Salesforce and Nvidia have unveiled Koa, a reasoning model built specifically for enterprise CRM tasks, giving Salesforce an alternative to routing complex Agentforce workloads through general-purpose models such as Claude or ChatGPT.
Koa is Salesforce’s first CRM reasoning model. Rather than building a foundation model from scratch, Salesforce post-trained Nvidia’s open-weight Nemotron 3 Super 120B using supervised fine-tuning and reinforcement learning. Its training data combines public material with synthetic scenarios modeled on nearly three decades of Salesforce CRM experience. Salesforce says no customer data was used.
Those simulations reproduce multi-step workflows such as qualifying sales opportunities, routing service cases, scheduling follow-ups and selecting the correct software tools across more than 14 industries. The objective is not to outperform frontier models at mathematics, coding or general knowledge, but to spend less computation solving the narrower tasks businesses repeatedly need.
Salesforce claims Koa matches or exceeds leading models on its own CRM benchmark while producing three times fewer errors. Its accompanying research paper offers an important qualification: Koa surpasses a strong proprietary baseline on some evaluations but remains behind the strongest frontier models overall. The results have not yet been independently validated.
That distinction explains why Koa could still matter to OpenAI, Anthropic and other model providers. Enterprises do not necessarily need the most capable general model for every request. A specialized model that uses fewer tokens, runs inside Salesforce’s infrastructure and keeps inference data within the company’s trust boundary could handle a significant portion of everyday agent workloads more cheaply.
Salesforce is not abandoning external AI providers. It recently expanded its Anthropic partnership through Claudeforce and Agentforce can continue routing jobs between different models. Koa is already being piloted with customers including Formula 1, Xero and UChicago Medicine. General availability in US regions is expected in winter 2026.