Salesforce is expanding its enterprise AI capabilities with Koa, its first CRM reasoning model for Agentforce. Built on NVIDIA Nemotron, Koa is designed for complex CRM work.

Many CRM tasks require more than one action. An AI agent may need to understand customer context, choose the right action, use a tool, update a record, and continue the process. Koa is designed to handle this type of multistep work.

Unlike a general-purpose AI model, Koa focuses on customer relationship management and enterprise workflows. Salesforce says the model uses 27 years of CRM intelligence and was post-trained with NVIDIA Nemotron 3 Super.

For businesses exploring Salesforce AI, Agentforce, and CRM automation, Koa shows how AI models are becoming more specialized. Instead of focusing only on general conversations, these models are being trained to support specific business processes.

Meet Koa

Koa was developed through collaboration between Salesforce and NVIDIA. Salesforce created a synthetic training dataset based on real-world enterprise processes, workflows, and business policies.

According to Salesforce, customer data was not used to train the Koa reasoning model. Instead, the company used synthetic scenarios to teach the model how CRM processes work.

The goal is to help Agentforce agents complete complex tasks more effectively.

For example, an AI agent may need to understand a customer request first. Next, it may retrieve information from Salesforce. Then, it may select an action, update a record, and continue the workflow.

Koa is designed to support this structured process.

This specialization is important because CRM AI needs more than good language skills. It must also understand how sales, service, and customer processes move from one step to another.

Built for Complex CRM Work

Koa is trained using scenarios that reflect real CRM workflows. These scenarios focus on reasoning, tool use, and decision-making.

Salesforce says the training scenarios cover more than 14 industries, including financial services, healthcare, manufacturing, and travel.

Typical CRM tasks include:

• Lead generation and qualification: Support lead activities while following defined sales processes.
• Opportunity management: Update opportunities and select the right actions during sales workflows.
• Case routing: Route customer service cases to the right team or process.
• Follow-up scheduling: Plan next steps while keeping customer and account information in context.
• Multistep CRM workflows: Complete processes that require several decisions, tools, and actions.

These capabilities make Koa relevant for organizations building enterprise AI agents inside Salesforce.

Instead of only answering questions, these agents can support real business processes and take actions within CRM workflows.

Built for Agentforce

Koa is designed to work with Salesforce Agentforce.

According to Salesforce, organizations can use Koa as a managed LLM. They can also select it as an Agentforce model provider. In addition, businesses can use it at the agent or sub-agent level.

Salesforce has also shared results from its CRM Bench evaluations.

According to the company, Koa provides:

• 11% greater precision when selecting the right action
• 2.1 times higher reliability when recalling customer context
• 15% better context retention during longer conversations

Salesforce also says Koa can match or exceed leading models on CRM actions while producing fewer errors.

However, these results come from Salesforce’s own testing. Therefore, businesses should still test Koa against their own workflows, data, and business requirements before wider adoption.

Trust and Enterprise Control

Security and data control are important when businesses use AI inside CRM systems.

Salesforce says it controls Koa’s model weights. The company also states that post-training and inference run within Salesforce infrastructure.

As a result, customer data remains inside the Salesforce trust boundary during inference.

This can be important for businesses evaluating AI for Salesforce CRM, especially when security, governance, and data privacy are key concerns.

Koa is currently available to select Agentforce pilot customers. Salesforce expects general availability in U.S. regions in Winter 2026.

For companies planning an Agentforce strategy, Koa represents an important development in Salesforce AI. It shows a move toward purpose-built reasoning models that understand CRM workflows.

The goal is not simply to add AI to Salesforce. Instead, businesses can use models that are designed to understand tasks, follow processes, and support real CRM work.