Salesforce and NVIDIA have introduced Koa, a CRM-focused reasoning model designed to help Agentforce AI agents handle complex, multistep workflows and take actions across enterprise processes.
Koa was developed by post-training NVIDIA Nemotron 3 Super using a proprietary synthetic dataset modeled on enterprise CRM workflows accumulated from nearly three decades of Salesforce deployments. Salesforce said no customer data was used to train the model.
The model is designed to go beyond generating responses by helping AI agents determine the appropriate sequence of actions and tool calls for tasks such as updating opportunities, routing service cases, generating leads, and scheduling follow-ups.
On Salesforce’s CRM benchmark, which includes real-world CRM tasks, Koa matches or exceeds leading model performance on CRM actions while producing three times fewer errors, according to Salesforce.
Salesforce used Supervised Fine-Tuning (SFT) and reinforcement learning through Group Relative Policy Optimization (GRPO), using NVIDIA’s NeMo tools to specialize Koa for enterprise workflows spanning more than 14 industries.
The model also gives Salesforce control over its model weights and deployment environment. Koa is post-trained and runs within Salesforce’s infrastructure, keeping customer data within its trust boundary during inference.
Koa is already being used internally, including through an Agentforce agent in Slack, and is moving into customer pilots 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.
The collaboration also extends to Missionforce, Salesforce’s platform for government and regulated organizations. NVIDIA models and accelerated computing will support mission-specific AI deployments in private clouds, classified environments, and air-gapped networks.
Missionforce Operations is designed to automate government workflows such as procurement, supplier management, and logistics. Post-trained NVIDIA models are expected to become available to select customers in October 2026.
