Salesforce unveiled its first reasoning model, Koa, at Dreamforce, signaling a shift among enterprise customers from general-purpose large models toward solutions that prioritize cost efficiency, data control, and specific business tasks. Koa was co-developed by Salesforce and NVIDIA, and is post-trained on NVIDIA’s open-weight Nemotron model.
For sales and customer service tasks
Koa will be offered as an optional model within the Salesforce Agentforce platform, primarily used for sales, marketing, and customer support scenarios to help businesses build automated agents that handle repetitive tasks such as customer service inquiries and appointment scheduling.
Jayesh Govindarajan, Executive Vice President of Salesforce AI, said that previously, when agents needed to handle multi-step or long-chain tasks, the system typically routed requests through Agentforce’s AI gateway to advanced models like Claude or ChatGPT. With the launch of Koa, Salesforce is now bringing some of this reasoning capability back into its own model ecosystem.
Emphasize data isolation and cost efficiency
Salesforce stated that Koa's training did not use any real customer data, but instead simulated enterprise scenarios using synthetic data, including customer service centers handling emotionally upset users and sales representatives closing deals.
- Adopt an open-weight approach as an alternative to closed frontier models.
- Train for specific job tasks rather than seeking to demonstrate general capabilities.
- Do not directly access customers' real data to reduce the risk of data leaks.
Salesforce also stated that Koa uses fewer tokens when handling similar tasks, helping to reduce the cost of AI adoption for enterprises. Kari Ann Briski, Vice President of Enterprise Generative AI Software at NVIDIA, said that Nemotron’s inference architecture places greater emphasis on token efficiency and response speed.
The enterprise model roadmap is beginning to diverge.
This release also highlights the growing divergence between enterprise software companies and cutting-edge AI labs. For many enterprise customers, the focus is no longer solely on the upper limits of model capability, but on whether the model can be deployed within existing systems, meet data requirements, and control long-term usage costs.
Salesforce executives said the company had long wanted to train its own enterprise-grade frontier models but lacked a suitable pretrained foundation. The arrival of Nemotron has now provided this capability.
However, Salesforce has not abandoned collaboration with external model providers. The company recently announced ClaudeForce in partnership with Anthropic, enabling enterprises to use Claude as an AI interface while keeping their business data within Salesforce’s systems and infrastructure.