Salesforce has introduced Koa, its first reasoning AI model, developed in collaboration with Nvidia. The model is built on Nvidia's open-weight Nemotron architecture and is specifically fine-tuned for sales, marketing, and customer service tasks.
The training process for Koa relied exclusively on synthetic datasets rather than real customer data, which Salesforce claims eliminates data leakage risks. The model is designed to be integrated into the Agentforce platform to handle repetitive tasks like scheduling and customer support inquiries.
Salesforce executives noted that Koa aims to reduce operational costs by consuming fewer tokens compared to third-party frontier models. The company previously relied on external providers like OpenAI and Anthropic for complex reasoning tasks, but now seeks more control over its AI infrastructure.
Despite the launch of Koa, Salesforce continues its partnerships with other AI providers. The company announced the ClaudeForce project, which allows enterprise users to utilize Claude as an interface while keeping business data secured within the Salesforce ecosystem.
