Salesforce and Nvidia announced Koa in San Francisco on September 15, introducing a specialized AI reasoning model for customer relationship management, or CRM. Salesforce says the Agentforce model is intended to handle complex business tasks requiring several steps, including sales and customer service workflows.
According to Salesforce, the model builds on Nvidia Nemotron 3 Super and uses synthetic scenarios reflecting 27 years of CRM experience. Its researchers say training combined public and artificially generated data without customer data. The approach aims to improve tool use while preserving general capabilities.
In a paper posted on arXiv, the researchers describe reinforcement learning that rewards successful tool use. Business workflow instructions become simulated tasks involving repeated exchanges, allowing the model to learn how to complete requests that depend on data. The training links rewards to task completion.
The paper reports improvements over the underlying open model, especially when tools must be used across several exchanges. Koa surpassed a strong proprietary comparison model but remained below the strongest frontier models. These are the authors’ reported test results; they do not establish superiority across all AI tasks.
Salesforce says it controls Koa’s model weights and runs the system within its own infrastructure, keeping customer data inside its controlled environment during operation. The announcement offers businesses another model option for Agentforce. TechCrunch also reported the launch, highlighting the focus on specialized enterprise reasoning.
Koa is already used internally, including in a Slack agent, Salesforce says. Pilot customers include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero. The company expects general availability in US regions in winter 2026; access currently remains limited to selected pilot customers.
Comments