Koa Decisions: Salesforce's Decision Model Runs in 26ms vs. 1.4s for a General-Purpose LLM

On November 6, 2026, Salesforce published on its official site (salesforce.com/news), bylined by Darvish Shadravan, Senior Director of Product Management, the announcement of Koa Decisions — a model distinct from the Koa unveiled in September with Nvidia. Where Koa reasons through complex, multi-step CRM tasks, Koa Decisions addresses a different need: making a single decision, reading a request and a list of categories defined by the agent's builder, as fast as possible.

The Successor to HyperClassifier, Not a New Version of Koa

Koa Decisions doesn't replace Koa: it succeeds HyperClassifier, the routing model already used internally in Agentforce to direct a request to the right agent or tool. The two models — Koa and Koa Decisions — now coexist within the Koa family, with distinct roles: multi-step reasoning for one, instant single-step decisions for the other.

A Different Foundation: Gemma Instead of Nemotron

• Koa (September 2026) is built by post-training NVIDIA Nemotron 3 Super.
• Koa Decisions is built from Gemma, Google's open model, using the Salesforce data science team's own training techniques.
• The announcement points to a research paper on arXiv for the full methodological detail, without reproducing that data in the announcement itself.

The Latency Figures Salesforce Disclosed

In internal Agentforce routing tests, Salesforce states Koa Decisions responds in about 26 milliseconds, versus about 1.4 seconds for a general-purpose LLM on the same task. No accuracy metric (error rate, accuracy) is published in the announcement: the confidence scores shown in the demo (95%, 99.5%) are illustrative product examples, not a benchmark result. As with the "3x fewer errors" benchmark disclosed in September for Koa, this latency figure remains a vendor-reported number, not independently audited — test it against your own request volume before drawing any conclusion.

Supported Inputs: Text, Image, and Audio

Text remains the base input, but Salesforce adds multimodal support: Koa Decisions can also take an image or an audio clip as input to make its decision — a point that goes beyond HyperClassifier's plain text classification.

Availability and Announced Roadmap

• No pricing is disclosed; a footnote explicitly states that pricing and packaging are subject to change, and availability varies by region and customer agreements.
• Anyone can try a free demo with a Trailhead account, at labs.agentforce.com/koa-decisions.
• Salesforce says it plans to extend the model "well beyond routing" into Flows, Apex, and prompt templates for agent builders, plus an API for developers — these integrations are not yet shipped as of the announcement.

What This Actually Means for Your Agentforce Deployment

• If your Agentforce agent already does routing or simple classification (directing a ticket, picking a tool from a closed list of categories): Koa Decisions targets exactly that case, with an announced latency gain that can matter at high request volumes — but with no pricing disclosed, the ROI math isn't possible yet.
• Don't conflate Koa and Koa Decisions when talking with your Salesforce contact: the first reasons through complex tasks, the second settles a simple decision in milliseconds — these are two separate models to activate and evaluate independently.
• The announced Flows/Apex/prompt template integrations and the developer API aren't available yet: worth tracking for your roadmap, but not to build into a project in progress today.
• Image and audio support opens routing use cases beyond text (an attachment, a voice message) — worth checking concretely against your own volume and formats before any decision, since the published latency figures are for text.

Do you have repetitive routing or classification rules in your Agentforce or Service Cloud setup? Our automation offering audits your existing flows to identify what's worth handing to a decision model before you evaluate its cost.