Drawn from synthetic scenarios that simulate real-world enterprise workflows across more than 14 industries — including manufacturing, financial services, healthcare, and travel — the training corpus was designed to reflect the reasoning, tool use, and decision-making skills agents perform across CRM workflows.

“The genesis of this was looking at all the things we’ve learned in building out our products, the strategy we’ve created, and asking how could we augment a model to make it very specific to this job,” Kumar said.

Under the hood

Salesforce post-trained the model by applying Supervised Fine-Tuning (SFT) and reinforcement learning to Group Relative Policy Optimization (GRPO) and NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel. The idea is to leverage Koa agents for complex sales and service tasks.