Salesforce and Nvidia introduce Koa, a reasoning model for enterprise work

Salesforce has introduced Koa, a reasoning model designed for enterprise tasks, at its Dreamforce conference. According to TechCrunch, it is Salesforce’s first reasoning model and was built with Nvidia on top of the open-weight Nemotron model. The companies post-trained Koa for work such as sales, marketing and customer support rather than presenting it as a general model for every possible problem.

The announcement is significant because Salesforce has traditionally depended on external frontier models when an Agentforce workflow needed to handle a long or multi-step request. Those requests could be routed through Salesforce’s AI gateway to models such as Anthropic’s Claude or OpenAI’s ChatGPT. Koa gives Salesforce another option inside that routing layer, alongside the smaller task-specific models already used in Agentforce.

Salesforce told TechCrunch that the post-training did not use actual customer data. Instead, the companies simulated customer-service and sales situations and created synthetic training examples. That distinction matters for enterprise buyers: a model that has not been trained on customer records may reduce one route by which sensitive information could be carried into model development, although deployment still depends on permissions, data connections and operational safeguards.

Salesforce also positions Koa as a more token-efficient option. If it can complete the intended business tasks with fewer tokens than a larger frontier model, companies could reduce inference costs for high-volume agent workflows. That is a claim about the model’s intended economics, not independent proof that every workload will be cheaper or more accurate.

The partnership does not mean Salesforce is abandoning closed model providers. The company has also announced Claudeforce, which lets customers use Claude as an interface while Salesforce remains the system of record. The broader message is that enterprise AI is becoming a routing and governance problem as much as a model-selection problem. For users, makers and IT teams, Koa could make it easier to keep routine work near business data while reserving larger models for tasks that genuinely need them. The practical test will be whether Salesforce publishes clear evaluations for quality, latency, cost and data controls across real Agentforce use cases.