Salesforce announced Koa, its first reasoning model, at this week’s Dreamforce conference. Koa is built on Nvidia‘s open-weight Nemotron model, with the two companies jointly post-training it to excel at sales, marketing, and customer-support tasks. TechCrunch frames Koa as a marker of how enterprise AI needs are diverging from what frontier labs offer. Proprietary labs, the article says, want enterprises to upload files, code, prompts, and feedback directly into their models and agents, and spend millions doing so. Koa instead gives Salesforce customers an open-weight alternative to closed frontier models, plus a model trained for specific work tasks rather than for solving hard math problems.

Salesforce also emphasizes that Koa has not ingested any actual customer data and therefore cannot leak it; that it helps reduce AI spending because it uses fewer tokens to achieve the same work; that it can be automatically routed via an AI gateway depending on the need; and that it follows all customer data requirements and security embedded in Salesforce. Koa is offered as an alternative to other models in Agentforce, Salesforce‘s platform where customers build agents to handle rote tasks such as answering customer service questions or scheduling appointments.

Before Koa, when an agent had to reason through a long-running or multi-step task, Agentforce‘s AI gateway—which decides which model handles each request—would route the prompt to a frontier model like Claude or ChatGPT. Salesforce‘s EVP of AI, Jayesh Govindarajan, told TechCrunch that the company had built many small task-specific language models in Agentforce‘s portfolio, but reasoning had always come from frontier model providers. “Until now,” he said.

Govindarajan also explained that Salesforce had wanted to train its own enterprise-grade frontier model but was blocked by the lack of a suitable pre-trained base. “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 added that the company has no idea what Qwen, Alibaba’s popular Chinese open-weight model, trains on.

For post-training, Salesforce and Nvidia did not use any actual customer data. Instead, they crafted synthetic data that mimicked customer patterns. Govindarajan described simulating a customer service environment with personas from irate customers calling in to sales professionals trying to close a deal. The goal is for Koa to be better at the work tasks Salesforce customers want an agent to do—and cheaper in terms of token burn—than sending those same tasks to Claude or ChatGPT.

Nvidia‘s vice president of generative AI software for enterprise, Kari Ann Briski, highlighted Nemotron‘s unique inference architecture for token efficiency. She called the combination of sovereign AI, time to first token, and efficient reasoning “the trifecta of things that you need to have” for tokenomics.

Even with Koa, Salesforce is not leaving Anthropic or OpenAI behind. It announced a partnership with Anthropic called ClaudeForce, which lets companies use Claude as their AI interface while their data remains in Salesforce‘s system of records, secured by Salesforce‘s infrastructure.