Salesforce has announced Koa, its first reasoning model built specifically for CRM, at Dreamforce 2026.
Developed in collaboration with NVIDIA, Koa is built by post-training NVIDIA Nemotron 3 Super for the kind of complex, multi-step work that Salesforce wants Agentforce agents to handle.
The model is designed to help agents reason through CRM workflows, use the right tools, and take actions to reach a goal, rather than simply generate an answer. Let’s learn more about this model.
What Is Salesforce Koa?
Koa is a CRM reasoning model for Agentforce. Its purpose is to handle the reasoning involved when an agent has to work through several steps before it can complete a CRM task.
That distinction is important because an enterprise agent often has to do more than understand a request. It may need to interpret the available CRM context, determine which information is relevant, choose a tool or action, use the result, and continue from there.
Salesforce has therefore trained Koa around the reasoning, tool use and decision-making involved in CRM workflows across the customer lifecycle. The company describes these workflows as including areas such as lead generation, opportunity qualification, and service resolution. (Source: Salesforce Investor Relations)
In practical terms, Koa is meant to help Agentforce move from understanding a CRM task to working through the steps needed to complete it.
How Is Salesforce Koa Built?
Koa is based on the NVIDIA Nemotron 3 Super, but Salesforce has post-trained the model specifically for CRM.
NVIDIA provides the Nemotron open-model foundation, while Salesforce uses its accumulated CRM knowledge to specialize that foundation for enterprise workflows. Salesforce describes the result as a model grounded in the processes, workflows and operational policies involved in how businesses actually use CRM.
The training corpus is entirely synthetic. Salesforce says it was modeled on nearly three decades of CRM deployments and covers enterprise workflows across more than 14 industries, including manufacturing, financial services, healthcare and travel. No customer data was used to train Koa.
There is another important part of this architecture. Salesforce controls Koa’s model weights and runs the model within its own infrastructure, keeping inference inside Salesforce’s trust boundary.
That gives Salesforce control over the specialized model rather than simply routing CRM requests to an external general-purpose model.
How Is Koa Trained for CRM?
Salesforce’s approach to training Koa is centered on how an agent performs work, rather than only on the information an agent needs to know.
The synthetic scenarios simulate the reasoning, tool use, and decision-making that Agentforce agents perform. Salesforce says each scenario pairs a persona with specific tasks and maps the sequence of actions and tool calls required to complete them.
Salesforce then used Supervised Fine-Tuning (SFT) and reinforcement learning with Group Relative Policy Optimization (GRPO) during post-training. The company used NVIDIA’s NeMo RL, NeMo Gym and NeMo AutoModel as part of that process.
The result Salesforce is aiming for is a model that can do more than produce the right response. It needs to select the right action and work through a task step by step.
Salesforce measures Koa using CRM Bench, a benchmark made up of real-world CRM tasks. The benchmark includes tasks such as updating an opportunity, routing a case, and scheduling a follow-up.
The benchmark results Salesforce has published provide a clearer picture of what it expects the specialization to achieve. Compared with the general-intelligence models Salesforce tested, the company reports:
- 3× fewer errors on CRM actions
- 11% greater precision when selecting the right action
- 2.1× greater reliability in recalling customer context
- 15% better context retention in long conversations
These are Salesforce’s own benchmark results, so they should be read as vendor-reported performance rather than independent comparative testing.
How Does Koa Work with Agentforce?
Koa is a model option within the Agentforce architecture rather than a replacement for Agentforce itself.
Agentforce provides the surrounding agent environment, including the instructions, grounding data, tools and actions available to an agent. And, Koa provides a specialized reasoning model that can be used within that environment.
Salesforce has made Koa available in several parts of its AI stack. It can be selected as a managed LLM in the Data Cloud generative models catalogue, enabled as an organization-wide model provider in Agentforce, or selected at the agent and sub-agent level. Salesforce also says an agent router can use Koa.
That means customers don’t have to treat Koa as a separate AI platform. It becomes another model choice within Salesforce’s existing Agentforce setup.
Salesforce Is Still Keeping Model Choice Open
There is another important detail in the timing of Koa. Salesforce is not moving toward a model where every Agentforce task has to run on a Salesforce-owned model.
The company continues to work with external models, including Anthropic’s Claude through Claudeforce, while adding Koa as a model built specifically for CRM reasoning.
That makes Koa another option at the model layer, not a replacement for the others. It also puts more emphasis on a practical question for Agentforce: which model should handle which kind of work?
For CRM-heavy tasks involving Salesforce data, tools and actions, a specialized model such as Koa may have a role. Other workloads may call for a different model.
That model-level flexibility could become an important part of how enterprises design Agentforce applications.
When Will Salesforce Koa Be Available?
Koa is available to select pilot customers now through Agentforce. Salesforce expects general availability in U.S. regions in winter 2026, with an open beta planned shortly afterward.
For Salesforce customers, the more immediate development is that Koa is already becoming part of the model choices available within Agentforce rather than remaining on a research project outside the platform.
Final Thoughts
Koa is still at an early stage, and there is more to learn as Salesforce moves it from pilots toward broader availability.
But its launch is already one of the more interesting developments around Agentforce at Dreamforce 2026, particularly as Salesforce expands the role of specialized models alongside external model choices.
At Synexc, we’re closely following that evolution across Salesforce, Agentforce, integrations and AI automation.
Have a Salesforce or Agentforce question you’re working through? Talk to Synexc!!
Frequently Asked Questions
Q1. What is Salesforce Koa?
Koa is Salesforce’s first CRM reasoning model, built specifically for the reasoning, tool use and multi-step actions involved in CRM workflows. It is designed to work within Agentforce rather than being a standalone CRM product.
Q2. How is Koa different from a general-purpose AI model?
Koa is specialized for CRM work. Salesforce trained it on synthetic CRM scenarios covering tasks, tool calls and decision-making across different industries, rather than optimizing it primarily for broad general-purpose reasoning.
Q3. Can Agentforce use Koa and other AI models together?
Yes. Koa is an additional model option within Agentforce, while Salesforce continues to support external models such as Anthropic’s Claude. This gives organizations the ability to use different models for different Agentforce workloads.