At Dreamforce, Salesforce announces Koa, a specialized CRM reasoning model built on Nvidia's Nemotron architecture, designed to power AI agents in Agentforce (Mike Wheatley/SiliconANGLE)

Salesforce and Nvidia unveiled a groundbreaking AI model called Koa at this week's Dreamforce conference. Koa is Salesforce's first reasoning model, constructed on Nvidia's open-weight Nemotron model. The collaboration between the two companies aimed to train Koa to specialize in sales, marketing, and customer support-related tasks.

This new model represents a significant shift in the AI landscape, as it caters to the unique needs of enterprises, which differ from the demands of frontier labs. Salesforce's approach differs from other companies in the industry, which prefer to upload files, code, prompts, and feedback directly into their models and agents, investing heavily in the process.

In contrast, Salesforce offers Koa as an alternative to the models available in its Agentforce platform, where customers create agents to manage routine tasks such as answering customer inquiries or scheduling meetings. Jayesh Govindarajan, EVP of Salesforce AI, explained that while Salesforce has developed numerous task-specific language models for Agentforce, reasoning capabilities were previously dependent on frontier model providers like Claude or ChatGPT.

However, Koa aims to address this limitation by training their own enterprise-grade model. The challenge was the lack of a pre-trained base model that met the necessary criteria, including being a sovereign American model, state-of-the-art, and having clear data provenance. Govindarajan noted that they addressed this challenge by employing synthetic data that mimicked customers' patterns, without using any actual data from Salesforce's customers.

Koa is designed to excel at the work tasks Salesforce customers expect from their agents, while being more cost-effective in terms of token usage compared to sending the same tasks to Claude or ChatGPT. Nvidia's Koa boasts a unique architecture for efficient inference and token economics, ensuring optimal performance. Despite this groundbreaking achievement, Salesforce has not abandoned its partnerships with Anthropic and OpenAI.

Instead, they have launched a partnership called ClaudeForce, which allows companies to utilize Claude as their AI interface while keeping their data within Salesforce's secure infrastructure.