Yesterday, Salesforce and Nvidia announced Koa, Salesforce’s first CRM reasoning model for Agentforce.
Built on Nvidia’s post-training Nemotron 3 Super, Koa is designed to handle multistep CRM tasks such as updating sales opportunities, routing support cases, and scheduling follow-ups. Salesforce built the model using a proprietary synthetic dataset based on nearly three decades of CRM experience. The scenarios cover more than 14 industries, including financial services, health care, manufacturing and travel.
The company said no customer data was used to train Koa. Salesforce also controls the model weights and handles post-training and inference inside its own trust boundary. That setup is important for companies that want AI agents to take actions across sensitive business systems without sending customer information outside their controlled environment.
Salesforce said Koa matches or exceeds leading models on its CRM benchmark while producing three times fewer errors on tasks such as case routing and opportunity updates.
But the results come with an important qualification. A research paper by Koa’s developers found that the model substantially outperforms its Nemotron base and performs strongly on enterprise agentic tasks, while still falling short of the strongest frontier models.
That distinction helps explain Salesforce’s broader strategy. Koa is not being positioned as the only model Agentforce customers will need. Instead, Salesforce can use a specialized model for repetitive, high-volume CRM work and potentially reserve more capable frontier models for difficult tasks. That could reduce the number of expensive model calls needed for routine workflows without forcing customers to give up access to outside AI models altogether.
Training agents to take the right steps
Koa’s specialization comes from how Salesforce trained it. The company used supervised fine-tuning and reinforcement learning with Group Relative Policy Optimization, supported by Nvidia’s NeMo tools.
Rather than simply teaching the model what a good answer looks like, the synthetic scenarios were built around sequences of actions and tool calls. This lets training focus on whether an agent correctly completes a workflow.
Marc Benioff, Salesforce chair and CEO, said, “With Koa, the knowledge is put inside the model itself.”
The approach could prove particularly useful in CRM, where a successful AI interaction often requires several connected actions rather than a single response.
Koa is already being used internally at Salesforce, including in a Slack agent that helps employees find information and complete tasks. Customer pilots include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.
Salesforce and Nvidia are also extending their work to Missionforce, bringing Nemotron-based models and accelerated computing to government and regulated organizations. The models are intended for environments including private clouds and air-gapped networks.
Koa is available to select Agentforce pilot customers, with general availability expected in winter 2026 in U.S. regions. Missionforce Operations is already generally available, while post-trained Nvidia models are expected to reach select customers in October.
The bigger business bet
Koa shows Salesforce trying to solve a practical problem in enterprise AI: not every task needs the most powerful model available.
A specialized model that makes fewer mistakes, uses fewer resources, and stays within Salesforce’s infrastructure could make agent-based automation easier to deploy at scale. The tradeoff is that narrower models may struggle when workflows become unusually complex, making model routing and access to stronger systems important.
For channel partners, that could shift the Agentforce opportunity beyond simply helping customers choose an AI model. Partners may increasingly be asked to identify which business processes are predictable enough for specialized models, which require frontier models, and where human approval should remain part of the workflow.
That creates work around process mapping, CRM integration, governance, and model routing. A sales organization, for example, might use Koa for routine opportunity updates and follow-ups while escalating unusual requests to a more capable model or an employee. The value comes from designing that handoff correctly rather than assigning every task to the most expensive model available.
The economics will matter, too. If Salesforce can use Koa to handle large volumes of routine CRM actions with fewer errors and less compute, customers could have a clearer path to expanding agent deployments without allowing AI costs to rise alongside every new automated workflow.
For partners building Agentforce practices, Koa therefore represents more than another model choice. It points toward an enterprise AI architecture in which specialized and general-purpose models work side by side, with partners helping customers decide which model should handle which job.
The interesting test for Salesforce will be whether Koa can turn CRM-specific knowledge into a measurable advantage in cost, reliability, and scale while preserving the flexibility enterprises expect from modern AI platforms.
Related reading: For more on how partners are turning Salesforce AI deployments into real-world enterprise workflows, read how OSF Digital is helping customers move AI from pilots to production.