The new AI model, called Koa, is one of the biggest announcements Salesforce made this week at the huge tech conference Dreamforce. Koa is the company’s first inference model built on Nvidia’s open-weight Nemotron model. The two companies collaborated on Koa’s post-training to help him excel in sales, marketing, and customer support-related tasks.
Koa is a shining example of how the corporate world’s AI needs are disconnected from what Frontier Labs offers. Its own AI lab allows companies to upload files, code, prompts, and feedback directly to models and agents that they hope to spend millions of dollars on.
However, with this model, Salesforce provides enterprise customers with:
Open-weight alternatives to closed-frontier models Models that are trained to perform a specific work task (rather than solving impossible math problems) Models that do not ingest any real customer data, so they cannot leak data to other users Models that help reduce AI spending because they use fewer tokens to do the same work Models that can be automatically routed through AI “gateways” depending on your needs Customer data requirements and Salesforce A model that follows all of the security built into it.
Koa is offered as an alternative to other models offered by Salesforce on its Agentforce platform, where the company’s customers build agents to handle rote tasks such as answering customer service questions and scheduling appointments.
“We’ve been building a lot of small, task-specific language models that are part of the Agentforce portfolio,” Jayesh Govindarajan, vice president of Salesforce AI, told TechCrunch. “But inference has always relied on frontier model providers. Until now.”
Before Koa, when agents needed to reason about long-running or multi-step tasks, those prompts were routed through Agentforce’s AI gateway (the system that decides which models handle which requests) to frontier models like Claude and ChatGPT.
“One of the reasons we haven’t done this before is to train our own enterprise-grade Frontier models, and we’ve always wanted to do that. But the challenge has always been the lack of pre-trained base models. Until Nemotron came along, there were no sovereign U.S. pre-trained models available. One and two were state-of-the-art, and three were the most advanced. The first one had a clear provenance of data. I had no idea what Qwen was training on,” Govindarajan said. A popular Chinese open weight model produced by Alibaba.
Post-training such a model means moving the model from a general-purpose system to a system familiar with sales and customer support knowledge. To do so, Salesforce and Nvidia did not use any actual data from Salesforce customers. Instead, we created synthetic data that mimics our customers’ patterns.
“We actually simulated a customer service environment with customer service professionals in our personas, from an irate customer calling a customer service center to a salesperson trying to close a deal,” Govindarajan explained.
Koa is intended to help Salesforce customers better perform the work tasks they want their agents to perform, and to be cheaper in terms of token consumption than sending the same tasks to Claude or ChatGPT.
Nemotron “gives us a unique architecture for efficiently inferring tokens,” Kari Ann Briski, Nvidia vice president of enterprise generative AI software, told TechCrunch. “It’s like a combination of three things you need: sovereign AI, time to first token, efficient inference, and tokenomics of all of that.”
However, Salesforce isn’t completely abandoning Anthropic or OpenAI. The company just announced a partnership with Anthropic called ClaudeForce. This allows businesses to use Claude as an AI interface, but the data remains in Salesforce’s system of record and protected by its infrastructure.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
