Salesforce has introduced Koa, its first reasoning model specifically engineered for customer relationship management, during the Dreamforce conference. The model is built on NVIDIA Nemotron 3 Super and was trained using a proprietary synthetic dataset derived from nearly thirty years of enterprise CRM deployments. Salesforce states that the model is designed to handle complex CRM workflows, including updating opportunities, routing cases, and scheduling follow-ups.
The development process utilized NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel to perform supervised fine-tuning and reinforcement learning. According to Salesforce, the training corpus encompasses synthetic scenarios across more than 14 industries, such as healthcare, manufacturing, financial services, and travel. Internal benchmarks from Salesforce’s CRM Bench indicate that Koa matches or exceeds the performance of leading models while producing three times fewer errors on specific CRM tasks.
Because the underlying NVIDIA Nemotron model is open, Salesforce was able to fine-tune and deploy Koa entirely within its own infrastructure. Rohan Kumar, president of platform and engineering at Salesforce, confirmed that no customer data was utilized during the training or inference phases. The company maintains control over the model weights, and the system operates locally within Salesforce’s existing environment.
Koa is currently active within Salesforce, powering an employee agent integrated into Slack. Customer pilots are scheduled to begin in October, featuring participants such as Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine, and Xero. General availability for the model is expected in U.S. regions by winter 2026.
During the announcement, NVIDIA CEO Jensen Huang emphasized the importance of safety in AI development, describing it as an engineering challenge that must be addressed alongside innovation. Huang noted that companies should prioritize building reliable systems, suggesting that if a product’s safety or functionality is in doubt, development should be paused to ensure the technology is properly refined before release.