Salesforce's Koa Decisions: A Gemma-Based Model That Answers in One Token

Salesforce introduced Koa Decisions on November 6, 2026, a decision model built from Google's open-weight Gemma that accepts text, images, and audio — and it plans to push the technology well beyond its current role routing Agentforce requests into broader Salesforce workflows and developer tools.

The release matters because it draws a sharper line inside enterprise AI stacks. Instead of asking one large model to do everything, Salesforce is betting that fast, cheap decision models can handle routing, checks, and scoring while general-purpose LLMs handle writing and reasoning.

What is Koa Decisions?

Koa Decisions is the next version of HyperClassifier, a model Salesforce already runs in production inside Agentforce. It reads a request plus a user-defined list of categories, then returns a selection, a yes-or-no decision, or a score — in a single step. Salesforce describes that operation as taking milliseconds.

The model joins the Koa family alongside a separate Koa reasoning model that Salesforce built with NVIDIA.

The core idea is a division of labor. A general LLM can write, explain, and reason. A decision model can make the same repeatable call thousands of times a day at a fraction of the compute cost.

How does it differ from an LLM?

The distinction lies in how the answer gets produced. A large language model generates text token by token, with each new token conditioned on everything before it. A decision model reads the whole request and emits a compact answer — often a single token — in one step.

Because the answer is short, it requires less computing to produce, and other software can act on it immediately. In Salesforce's design, customers do not need to manually pick a model for each job.

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. An agent can then route the ticket while reserving a general LLM for tasks that actually need one.

What does the research show?

Salesforce's Koa-decision research preprint frames 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 reported numbers are concrete:

• 85.5% accuracy on a production intent-routing benchmark
• Responses in about half a second
• Median speed up to roughly 7.5 times faster than the frontier models tested under identical serving conditions
• Against the dedicated Jev/TypeSafe system: competitive accuracy and faster median responses, plus multimodal inputs and multi-label outputs that single-label text systems do not handle

Those are research results for the Koa-decision approach, not shipped product guarantees. But they show why Salesforce thinks the one-token architecture can compete with both general LLMs and purpose-built classifiers.

What did HyperClassifier already prove in production?

Koa Decisions builds on a predecessor with a real track record. 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 how long the class names are.

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.

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

Can you try it?

Yes. 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 the category list and see confidence scores for the options.

The demo points at the model's practical pitch: adding a category such as "returns" requires no retraining. Salesforce contrasts that with custom predictive-classification projects that took months for many enterprise customers; a decision-model implementation can sometimes reach production in hours.

Where will it plug in?

One documented workflow example shows the pattern Salesforce has in mind. 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
• A developer API

Buyers should read the fine print. 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.

Why does this matter?

The practical significance is a clearer division of labor inside enterprise software. Decision models handle routing, checks, and scoring. 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.