Salesforce and NVIDIA have announced Koa, Salesforce’s first CRM reasoning model for Agentforce, developed through a technical collaboration using NVIDIA Nemotron.
Designed for complex, multistep workflows, Koa is built to help Agentforce agents reason through business processes and select the appropriate tools to complete tasks. Salesforce post-trained NVIDIA Nemotron 3 Super using a proprietary synthetic dataset modeled on enterprise knowledge gathered from nearly three decades of CRM deployments.
In Salesforce’s CRM benchmark, which includes real-world tasks such as updating opportunities, routing service cases, and scheduling follow-ups, Koa matches or exceeds leading model performance on CRM actions while recording three times fewer errors, according to Salesforce.
Built For Enterprise Workflows
Koa was trained entirely on synthetic scenarios, with no customer data used in its training. The scenarios simulate enterprise workflows across more than 14 industries, including manufacturing, financial services, healthcare, and travel.
Salesforce applied Supervised Fine-Tuning (SFT) and reinforcement learning using Group Relative Policy Optimization (GRPO) with NVIDIA’s NeMo tools. The approach focuses on teaching the model not only to provide an answer but also to determine the appropriate sequence of actions and tool calls needed to complete a task.
Salesforce controls Koa’s model weights and performs post-training and inference within its own infrastructure, allowing the company to offer a Salesforce-hosted reasoning model designed to operate within its trust boundary.
Koa is already being used within Salesforce, including an agent in Slack that helps employees find information and complete everyday tasks. The model is also entering customer pilots with organizations including 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
Expanding AI For Regulated Environments
The collaboration also extends to Missionforce, Salesforce’s platform for government and highly regulated organizations. Salesforce and NVIDIA are bringing Nemotron-based models and accelerated computing to environments such as private clouds and air-gapped networks, where organizations require greater control over models, data, and deployment.
Missionforce Operations is designed to digitize and automate government workflows including procurement, supplier management, and logistics. Post-trained NVIDIA models can be adapted using an organization’s operational data and terminology while operating within its own environment.
Koa is currently available to select Agentforce pilot customers, with general availability expected in winter 2026 in U.S. regions. Post-trained NVIDIA models for Missionforce Operations are expected to become available to select customers in October 2026.
For more information, visit Salesforce and NVIDIA.