Salesforce and Nvidia debut Koa, a reasoning model for enterprise agents

Salesforce unveiled Koa, its first reasoning model, at the Dreamforce conference on September 15. Built on Nvidia's open-weight Nemotron model and post-trained jointly by the two companies, Koa is designed for sales, marketing and customer-support tasks.

Koa serves as an open-weight alternative to closed frontier models in Salesforce's Agentforce platform, where customers build agents to handle routine work such as answering service questions or scheduling appointments. According to Salesforce, Koa can be automatically routed through an AI gateway based on need, uses fewer tokens to complete the same work, and does not ingest customer data. It also follows customers' data requirements and security controls embedded within Salesforce.

Previously, agents in Agentforce that required multi-step reasoning would route prompts to frontier models such as Claude or ChatGPT. Jayesh Govindarajan, EVP of Salesforce AI, said the company had long wanted to train its own enterprise-grade model but lacked a suitable pre-trained base model. "Until Nemotron came along, there was no sovereign American pre-trained model that was available, one, and two, that was state of the art, and, three, that had clear data provenance," he said.

Salesforce and Nvidia said they did not use any actual customer data during post-training. Instead, they created synthetic data simulating customer service and sales scenarios, including irate callers and sales professionals working to close deals. Kari Ann Briski, Nvidia's VP of Generative AI Software for Enterprise, said Nemotron offers a unique architecture for token-efficient inference.