Salesforce Unveils 'Koa,' Its First Reasoning Model Based on NVIDIA Nemotron, Reducing CRM Errors by Two-Thirds

Salesforce unveiled Koa, its first CRM reasoning model based on NVIDIA Nemotron 3 Super, at Dreamforce. The model underwent post-training on a synthetic dataset derived from approximately 30 years of enterprise CRM deployment data and is fully controlled within Salesforce infrastructure.

Technical Implementation and Performance

Koa was developed by combining Supervised Fine-Tuning and Reinforcement Learning, utilizing NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel. Synthetic enterprise scenarios spanning more than 14 industries, including manufacturing, finance, healthcare, and travel, were used as training data.

According to CRM Bench, Salesforce's proprietary benchmark, Koa demonstrated performance equal to or better than existing major models in real-world CRM tasks such as opportunity updates and case routing. Notably, the error rate decreased to one-third of previous levels.

Data Security and Deployment Plans

Salesforce stated that it manages model weights within its own systems and does not use customer data during training or inference. Rohan Kumar, Salesforce EVP and General Manager of Platform and Engineering, emphasized, "Not a single byte of customer data was used."

Koa is currently operational within Salesforce's internal Slack agent and will be deployed in customer pilot tests with clients such as Formula 1 and UChicago Medicine starting in October. General Availability in the United States is scheduled for Winter 2026.

Jensen Huang's Vision

NVIDIA CEO Jensen Huang argued that AI, as an infrastructure layer, has ushered in an era where it "knows everything and can do everything." He defined AI safety as an engineering problem, emphasizing that innovation and safety are not mutually exclusive. He also noted that the adoption rate of open models has surged from 30% in early last year to approximately 70% currently, predicting that all companies will become agentic AI companies.