✉ The Friday AI Brief: the week's 5 best AI stories, tools & comparisons — in your inbox every Friday morning.

Diagram of Salesforce Koa Decisions taking text, audio and image inputs and returning routing, permit-or-block, and scoring decisions

Salesforce Introduces Koa Decisions: A Decision Model That Reads Text, Images and Audio — and Answers in Milliseconds

Salesforce has introduced Koa Decisions, the next version of the HyperClassifier model that routes requests inside Agentforce — and this one can read images and listen to audio, not just text. Announced November 6, it takes a request plus a list of categories you define and returns a single decision — a pick, a yes-or-no, or a score — in one step, in milliseconds. Why it matters: a large share of what enterprise agents actually do is not writing, it is choosing, and Salesforce is betting a small, specialized model should make those calls instead of a full-size language model.

November 6, 2026

What happened

Salesforce’s data science team built Koa Decisions by applying new training techniques to Gemma, Google’s open-weight model. It succeeds HyperClassifier, the routing model already embedded in Agentforce, which sends each incoming request — “Where’s my package?”, for example — to the part of the agent built to handle shipping, FAQs, promotions, or an escalation. Koa Decisions keeps that job and expands it: it adds multimodal support for images and audio alongside text, and returns its answer with a confidence score that other software can act on immediately.

The details

The speed gap is the pitch. Salesforce says that when it measured routing requests in Agentforce, its decision model answered in roughly 26 milliseconds, compared with about 1.4 seconds for a general-purpose LLM. The team’s research paper, summarized by Unite.AI, reports 85.5% accuracy on a production intent-routing benchmark and median responses up to roughly 7.5 times faster than the frontier models tested under the same serving conditions — vendor-side research figures, not independent benchmarks. One practical difference from older classifiers: a team can add a brand-new category — Salesforce’s demo adds “Returns” to a routing list — and the model starts using it immediately, with no retraining and no labeled dataset to build first.

Anyone can try it: the Koa Decisions demo is public through Agentforce Labs with a free Trailhead account, no Salesforce contract required. Salesforce also placed the launch in context itself, noting that decision models are having a moment — TypeSafe AI’s Jev came out of stealth in mid-September, and OpenAI made its Decisions API generally available a month later. Koa Decisions sits in the wider Koa family next to a separate Koa reasoning model Salesforce built with NVIDIA.

What is not live yet

The broader rollout is a plan, not a shipping product. Salesforce says it intends to extend Koa Decisions well beyond routing — into Flows, Apex, and prompt templates for builders, plus an API for developers — but its own fine print warns that pricing and packaging can change and availability varies by region and customer agreement. Treat the Agentforce routing role and the public demo as what exists today; the rest is roadmap.

Why it matters

Here is the subscriber’s take: the interesting shift is not another giant model, it is the unbundling of one. Routing, scoring, and permit-or-block checks are high-volume, low-glamour decisions that burn frontier-model tokens and add a second of latency to every interaction. A purpose-built model that answers in 26 milliseconds — and can now look at a photo or a voice note while it decides — makes agent workflows faster and cheaper in a way users will actually feel. If Salesforce delivers the Flows, Apex, and API expansion, “which model makes the small calls” becomes a design choice every enterprise builder has to make, not just a Salesforce detail. Everything AI, tested — and increasingly, routed by something much smaller.

Frequently asked questions

What is Salesforce Koa Decisions?

It is a decision model introduced on November 6, 2026 — the next version of HyperClassifier, the model that routes requests inside Agentforce. Built from Google’s open-weight Gemma, it reads text, images, and audio, then returns a category pick, a yes-or-no answer, or a score in a single step.

How fast is Koa Decisions compared with a regular LLM?

Salesforce measured its decision model at roughly 26 milliseconds on Agentforce routing requests, versus about 1.4 seconds for a general-purpose LLM. Its research paper reports 85.5% accuracy on a production intent-routing benchmark.

Can I try Koa Decisions without being a Salesforce customer?

Yes. Salesforce offers a public demo through Agentforce Labs; you only need a free Trailhead account.

When will Koa Decisions reach Flows, Apex, and a developer API?

Salesforce says those expansions are planned, but it has not given a general-availability date, and it cautions that pricing, packaging, and regional availability can change.

Sources: Salesforce (official announcement, “Meet Koa Decisions,” November 6, 2026), Unite.AI (research-paper summary).

Leave a Comment

Your email address will not be published. Required fields are marked *

Get the 5 best AI tools every week

Top AI news, tools, and prompts — one short email. Free, unsubscribe anytime.

Run a newsletter of your own? Monetize and grow it with SparkLoop →

Scroll to Top