Details
- Salesforce and NVIDIA unveiled Koa, a CRM-specific reasoning model for Agentforce, built on NVIDIA Nemotron 3 Super and trained on 27 years of Salesforce CRM intelligence using proprietary synthetic data rather than customer data.
- The collaboration combines Salesforce’s enterprise workflow expertise with NVIDIA’s open models and accelerated computing, giving Salesforce full control of Koa’s weights, post-training, and inference inside its own trust boundary.
- Koa is post-trained with supervised fine-tuning and reinforcement learning via Group Relative Policy Optimization, using synthetic multi-step scenarios across more than 14 industries to teach agents step-by-step tool use and decision-making.
- In Salesforce’s CRM benchmark of real-world tasks, Koa matches or exceeds leading models on CRM actions while delivering roughly three times fewer errors, and is already used internally, including a Slack agent and pilots with customers such as 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
- NVIDIA Nemotron-based models and accelerated computing are also being integrated into Missionforce Operations to let government and regulated organizations train and deploy mission-specific agents on their own data in private clouds and air-gapped networks; Missionforce Operations is generally available now in U.S. regions, with post-trained NVIDIA models for select customers in October 2026 and Koa GA in winter 2026.
Impact
Koa signals a shift toward domain-specialized reasoning models that can reliably execute complex CRM workflows, challenging the dominance of generic frontier models in enterprise agents. By pairing open-weight Nemotron with Salesforce-hosted deployment and synthetic, industry-specific training data, the partnership strengthens the case for governed AI agents in highly regulated sectors and government. Over the next 12–24 months, this approach is likely to steer R&D and spending toward task-specific, controllable models embedded in existing enterprise platforms.
