Salesforce unveiled Koa, its first reasoning model, at the Dreamforce technology conference. Built on Nvidia’s open-weight Nemotron model, Koa was subsequently trained by the two companies to specialize in sales, marketing and customer support tasks.
The model will be offered as an option on the Agentforce platform, where Salesforce customers create AI agents for tasks such as answering customer questions or scheduling appointments. Salesforce Vice President of AI Jayesh Govindarajan said the company had previously developed small, task-specific language models but had so far relied on frontier model providers for operations requiring reasoning. Long-running or multi-step tasks were routed to models such as Claude or ChatGPT.
According to Govindarajan, the main reason Salesforce had previously been unable to develop its own enterprise-grade model was the lack of an advanced, US-based foundation model with transparent data provenance. Nemotron met this need. No real data from Salesforce customers was used in the training process. Instead, synthetic data was generated to simulate various scenarios, ranging from support staff dealing with angry customers to sales personnel trying to close deals.
Salesforce aims for Koa to outperform Claude or ChatGPT on the tasks customers want agents to carry out while costing less in terms of token usage. Nvidia Vice President of Enterprise Generative AI Software Kari Ann Briski said Nemotron’s inference architecture offers token efficiency, a fast initial response and effective reasoning.
Salesforce is not moving away from Anthropic and OpenAI entirely. The company also announced the ClaudeForce partnership. This arrangement allows companies to use Claude as their artificial intelligence interface while keeping data securely within Salesforce’s infrastructure.
Why it matters
Adding Koa to Agentforce gives Salesforce customers an alternative to frontier model providers for multi-step sales, marketing and support tasks. Salesforce’s shift from small, task-specific models to an enterprise-grade reasoning model shows how Nvidia’s open-weight foundation model can be used for in-house specialization. The decision to train with synthetic data ensures that real customer data is not used, while also requiring the model’s performance under real-world working conditions to be tested separately. The continuation of the ClaudeForce partnership indicates that Koa expands the options within Agentforce rather than entirely replacing Anthropic and OpenAI models. The key unanswered question is whether the targets of higher performance and lower token costs will be validated in customer use.
Background
Nvidia is not a new name in the FikirPilot archive: over the past 90 days, we have published 20 stories mentioning the company, the latest dated September 12, 2026.
Term: agent
An artificial intelligence agent is software that calls tools and carries out multi-step tasks to achieve a goal rather than producing a single response.