Salesforce released its first inference model, Koa, at the Dreamforce conference, indicating that enterprise customers' demand for AI is gradually shifting from general large models to products that place more emphasis on cost, data control, and specific business tasks. Koa was jointly developed by Salesforce and NVIDIA, and is further trained based on NVIDIA's open-weight model Nemotron.

For sales and customer service tasks

Koa is mainly used in sales, marketing, and customer support scenarios, and will serve as an optional model within the Salesforce Agentforce platform. This platform primarily helps enterprises set up automated agents to handle repetitive tasks such as customer service inquiries and appointment arrangements.

Salesforce AI Executive Vice President Jayesh Govindarajan stated that previously, if agents needed to handle multi-step or long-chain tasks, the system would usually route the requests through the Agentforce's AI gateway to forward them to cutting-edge models such as Claude or ChatGPT. After the launch of Koa, Salesforce began to reintegrate some of this reasoning capability into its own model system.

Emphasize data isolation and cost efficiency

Salesforce indicates that the training for Koa did not use any real customer data; instead, it simulated business scenarios using synthetic data, including how customer service centers deal with emotionally agitated users and how sales staff facilitate transactions.

• Adopting an open-weight scheme as an alternative to closed-frontier models
• Training is aimed at specific work tasks, rather than pursuing the demonstration of general abilities.
• Does not directly access customers' real data, reducing the risk of data leakage.

Salesforce is also said that Koa consumes less token when handling similar tasks, which helps to reduce the cost for enterprises using AI. Kari Ann Briski, vice president of NVIDIA's enterprise generative AI software, stated that the inference architecture of Nemotron places more emphasis on token efficiency and response speed.

Enterprise model paths begin to diverge

This release also reflects that the gap between enterprise software companies and cutting-edge AI laboratories is widening. For many corporate clients, the focus is no longer just on the upper limit of model capabilities, but on whether they can deploy within existing systems, meet data requirements, and control long-term usage costs.

Salesforce Executives stated that the company has always hoped to train its own enterprise-level cutting-edge models in the past, but lacked a suitable pre-trained foundation. Nemotron Once it appeared, this condition was finally met.

However, Salesforce has not given up on cooperating with external model manufacturers. The company recently also announced a collaboration with Anthropic to launch ClaudeForce, which allows enterprises to use Claude as an interface for interacting with AI, while continuing to retain business data within the systems and infrastructure of Salesforce.