Salesforce and Nvidia launch Koa for CRM agents

Salesforce and Nvidia have introduced Koa, Salesforce’s first CRM reasoning model for Agentforce. The model is built on Nvidia Nemotron and post-trained for business workflows such as updating opportunities, routing service cases and coordinating follow-ups.

Salesforce says Koa was trained on synthetic enterprise scenarios rather than customer data. The company controls the model weights and runs post-training and inference inside its own trust boundary, which gives customers a Salesforce-hosted option for agents that need to reason through longer workflows.

The practical point is that Salesforce wants more reasoning work to happen in a specialized enterprise model instead of always routing difficult tasks to general frontier models. TechCrunch reports that Koa is meant to be cheaper in token use and better aligned with sales, marketing and support tasks, while Salesforce will still support outside models through its Agentforce gateway.

Koa is already used internally at Salesforce and is moving into customer pilots, including 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero. Salesforce and Nvidia are also bringing Nemotron-based models into Missionforce for government and regulated customers that need private or air-gapped deployments.