Salesforce and Nvidia Launch Koa: Enterprise CRM Reasoning Model Built on Nemotron

On September 15, 2026 at Dreamforce, Salesforce and Nvidia announced Koa, a specialized reasoning model designed exclusively for CRM work inside the Agentforce platform. Rather than relying on a general frontier model, Salesforce post-trained Nvidia’s open-weight Nemotron 3 Super on a large synthetic dataset that encodes nearly three decades of enterprise CRM processes, policies, and multi-step workflows across more than a dozen industries. Critically, no real customer data was used in training or inference. The result is a model that matches or exceeds leading general-purpose models on Salesforce’s internal CRM Benchmark while producing three times fewer errors on common actions such as updating opportunities, routing cases, and scheduling follow-ups. Because Salesforce controls the weights and runs inference inside its own trust boundary, enterprises retain tight data residency and security guarantees. Koa is already powering an internal employee agent in Slack and is live with select pilot customers including Formula 1, UChicago Medicine, and Xero; broader availability is planned for winter 2026 in U.S. regions. The launch highlights a growing trend: domain-specific models that prioritize reliability, token efficiency, and compliance over raw general intelligence. For product and engineering teams building customer-facing agents or internal automation, this approach reduces both risk and operating cost. If your organization is evaluating specialized AI for sales or support workflows—or needs custom iOS, Flutter, or n8n automation that plugs cleanly into CRM systems—let’s discuss how the same principles of controlled, high-accuracy agents can be applied to your stack.