Salesforce Unveils Koa: A CRM Reasoning Model for Multi-Step Business Work
Salesforce has introduced Koa, its first CRM reasoning model for Agentforce, as a specialized system for multi-step business work rather than a general chatbot. Currently in pilot phase with select customers, Koa is expected to become generally available in U.S. regions by Winter 2026.
Koa is designed to handle CRM tasks that require several decisions and tool calls, such as updating opportunities, routing cases, scheduling follow-ups, selecting the right business tool, and completing multi-turn workflows. It operates with structured Salesforce workflows and can be selected in Data Cloud's generative-model catalog at the Agentforce organization level or for individual agents and sub-agents.
Koa was built by post-training NVIDIA's open-weight Nemotron-3-Super-120B model with Group Relative Policy Optimization (GRPO), a reinforcement-learning method. The training pipeline turns Agent Script workflow definitions into simulated, multi-turn CRM scenarios, and the model is trained to do the CRM job, not just sound like it understands the job.