Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear

Your AI-powered customer service tools can now reason through complex tasks without sending data to external frontier AI labs.

Salesforce and Nvidia have collaborated to develop Koa, a new reasoning model specifically for enterprise tasks. Built on Nvidia's open-weight Nemotron model and then post-trained by Salesforce, Koa is designed to excel in sales, marketing, and customer support functions. Unlike proprietary models from frontier AI labs that require enterprises to upload sensitive data, Koa uses synthetic data and is designed to be token-efficient, potentially reducing AI spending for businesses. It functions within Salesforce's Agentforce platform as an alternative to routing complex reasoning tasks to external models like OpenAI's ChatGPT or Anthropic's Claude. This move offers enterprises a specialized, secure, and cost-effective AI reasoning capability that adheres to their data security and compliance requirements without direct data exposure to third-party models.

The development of Koa signifies a strategic shift in enterprise AI, emphasizing specialized reasoning capabilities over general-purpose frontier models. By leveraging Nvidia's Nemotron, Salesforce addresses the enterprise demand for sovereign AI with clear data provenance, a stark contrast to the opaque training data of many leading models. The focus on synthetic data for training is a novel approach to ensure customer data remains secure and private within the Salesforce ecosystem, mitigating risks associated with proprietary AI data ingestion.

This initiative directly challenges the business models of major AI labs by offering a performant, cost-effective, and secure alternative for core enterprise reasoning tasks. Enterprises can now deploy AI agents that handle complex workflows internally, reducing reliance on expensive, generalist models and enhancing data control. The availability of open-weight, task-specific models like Koa signals an evolving market where customization and security are paramount for widespread AI adoption in business.