Salesforce announced Koa, its first reasoning AI model, at its annual Dreamforce conference in September 2026. Built on Nvidia’s open-weight Nemotron model, Koa was developed jointly by the two companies to handle sales, marketing, and customer-support tasks within Salesforce’s Agentforce platform.
Before Koa, when an Agentforce agent needed to work through a complex, multi-step task, those requests were routed to frontier models such as Anthropic’s Claude or OpenAI’s ChatGPT. Koa is designed to handle that reasoning workload directly, using fewer tokens — and therefore lower cost — than sending the same tasks to third-party providers.
Jayesh Govindarajan, EVP of Salesforce AI, said the availability of Nemotron was the key enabler. “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 told TechCrunch, contrasting it with Alibaba’s Qwen, whose training data origins he described as unclear.
Salesforce and Nvidia trained Koa using synthetic data rather than actual customer records, simulating scenarios including customer service calls and sales interactions. The approach means no real customer data was ingested into the model, reducing the risk of data leakage. Koa also operates within Salesforce’s existing security and data-compliance infrastructure.
Kari Ann Briski, Nvidia’s VP of Generative AI Software for Enterprise, described Nemotron’s architecture as offering “sovereign AI, time to first token, efficient reasoning” — what she called “the trifecta of things that you need to have.”
Salesforce is not abandoning its existing model partnerships. Alongside the Koa announcement, the company revealed a deal with Anthropic called Claudeforce, which allows enterprises to use Claude as their AI interface while keeping their data secured within Salesforce’s infrastructure. Koa will be offered as an alternative option within Agentforce, routed automatically through the platform’s AI gateway depending on the task at hand.
Source: TechCrunch