Salesforce Launches Koa Decisions, a Gemma-Based Router for AI Agents

By Daniel Okafor

Salesforce introduced Koa Decisions on November 6, 2026, a decision model built from Google's open-weight Gemma that accepts text, images, and audio. The company plans to expand the technology beyond Agentforce request routing into Salesforce workflows and developer tools.

The model is the next version of HyperClassifier. It reads a request and a user-defined list of categories, then returns a selection, a yes-or-no decision, or a score in one step. Salesforce describes that operation as taking milliseconds. Koa Decisions joins the Koa family alongside a separate Koa reasoning model built with NVIDIA.

How does a decision model differ from an LLM?

The mechanism matters. A large language model generates text token by token, with each new token based on those before it. A decision model reads the whole request and emits a compact answer, often a single token, in one step. Because that answer is short, it requires less computing to produce, and other software can act on it immediately.

Consider an account-login problem. The model can select the accounts team from a list that also includes engineering and sales, then return a confidence score. That lets an agent use a specialized model for repeatable decisions while reserving a general LLM for writing and reasoning. In Salesforce's design, customers do not need to manually select a model for each job.

What do the research numbers show?

The team's Koa-decision research preprint formulates classification, semantic endpointing, Boolean checks, and scoring as constrained generation with single-token outputs. Atomic label tokens and supervised fine-tuning reduce those tasks to deterministic one-step decoding.

The authors report:

• 85.5% accuracy on a production intent-routing benchmark
• responses in about half a second
• median speed up to approximately 7.5 times faster than the frontier models tested under identical serving conditions
• competitive accuracy against the dedicated Jev/TypeSafe system, with faster median responses

Against Jev/TypeSafe, the model also handles multimodal inputs and multi-label outputs that single-label text systems do not support.

What did HyperClassifier prove in production?

Koa Decisions builds on a predecessor already embedded in Agentforce. A Salesforce Engineering account describes how HyperClassifier interprets user intent and routes agents to action sets before a response is generated. In voice interactions, it determines whether someone has finished speaking.

Its earlier design represents each class with a unique special token and caches initial prompt computations, so repeated instructions are not processed again. That keeps output steps fixed regardless of class-name length.

HyperClassifier was fine-tuned on GPT-OSS-20B and can classify across as many as 200 labels, according to a July 8, 2026 deployment account. It became the default intent-detection and routing model for Agentforce Service and Employee Agent templates in Spring 2026.

Salesforce measured HyperClassifier's routing response at roughly 26 milliseconds, compared with 1,446 milliseconds for a general frontier model. That production measurement belongs to the predecessor and does not establish Koa Decisions' performance.

Who can try it, and what comes next?

Anyone, including people who are not Salesforce customers, can try the Koa Decisions demo after registering for a free Trailhead account. Users can edit the request and category list and see confidence scores for the options.

Salesforce says adding a category such as returns requires no retraining. It contrasts that with custom predictive-classification projects that took months for many enterprise customers; a decision-model implementation can sometimes reach production in hours.

One documented workflow example shows the division of labor concretely. A decision model flags an at-risk sales deal and selects the team that should intervene, while an LLM summarizes what changed, moving from a Sales Cloud record to a Slack alert.

Planned integrations include flows, Apex, prompt templates, and a developer API. Pricing and packaging can change, availability may vary by region, and access is governed by customer agreements. The demo and HyperClassifier's existing deployment do not establish general availability for every planned integration.

The practical significance is a clearer split inside enterprise software: decision models handle routing, checks, and scoring, while LLMs handle explanation and reasoning. Koa Decisions adds multimodal inputs and broader planned integration points to that architecture; its enterprise reach now hinges on how those integrations move beyond the demo and the routing functions HyperClassifier already performs.