforce Unveils Koa Decisions to Route AI Tasks in Milliseconds

Salesforce has announced the release of Koa Decisions, an upgraded version of its internal HyperClassifier routing model. The new tool is built to handle specific choices, such as selecting workflows, scoring requests, or providing yes-or-no answers, in a fraction of the time required by traditional large language models.

Speed and Efficiency for Enterprise Workflows

While general-purpose large language models generate text sequentially, decision models read entire requests at once and return a choice in a single step. In tests measuring routing requests within Salesforce’s Agentforce system, the decision model completed tasks in roughly 26 milliseconds, compared to about 1.4 seconds for a standard large language model.

Building on Open-Weight Foundation Models

The company’s data science team created Koa Decisions by applying new training techniques to Google’s Gemma, an open-weight model. The update introduces multimodal support, allowing the system to process images and audio alongside text. Salesforce plans to integrate the technology into developer APIs, prompt templates, and software building tools.

• Koa Decisions answers requests in milliseconds instead of seconds.
• The model uses Google’s Gemma as its foundation.
• Salesforce plans to expand the tool into developer APIs and flows.