Salesforce introduced Koa, its first CRM reasoning model, at Dreamforce this week, appearing onstage with NVIDIA CEO Jensen Huang and Salesforce CEO Marc Benioff. Koa was built by post-training NVIDIA's Nemotron 3 Super on a proprietary synthetic dataset drawn from nearly three decades of enterprise CRM deployments, using NVIDIA's NeMo RL, NeMo Gym and NeMo AutoModel tools. The training corpus covers synthetic enterprise scenarios spanning 14-plus industries, including manufacturing, financial services, healthcare and travel.
Because Nemotron is an open model, Salesforce fine-tuned and now runs Koa entirely within its own infrastructure. Salesforce controls the model weights, and the company says no customer data is used during either training or inference — a point Salesforce platform and engineering president Rohan Kumar emphasized when introducing Koa.
On Salesforce's CRM Bench, a benchmark built from real-world CRM tasks like updating an opportunity, routing a case or scheduling a follow-up, Koa already matches or exceeds leading model performance while producing three times fewer errors on those specific actions.
Koa is currently running inside Salesforce itself, powering an internal employee agent in Slack. It moves into customer pilots in October 2026 as a customer-selectable model within Agentforce, with an initial cohort that includes Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine and Xero. General availability is expected in winter 2026 for U.S. regions.
Huang used the appearance to argue that AI safety is 'an engineering problem' rather than a reason to slow down, telling Benioff that companies should 'run as fast as you can' but pause if a product isn't ready. He also noted that open-model adoption has grown from roughly 30% to about 70% among enterprises building custom AI systems, a trend he tied directly to Salesforce's decision to fine-tune and self-host Nemotron for Koa.
