Salesforce and NVIDIA today announced Koa, Salesforce’s first CRM reasoning model, developed in collaboration and unveiled at Dreamforce on September 15, 2026. Built on NVIDIA’s open‑weight Nemotron 3 Super model, Koa is trained using a proprietary synthetic dataset that encapsulates 27 years of enterprise CRM workflows—spanning industries such as manufacturing, financial services, healthcare, and travel. This enables agents to reason through multistep tasks like updating opportunities, routing cases, and scheduling follow‑ups with three times fewer errors compared to leading models in Salesforce’s CRM benchmark.(salesforce.com)
A key differentiator is that no customer data was used in training; instead, synthetic scenarios simulate real-world workflows, complete with personas, tool calls, and action sequences. Salesforce retains full control over model weights and inference, ensuring all operations occur within its trust boundary, addressing enterprise concerns around data privacy and governance.(salesforce.com)
Koa is already deployed internally—such as in a Slack-based agent—and is now entering customer pilot programs with organizations including 1‑800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero. General availability is expected in winter 2026 in U.S. regions.(salesforce.com)
This launch underscores a broader trend: enterprises are demanding AI that not only generates content but can reason and act within the context of business workflows, while maintaining control and compliance. By embedding decades of CRM intelligence into a reasoning model, Salesforce and NVIDIA are advancing agentic AI toward practical, enterprise-ready deployment.