Salesforce introduced Koa Decisions, the successor to its HyperClassifier routing model, and says it picked a route in about 26 milliseconds where a general-purpose LLM took about 1.4 seconds.
What Salesforce announced
In a newsroom post, Salesforce says Koa Decisions reads a request and a list of categories a customer defines, then returns a choice in a single step. Its data science team built it by applying new training techniques to Google’s Gemma, an open-weight model, and it adds support for images and audio. Salesforce plans to extend it into flows, Apex and prompt templates, with an API for developers.
The timing figures come from routing requests inside Agentforce, which Salesforce measured at roughly 26 milliseconds for the decision model against about 1.4 seconds for a general-purpose LLM.
Why it matters
An agent makes many small choices, such as which team gets a request or which article to show. Salesforce argues that a model returning one answer is cheaper to run than a large model writing text for each of those steps. The post places the release among other decision models: TypeSafe AI’s Jev in mid-September, and an OpenAI Decisions API that Salesforce says reached general availability a month later.
For a buying committee, the routing layer becomes a line item. The Agentforce price comparison in our report on Agentforce Coworker and the credit-based limits in our coverage of Agent Fabric both depend on how many model calls a task makes.
What the post leaves open
The 26 millisecond and 1.4 second figures are Salesforce’s own, from one routing workload. The post does not give an accuracy figure, name the LLM used in the comparison, or say how decision-model calls are priced, and it notes that pricing and packaging are subject to change. A buyer evaluating agents on speed or cost should ask for those three items, and for the share of agent steps that go to a decision model instead of a large one.
Source: Salesforce Newsroom
