Salesforce launches Koa: All about the tech firm's first CRM reasoning AI model powered by NVIDIA Nemotron

Salesforce has announced Koa, its first CRM reasoning model for Agentforce, built in collaboration with NVIDIA using the Nemotron 3 Super model. Koa is designed to handle complex CRM tasks and reportedly delivers better performance with three times fewer errors on Salesforce’s CRM benchmark.

Salesforce has announced Koa, its first CRM reasoning model for Agentforce, at its Dreamforce conference. Built in collaboration with NVIDIA, Koa is designed to handle complex, multi-step business workflows such as updating opportunities, routing customer cases and scheduling follow-ups.

Salesforce claims Koa was developed by a post-training NVIDIA Nemotron 3 Super customised with a synthetic dataset of Salesforce's own enterprise workflows. The company stated that it has not used any customer information to train the model. The inference process will take place in the infrastructure of Salesforce, which means that customer data will stay in the Salesforce trust zone.

Koa focuses on CRM tasks

Koa is different from other general AI models, as it is specially trained to handle enterprise CRM workflows. Salesforce said the training scenarios are synthetic and span over 14 industries, including financial services, healthcare, manufacturing, and travel.

The scenarios represent activities that a sales and customer service representative might do. They include generating leads, qualifying opportunities and resolving service cases. Supervised fine-tuning and reinforcement learning with NVIDIA's NeMo RL, NeMo Gym and NeMo AutoModel tools powered Salesforce.

Koa is expected to perform as well as or better than the top models on its CRM benchmark while using 3X fewer mistakes with CRM actions, according to Salesforce.

Designed for secure enterprise AI

According to Salesforce, it owns the model weights and conducts post-training and inference on its own infrastructure. This is designed to provide businesses with more control over their data and the use of AI.

Pragya Singha Roy