Salesforce's Koa Decisions: One Step, One Answer, No Paragraph

Nothing here is sponsored — no affiliate links on this page. This breaking article is not the daily buyer-intent post. Facts below come from the sources listed at the end, checked 6 November 2026. We have not tested Koa Decisions hands-on, and nothing here is a hands-on finding.

Here's the number worth sitting with: every agent you build makes dozens of small decisions per task. Which team gets this request? Is this deal actually at risk? Does this result satisfy what the user asked? And every single one of those decisions, today, usually costs a full LLM call — a model generating sentences just so you can parse a yes out of them.

Salesforce shipped something on 6 November 2026 that skips that whole dance. It's called Koa Decisions, and it doesn't generate text at all. You give it a request and a list of categories you define, and it hands you back a pick, a yes-or-no, or a score — in one step, in milliseconds. That's it. That's the whole product.

What it actually is

Koa Decisions is the next version of HyperClassifier, Salesforce's routing model that's been working inside Agentforce for a while. The new version is built by applying novel training techniques to Google's open-weight Gemma, and it adds something the predecessor didn't have: it accepts text, images, and audio as input.

The mechanics are the interesting part, and they're the opposite of how an LLM works. A large language model writes its answer one token at a time, each new token based on the ones before it. A decision model reads the whole request at once and produces its answer in one step — often a single token. A short answer takes less computing to produce, and other software can act on it right away.

Say a customer types "I can't log in to my account" into a chat window. Should that go to engineering, accounts, or sales? A general-purpose LLM can work it out, but it isn't built for that job, and it can take more than a second. Koa Decisions reads the request and your list of options, picks "accounts" in milliseconds, and returns a confidence score with it. No paragraph. No parsing.

It joins the Koa family of Salesforce models alongside a separate Koa reasoning model built with NVIDIA. The reasoning model does the thinking; Koa Decisions does the choosing.

The numbers, with the vendor label still attached

Salesforce's team published a research preprint alongside the announcement. The figures in it are Salesforce's own, from Salesforce's own measurements — nobody independent has reproduced them yet:

• 85.5% accuracy on a production intent-routing benchmark
• Responses in about half a second
• Median speed ~7.5x faster than the frontier models tested, under identical serving conditions

Against TypeSafe's dedicated Jev system, the preprint reports competitive accuracy with faster median responses — plus multimodal inputs and multi-label outputs that a single-label text system doesn't handle.

Now, the number you'll see quoted everywhere, and the one to be careful with: 26 milliseconds versus 1,446 milliseconds for a frontier model. That measurement belongs to HyperClassifier, the predecessor — Salesforce measured it on routing requests inside Agentforce. It does not establish Koa Decisions' performance. It's useful context for what this family of models has already done in production, but it's not this model's number, and anyone quoting it as Koa Decisions' speed is getting it wrong.

Decision models are having a moment

Salesforce says so itself in the announcement, and the timing backs it up. TypeSafe AI came out of stealth in mid-September with Jev, a model built to pick from a set of options. A month later, OpenAI made its Decisions API generally available. And this same week, Kakao shipped Kanana-2-v-decide, an open-source decision model that returns probabilities instead of text.

That's three vendors in about seven weeks, plus Salesforce, all converging on the same shape: a small model that answers a typed question — pick one, yes or no, score it — instead of generating a paragraph you then have to parse. The category is real now, even if the benchmarks are still vendor-reported.

Where this fits in work you're actually doing

Here's the practical translation. A decision model slots into the gaps in your agent loop — the places where you're currently calling a big model and asking it to answer in JSON so you can extract a field:

• Routing and triage. Classify the incoming request — support, billing, bug report, sales — with a probability attached, so you can set a confidence threshold and escalate the uncertain ones to a human instead of guessing.
• Yes-or-no gates. After a tool returns a result, ask: does this satisfy the original request? That's the cheapest guardrail in an agent loop.
• Scoring and flagging. When a sales deal starts to slip, a decision model can flag the risk and pick which team should step in, while an LLM summarizes what changed. By the time anyone opens the record, the right team already knows. Salesforce's own example runs that from a Sales Cloud record to a Slack alert.
• Permit or block. A decision model can approve or refuse an action — the kind of binary call that currently costs a full generation.

None of this replaces your main model. It replaces the small, expensive calls you make around your main model — the ones that add latency and cost without adding intelligence.

Try it yourself in ten minutes

You don't need to be a Salesforce customer. Sign up for a free Trailhead account and you can run the Koa Decisions demo. It starts with routing a support request: click through, and the model routes the default request to engineering, accounts, or sales and shows a confidence score for each option.

Then do the part that actually demonstrates the idea. Under Choices, add a fourth option — "Returns" is Salesforce's example, but you can add anything — then change the request to something matching, like "I need to send back the order that arrived yesterday." The model picks Returns. No retraining happened. In Salesforce's own demo walkthrough it picked returns with a 99.5% confidence score.

That no-retraining step is the whole pitch for enterprise teams. The old way to classify requests was a custom predictive machine learning project: define the outcome, gather and label data, train, check, repeat. For many enterprise customers, getting from idea to production took months. A decision model starts from a language model that has already learned how "send back the order" relates to "returns" — so a team can add a category and sometimes reach production in hours.

The limits — stated plainly

• No independent verification yet. The 85.5% accuracy, the half-second responses, the 7.5x figure — all of it is vendor-reported until someone else reproduces it. The preprint is public, so the claims are checkable, but nobody has checked them yet.
• The 26ms figure is the predecessor's. HyperClassifier's routing measurement, not Koa Decisions'. Don't let anyone sell it to you as this model's speed.
• Pricing and packaging are not announced. Salesforce's own announcement carries the standard caveat: pricing and packaging are subject to change, availability may vary by region, and access is governed by customer agreements.
• The planned integrations are planned, not shipped. Flows, Apex, prompt templates, and a developer API are on the roadmap. The demo and HyperClassifier's existing deployment inside Agentforce do not establish general availability for every one of those.
• It's a specialist, not a generalist. It won't reason about your problem, write your code, or explain its decisions. If you need explanations, you still need a generative model sitting next to it.
• Nothing here is hands-on. The demo and the model haven't been run for this article, and no figure here is our measurement. The demo is free with a Trailhead account — the fastest way to check Salesforce's claims is to run it yourself.

The bottom line is a measured one. Most of the agent stack is built on the assumption that every judgment needs a full language model. Koa Decisions is a bet that it doesn't — that a small model returning a choice and a confidence score is the right tool for the decisions that sit between the real work. If the numbers hold up under independent testing, the agents you already run get noticeably faster and cheaper, because the decision tax quietly disappears. If they don't, it's still a useful pattern to steal: stop paying generation prices for yes-or-no answers.

Where in your agent loop are you currently parsing a paragraph to get a boolean? That's the call this model is trying to replace.

Sources: Salesforce, "Less Reasoning. More Doing. Meet Koa Decisions." by Darvish Shadravan (6 November 2026, https://www.salesforce.com/news/stories/koa-decisions-salesforce-decision-models); Unite.AI, "Salesforce Introduces Koa Decisions with Image and Audio Support" (6 November 2026, https://www.unite.ai/salesforce-introduces-koa-decisions-with-image-and-audio-support/); Salesforce Engineering, HyperClassifier deployment account (8 July 2026). All facts come from the launch-day announcement and reporting, checked 6 November 2026. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.