Salesforce introduced Koa Decisions on 6 Nov, a decision model built on Google’s open-weight Gemma that reads a request and returns a choice — a category, a yes-or-no, or a score — in a single step, in milliseconds. It is the next version of HyperClassifier, the routing model already inside Agentforce, now with image and audio inputs alongside text, and Salesforce plans to extend it into flows, Apex, prompt templates and a developer API.
What Koa Decisions actually does
Most of an agent’s work is choosing, not writing. Koa Decisions takes a request and a list of categories the user defines, then picks one, answers yes or no, or returns a score. A package question goes to shipping, a login problem to accounts. The answer is often a single token, so other software can act on it at once.
The speed is the point. Salesforce measured its predecessor, HyperClassifier, answering Agentforce routing requests in 26 milliseconds, against about 1.4 seconds for a general-purpose model. Koa Decisions adds what that model could not do: it reads images and audio as well as text, so a photo of a damaged delivery or a voice message can drive the same one-step decision.
Salesforce built Koa Decisions by applying new training techniques to Gemma, Google’s open-weight model, and published the method in a research paper. The approach turns classification, checks and scoring into constrained generation with single-token outputs, so the model reads a request once and answers in one step.
The paper reports 85.5% accuracy on a production intent-routing benchmark, with answers in about half a second and median speed up to roughly 7.5 times faster than the frontier models tested. Those are the team’s own figures, not independent results, and the headline accuracy figure comes from a variant in the paper rather than the shipped model.
Where it goes next
Koa Decisions starts in Agentforce, where HyperClassifier already routes requests, but Salesforce plans to take it further: into flows, Apex and prompt templates for Salesforce builders, plus an API for developers. Anyone can try a demo with a free Trailhead account, adding categories such as returns and watching the model pick them with no retraining.
Pricing and packaging are still to be decided, and availability varies by region under customer agreements, Salesforce says. The demo and the predecessor’s existing deployment do not yet amount to general availability for every planned integration, so the wider rollout is a promise for now, not a finished product.
Our Read
Decision models are having a moment, and Salesforce is not early to it. TypeSafe came out of stealth in mid-September with Jev, and OpenAI made its Decisions API available a month later. Salesforce’s answer is to productize what it already runs: HyperClassifier has routed Agentforce requests since the spring, and Koa Decisions extends that proven job instead of starting over.
The economics are the story. A routing call the predecessor answers in 26 milliseconds instead of 1.4 seconds changes what an agent can afford on every turn, and a small model doing the choosing frees the large model for the writing. It is the same logic behind this week’s launches, from Ant’s 560B Ling to Thinking Machines’ Inkling 2 Small.
What to watch is execution. The paper’s speed and accuracy figures are Salesforce’s own, and the 26-millisecond measurement belongs to the predecessor, so independent tests will matter. So will the timetable: the flows, Apex and API integrations are planned, not shipped, and no date or price has been given.
Why It Matters
Enterprise agents make thousands of small decisions a day, each one currently costing a large-model call. A fast, cheap decision model changes the unit economics of running agents at scale. Whoever owns the routing layer owns a piece of every agent workflow that follows.
Background
Salesforce has built decision models for years: HyperClassifier became the default intent-detection model for Agentforce service and employee templates in spring 2026. Koa Decisions joins the Koa family alongside a separate reasoning model built with NVIDIA, and it arrives as Salesforce keeps buying into AI voice and agents, including its November deal for Cresta.
Source: Salesforce’s announcement and Unite.AI’s report
0 Comments