Salesforce Koa brings CRM reasoning in-house

Salesforce and NVIDIA announced Koa on September 15, giving Agentforce a Salesforce-controlled reasoning model built specifically for complex CRM work. Rather than sending every multi-step task to a general-purpose frontier model, Salesforce can now run a specialist whose weights, post-training, and inference remain within its own trust boundary. For now, however, Koa is limited to select customer pilots.

From open weights to a CRM specialist

Koa starts with NVIDIA’s open-weight Nemotron-3-Super-120B foundation model. Salesforce then specialized it through post-training with public and synthetically generated data, including CRM scenarios spanning more than 14 industries, and reinforcement learning. Salesforce says no customer data was used to train Koa.

Reasoning through CRM work step by step

Koa is designed to plan and use tools across multi-step workflows rather than simply generate an answer. Its target jobs include updating opportunities, routing and resolving service cases, scheduling follow-ups, generating leads, and qualifying opportunities. That focus gives Agentforce a model trained around the actions its agents are expected to perform.

The benchmark gain comes with limits

Salesforce says Koa matched or exceeded leading-model performance on CRM actions with three times fewer errors in its own CRM benchmark. The technical report offers a narrower result: Koa improved on its Nemotron base and surpassed GPT-4.1 in the reported tool-use and agentic-reasoning evaluations, while remaining below the strongest frontier models overall. Current pilots include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero. General availability in U.S. regions is expected in winter 2026.