Salesforce announces Koa, a CRM reasoning model built on NVIDIA Nemotron

Salesforce and NVIDIA announced Koa on September 15, 2026, Salesforce's first CRM reasoning model for its Agentforce platform. Salesforce built Koa by post-training NVIDIA Nemotron 3 Super on a synthetic dataset modeled on CRM deployments, and says no customer data was used. The model is intended to help agents work through multistep business tasks and choose the right tools. Koa is available to select pilot customers now, with general availability expected in winter 2026 in U.S. regions.

The details

• Training scenarios spanned more than 14 industries, including manufacturing, financial services, healthcare and travel.
• Training combined supervised fine-tuning with reinforcement learning, using NVIDIA NeMo RL, NeMo Gym and NeMo AutoModel.
• Salesforce says Koa matches or exceeds leading models on its internal CRM benchmark and makes three times fewer errors, without publishing absolute scores.
• Salesforce controls the weights and runs post-training and inference inside its own infrastructure.
• Pilot customers are 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero, and Koa already powers an internal Slack agent.

Context

Koa was announced on the same day as AIforce, which brings Salesforce into Claude, Slack and Lightning, and follows the August 26 Claudeforce partnership with Anthropic. The release also extends Nemotron-based models to Missionforce for government and regulated organizations, with Missionforce Operations generally available in U.S. regions. In its Q2 fiscal 2027 results Salesforce reported Agentforce ARR above $1.5 billion. Marc Benioff said Salesforce's most valuable asset "isn't our platform."

Not disclosed

• Pricing for Koa usage
• Absolute benchmark scores