Salesforce introduced Koa, its first reasoning AI model, during the Dreamforce conference. The model is built on Nvidia’s open-weight Nemotron architecture and is specifically optimized for sales, marketing, and customer support tasks.
Koa functions as an alternative to existing models within the Agentforce platform. Previously, Salesforce relied on third-party frontier models like Claude or ChatGPT to handle complex, multi-step tasks through its AI gateway.
To train Koa, Salesforce and Nvidia utilized synthetic data that simulated professional environments, such as customer service interactions and sales deal closures. The companies confirmed that no actual customer data was used during this process.
The model aims to provide more efficient reasoning and lower token costs compared to general-purpose frontier models. Nvidia noted that the architecture focuses on sovereign AI, reduced latency, and efficient tokenomics.
Despite the launch of Koa, Salesforce continues its collaboration with other providers. The company recently announced ClaudeForce, a partnership with Anthropic that allows users to integrate Claude while keeping their data within Salesforce infrastructure.
