Salesforce and Nvidia Launch Koa Reasoning Model for Enterprise AI

Salesforce is expanding its enterprise AI strategy with Koa, its first reasoning model, developed with Nvidia and built on the chipmaker’s open-weight Nemotron model.

The companies worked together to post-train Koa for business tasks involving sales, marketing and customer support. The model is designed to give Salesforce customers another option inside Agentforce, the company’s platform for building AI agents that can handle tasks such as answering customer questions and scheduling appointments.

Salesforce brings reasoning into its own model portfolio

Salesforce has previously developed smaller, task-specific language models for Agentforce, but more complex reasoning work has depended on models from frontier AI providers.

Jayesh Govindarajan, executive vice president of Salesforce AI, said reasoning had been an area where Salesforce relied on frontier model providers until Koa. Before the new model, Agentforce could route long-running or multi-step reasoning requests through its AI gateway to outside models such as Anthropic’s Claude or OpenAI’s ChatGPT.

Govindarajan said Salesforce had wanted to develop an enterprise-grade reasoning model but lacked a suitable pre-trained foundation model. He pointed to Nemotron as offering an open model with the performance and data provenance Salesforce wanted.

That open-weight foundation gives Salesforce more control over how the model is adapted for enterprise work, while giving customers an alternative to relying exclusively on closed frontier systems.

Koa was trained without Salesforce customer data

Salesforce and Nvidia used synthetic data to post-train Koa rather than feeding actual Salesforce customer data into the model. The synthetic material was designed to reproduce business scenarios and patterns that the model could encounter while handling enterprise tasks.

Govindarajan said the company’s simulated environments ranged from customer service interactions, including conversations with frustrated customers, to scenarios involving sales professionals working to close deals.

The approach is intended to give Koa specialised knowledge of sales and customer-support workflows without using customer information as training material. Salesforce is also positioning the model to operate within the company’s existing data and security requirements.

Salesforce targets lower token use for enterprise tasks

Another part of the strategy is efficiency. Salesforce says Koa is intended to perform the kinds of tasks its customers assign to agents while consuming fewer tokens than sending the same work to general-purpose frontier models.

Kari Ann Briski, Nvidia’s vice president of generative AI software for enterprise, said Nemotron was designed with an inference architecture focused on token efficiency. She described the combination of sovereign AI, faster time to the first token and efficient reasoning as important to the economics of enterprise AI deployments.

Agentforce’s AI gateway can determine which model should handle a request, meaning Koa can exist alongside other models rather than replacing them outright. That lets Salesforce customers use a more specialised model for certain business workflows while still accessing frontier systems when needed.

Koa does not mean Salesforce is leaving Claude behind

Salesforce is continuing to work with Anthropic even as it introduces its own reasoning model. The company also announced Claudeforce, its partnership with Anthropic, designed to let companies use Claude as an AI interface while keeping business data governed through Salesforce’s systems and infrastructure.

The two initiatives illustrate Salesforce’s broader model strategy: customers can use specialised models such as Koa for enterprise workloads while still accessing models through the same platform.

For Salesforce, Koa adds a reasoning model built specifically around the business tasks Agentforce is expected to perform, rather than depending on a single outside model provider for every complex request.