Salesforce’s Koa Decisions is available to test—but its builder and API surfaces are still a roadmap

Salesforce has introduced Koa Decisions, a Gemma-derived model designed to choose among supplied options in one decoding step instead of asking a general-purpose language model to reason through every routing decision. The launch says the model accepts text, image and audio inputs and can return a category, Boolean choice or score. A public demo is available through a free Trailhead account. Salesforce launch

That makes Koa relevant to agent builders who currently spend a large-model call on narrow tasks such as selecting a subagent, permitting or blocking an action, or routing a support request. But the current product boundary matters: Salesforce says it plans to extend Koa Decisions into Flow, Apex, prompt templates and a developer API. Those are future surfaces, not capabilities the announcement establishes as generally available today. Pricing, packaging and regional availability are also explicitly left subject to change. Salesforce launch

Do not turn 26 milliseconds into a Koa SLA

The launch cites a roughly 26 ms routing measurement versus about 1.4 seconds for a general-purpose LLM. Its wording points back to Salesforce’s earlier decision-model work inside Agentforce, rather than presenting that number as a fresh Koa Decisions benchmark. The safer reading is that Salesforce is illustrating the decision-model category, not promising that every Koa deployment will complete in 26 ms. Salesforce launch

The linked research paper gives a different, more controlled comparison. On Salesforce’s production intent-routing benchmark, the authors report 85.5% accuracy and latency of about half a second under identical serving conditions, up to roughly 7.5× faster at the median than the frontier models tested. The paper also says Koa supports multimodal inputs and multi-label outputs. Those figures are vendor-authored research results, not an independent benchmark or production service-level commitment. Koa-decision paper

What to validate before replacing an LLM router

Teams evaluating Koa should separate three questions:

1. Decision quality: test the real label set, including overlapping and newly added categories, rather than relying on the launch demo.
2. Latency boundary: measure end-to-end service time—including network and Salesforce orchestration—not only model decoding.
3. Integration availability: confirm that the required invocation surface exists under the customer’s contract. The announced Flow, Apex and developer API paths remain planned.

Koa Decisions points toward a useful two-model architecture: a compact decision model handles constrained choices, while a larger model remains responsible for generation and multi-step reasoning. Salesforce has made that architecture tangible enough to test. It has not yet made every promised integration surface available, nor supplied a Koa-specific 26 ms production guarantee. Salesforce launch