Salesforce Koa: the CRM reasoning model, explained

By Giorgio Bonan. Updated October 21, 2026, published October 21, 2026.

Salesforce Koa is Salesforce's own reasoning model for Agentforce, built with NVIDIA by post-training Nemotron 3 Super on synthetic CRM scenarios and hosted inside Salesforce's trust boundary. It is in pilot with select customers, with general availability expected in U.S. regions in Winter 2026.

Koa was revealed at Dreamforce in September 2026, a few weeks after Claudeforce made Claude a default model across Agentforce.

Why Salesforce built its own model

Salesforce's newsroom explains the history. For most of Agentforce's life, one general model did almost all the reasoning. Over eighteen months Salesforce moved narrow jobs (intent detection, prompt-injection screening, toxicity, reranking) to smaller dedicated models, but left multi-step reasoning to the general model. Koa is the step that brings core reasoning in-house for the most common enterprise tasks.

The commercial logic is cost and control: a CRM-tuned model hosted by Salesforce handles routine, well-scoped work, and frontier models are kept for tasks that need broader judgment.

How Koa was trained

What Salesforce claims, and how much weight to give it

All of these come from Salesforce. CRM Bench is a public benchmark run by Salesforce AI Research, with tasks such as updating an opportunity, routing a case or scheduling a follow-up. No independent evaluation was published by 21 October 2026. Treat the numbers as a direction, and test on your own use cases during a pilot.

Which model does what in Agentforce now

This split is Salesforce's stated direction. Which model actually runs a given agent depends on configuration and on what reaches GA.

Timeline

How to read "11% more precise"

Calling the right action means that, given a request, the agent picks the operation that should run: update this opportunity rather than create a new one, route the case to this queue rather than that one. A gain in that rate matters because a wrong action is often worse than no action: it changes data. The figure compares Koa with "today's default general intelligence models" on Salesforce's own tasks; it does not say which models, and it does not measure your agents. Your pilot should measure the same thing on your tasks: out of 100 requests, how many led to the right action.

Questions to ask in a Koa pilot

1. Which of our agents will run on Koa, and which stay on a frontier model? The answer changes cost and behavior.
2. How is the choice made at runtime? Salesforce describes routing routine work to Koa and judgment work to frontier models; ask what "routine" covers for your agents.
3. What do our own test cases show? Run your ten most common agent tasks on both models and compare the wrong actions, not the average score.
4. What happens to price? No separate price is published; ask whether a Koa-run action draws the same Flex Credits as any other action.
5. Where is it hosted for us? GA is announced for U.S. regions; for other regions, ask for a date in writing.

What it means for buyers

• Price. Salesforce has not published a separate price for Koa. Agentforce usage remains billed through its existing models, see Agentforce pricing.
• Data residency. Koa runs inside Salesforce's infrastructure, which may simplify reviews for teams that hesitate to send CRM data to a third-party model.
• Timeline. GA is announced for U.S. regions in Winter 2026; other regions are not dated.
• The data still decides. A model that calls the right action more often still acts on the fields it reads. Stale stages and close dates produce confident wrong actions, whichever model reasons over them.

For how Koa fits next to older Einstein features, see Einstein vs Agentforce. For the autonomous assistant it is likely to power, see Agentforce Coworker.