Salesforce Unveils Koa: A Custom AI Model for Enterprise Work

Salesforce has developed Koa, its first AI reasoning model optimized for enterprise work. Koa runs inside Salesforce's trust boundary and was co-engineered with NVIDIA. It is post-trained on their open model, Nemotron 3 Super.

A general model reasons about tasks from first principles, which can lead to inconsistent results. In contrast, a frontier model learns specific tasks and applies them consistently. Koa falls into the latter category, having been trained on simulated enterprise workflows across various industries.

The model's training involved reinforcement learning, where it tried thousands of times to complete tasks correctly. This approach improved its ability to carry out multi-turn tool use, a critical aspect of enterprise work.

Koa is already being used in employee workflows and is being piloted with customers across regulated industries. Early benchmarking results show that Koa outperforms strong general-purpose models on CRM-specific tasks.