Salesforce Koa brings reasoning in-house on Nvidia Nemotron

Koa arrives as Salesforce's first reasoning model, built on Nvidia's open-weight Nemotron and post-trained for sales, marketing and customer-support tasks. Before it, Agentforce agents that needed to reason through long or multi-step work were routed to frontier models like Claude or ChatGPT through Salesforce's AI gateway, the system that decides which model handles which request. Salesforce's Jayesh Govindarajan said the company had wanted to train its own enterprise-grade frontier model but lacked a suitable pre-trained base, until Nemotron offered a sovereign American model that is state of the art with clear data provenance, unlike Qwen.

The model was trained without actual Salesforce customer data; instead, Salesforce and Nvidia crafted synthetic data that simulated a customer service environment, from irate callers to a sales professional trying to close a deal. Koa will sit alongside other models in Agentforce as an alternative, and Salesforce simultaneously announced a partnership with Anthropic called ClaudeForce, which lets companies use Claude as their AI interface while data remains in Salesforce's system of records.

What this hinges on is whether Koa's token efficiency and open-weight design persuade Agentforce customers to route their reasoning tasks away from frontier models. If it performs well enough, enterprises gain a cheaper, more controllable option; if not, the gateway will keep sending that work to Claude or ChatGPT. Salesforce is not abandoning those partnerships, so Koa's role may be as a complement rather than a replacement.