Salesforce and Nvidia challenge frontier AI labs with Koa

The introduction of Koa signals a growing divide between the requirements of the enterprise sector and the offerings of major AI laboratories. While companies like OpenAI and Anthropic prioritize expansive, general-purpose models, Salesforce is betting on a more focused approach. Koa is engineered to navigate multi-step business processes—such as managing irate customer service interactions or closing complex deals—without the need for the massive compute power and token consumption associated with larger, closed-source competitors.

Crucially, the development of Koa relied on synthetic data rather than actual customer information, a design choice intended to mitigate data privacy concerns and prevent the leakage of proprietary corporate details. Jayesh Govindarajan, Salesforce’s EVP of AI, noted that the company previously lacked a suitable, transparent base model to build upon before the arrival of Nemotron. By using this architecture, Salesforce claims it can offer a more secure, cost-effective alternative that integrates directly into its Agentforce platform. While this move reduces dependence on third-party providers, Salesforce maintains its partnerships with companies like Anthropic, ensuring that businesses can still utilize external models while keeping their core records secured within Salesforce’s own infrastructure.