Salesforce has introduced Koa, its first reasoning model, developed with Nvidia and unveiled at the company's Dreamforce conference. Built on Nvidia's open-weight Nemotron model, Koa is designed to support enterprise workflows with a sharper focus on sales, marketing, and customer service.
Rather than aiming for broad, general-purpose performance, Koa is tuned for practical business tasks. Salesforce says the model can handle multi-step requests more efficiently, use fewer tokens, and fit more naturally into its Agentforce platform, where companies build AI agents for routine operations such as answering client questions or managing appointments.
One of Koa's key advantages is its enterprise-first design. Salesforce says the model was trained without using customer data, relying instead on synthetic scenarios that simulate real workplace interactions. That approach is intended to support privacy, security, and compliance while still improving performance in business settings.
Nvidia describes Nemotron as a token-efficient architecture built for faster inference and more efficient reasoning. Salesforce also noted that Koa can be routed through its AI gateway depending on the task, giving customers more flexibility in how different models are used across workflows.
The company is not replacing its broader AI ecosystem. Salesforce continues to work with other model providers, including Anthropic, through its ClaudeForce partnership. The result is a more modular strategy: businesses can choose the model that best matches each task while keeping data inside Salesforce's secure environment.
As enterprise AI moves toward specialized, efficient, and privacy-aware systems, models like Koa may help define the next stage of workplace automation.
