Announcing Koa: Salesforce’s First CRM Reasoning Model, Built on NVIDIA Nemotron

Salesforce and NVIDIA announced Koa, Salesforce’s first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. Developed through collaboration, Koa is trained using nearly three decades of CRM data to help agents navigate multistep workflows, such as updating opportunities or routing cases, with greater accuracy. The model is grounded in real-world enterprise processes and policies, reflecting how businesses operate across industries. Salesforce controls the model weights and performs inference within its own infrastructure, ensuring customer data remains within the trust boundary.

Koa was post-trained on NVIDIA Nemotron 3 Super using a proprietary synthetic dataset modeled after enterprise knowledge from Salesforce CRM deployments. The training focused on simulating real-world workflows across 14 industries, including manufacturing, financial services, and healthcare. By applying supervised fine-tuning and reinforcement learning, the model learned to take precise actions step-by-step to achieve goals, not just produce correct answers. In Salesforce’s CRM benchmark, Koa matched or exceeded leading model performance on CRM tasks while reducing errors by two-thirds.

The model is already in use within Salesforce, including an agent in Slack that assists employees with tasks, and is expanding to customer pilots with companies like 1-800Accountant, Baxter Credit Union, and Xero. Salesforce executives and pilot customers highlight Koa’s ability to handle complex, industry-specific workflows, such as accounting tax rules or healthcare coordination, by reasoning through multi-step processes. Salesforce emphasizes that Koa’s training corpus was built entirely from synthetic scenarios, avoiding the use of any customer data.

Salesforce is also integrating NVIDIA Nemotron-based models into Missionforce, extending control over model, data, and deployment to government and regulated organizations. Missionforce Operations, now generally available, digitizes and automates government workflows like procurement and logistics. Post-trained NVIDIA models will be available to select customers in October 2026, enabling deployment in secure environments such as private clouds and air-gapped networks.