Salesforce and NVIDIA said on September 15, 2026, that they are launching Koa, Salesforce’s first CRM reasoning model for Agentforce. The companies said Koa is built on NVIDIA Nemotron and is designed to reason through multistep CRM work, while Salesforce keeps the model weights and all post-training and inference inside its own trust boundary.
The timing matters because Salesforce is not pitching Koa as another generic AI layer. It is framing the model as a specialized tool for users who need software that can think through CRM actions inside Salesforce CRM, with the company saying Koa already matches or exceeds leading models on CRM actions with three times fewer errors in its benchmark.
That benchmark claim is doing a lot of the work. Salesforce said Koa was post-trained from NVIDIA Nemotron 3 Super using a proprietary synthetic dataset, with no customer data used at all. Yet the company also says the model was grounded in 27 years of Salesforce CRM intelligence and in synthetic scenarios that mirror how agents work across CRM workflows, from generating leads and qualifying opportunities to resolving service cases, each one built around a persona, a task list and the tool calls needed to finish the job. In practice, that means the model is being sold as both safely isolated and deeply experienced at the same time.
Marc Benioff put that pitch plainly, saying the most valuable thing Salesforce has built is the accumulated knowledge of how enterprise business actually works, and that Koa places that knowledge inside the model itself. He said the system runs entirely inside the customer trust boundary. Jensen Huang said every company needs useful AI tailored to its knowledge, expertise and work, adding that NVIDIA Nemotron open models give Salesforce the base to turn decades of enterprise expertise into specialized AI that can reason and securely take action.
Salesforce and NVIDIA also said they are bringing Nemotron-based models and accelerated computing into Missionforce, where government and regulated organizations would get control over their model, data and deployment environment. The companies said that version of the effort is meant for private clouds, air-gapped networks and other secure settings, extending the same approach beyond CRM into mission-specific AI.
What remains unanswered is how much of Koa’s edge comes from the synthetic training data and how much comes from the Nemotron base itself. Salesforce has shown the outline of a product that is tightly controlled, commercially targeted and built to handle enterprise workflows without exposing customer records, but it has not given a rollout date or explained how broadly the model will be deployed beyond the launch announcement. For now, the clearest fact is that Salesforce is betting its next CRM push on a model that learns from its own history without touching a customer’s.
