Artificial intelligence is changing fast, and businesses want AI that actually understands how their work gets done, not just AI that sounds smart. That is exactly the gap Salesforce is trying to close with its newest release. On September 15, 2026, Salesforce introduced Salesforce Koa, its first CRM reasoning model built specifically for Agentforce. Developed together with NVIDIA and based on the Nemotron 3 Super model, Koa is designed to help AI agents think through complex, multi-step business tasks instead of just answering simple questions. In this post, we will break down what Koa is, how it was built, why it matters, and where it is headed next.
What Is Salesforce Koa?
Salesforce Koa is a reasoning model made for one specific job: helping Agentforce agents handle real business workflows. Unlike a general-purpose AI model, Koa is trained to understand tasks like updating a sales opportunity, routing a customer service case, or scheduling a follow-up call. These are not simple one-step actions. They usually involve several decisions, tool calls, and pieces of information working together, and Koa is built to reason through that entire process step by step.
Salesforce has spent nearly three decades helping companies manage customer relationships, and that experience has taught the company a lot about how enterprise processes actually work. With Koa, Salesforce is putting that accumulated knowledge directly into the model itself, so agents can reason with a real understanding of how a deal moves forward or how a support case should be resolved.
Built on NVIDIA Nemotron: How Koa Was Trained
Koa was not built from scratch. It is based on NVIDIA’s Nemotron 3 Super model, which Salesforce then post-trained using its own data and methods. Instead of using real customer information, Salesforce built a synthetic training dataset made up of realistic enterprise scenarios. These scenarios cover more than 14 industries, including manufacturing, financial services, healthcare, and travel.
Each scenario in the training data followed a clear pattern. A persona was given a specific task, and the model was trained on the full sequence of actions and tool calls needed to complete that task successfully. This approach helped Koa learn not just what the right answer looks like, but how to actually get there through a logical chain of steps.
To fine-tune the model, Salesforce used a mix of Supervised Fine-Tuning (SFT) and reinforcement learning through a method called Group Relative Policy Optimization (GRPO). This training was carried out using NVIDIA’s NeMo RL, NeMo Gym, and NeMo AutoModel tools. Because the training data was entirely synthetic, no real customer data was ever used to teach the model.
Why Koa Performs Better on CRM Tasks
Because Koa was trained specifically around CRM workflows, it performs differently than a general AI model would. Salesforce tested Koa using its own CRM benchmark, which includes realistic tasks such as updating an opportunity, routing a case, or scheduling a follow-up. According to Salesforce, Koa already matches or outperforms leading models on these CRM actions, while making three times fewer errors.
This kind of accuracy matters a lot in business settings. A small mistake in a sales pipeline or a mishandled support case can create real problems for a company and its customers. By focusing training specifically on enterprise tasks, Koa aims to reduce these errors and make AI agents more dependable for everyday business use.
Keeping Customer Data Safe: The Trust Boundary
One of the most important parts of Koa’s design is how it handles data. Salesforce controls the model weights and runs both training and inference within its own systems. This means customer data never leaves Salesforce’s trust boundary during use. For businesses that rely on Agentforce, this is a meaningful detail, since it shows that performance improvements do not come at the cost of data privacy or security.
Bringing AI to Missionforce and Regulated Industries
Salesforce and NVIDIA are not stopping at standard business use cases. The same Nemotron-based technology is also being brought into Missionforce, a set of tools built for government agencies and other highly regulated organizations. These organizations often need to keep full control over their models and data, sometimes even running systems on private clouds or fully air-gapped networks that never connect to public infrastructure.
As part of this effort, post-trained NVIDIA models will help power Missionforce Operations, a product built to digitize and automate complicated government workflows such as procurement, supplier management, and logistics. These models will be trained using an organization’s own operational data and language, allowing agents to reason through back-office processes securely, even in fully isolated environments.
Early Adopters: Who Is Already Using Koa
Koa is already being tested in real business environments. Inside Salesforce itself, an internal Slack agent powered by Koa is helping employees find information and complete everyday tasks more efficiently. Beyond internal use, Koa has also moved into customer pilot programs with several well-known organizations, including 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
These early adopters come from very different industries, from accounting and banking to healthcare and business travel, which shows how flexible Koa’s reasoning approach is meant to be. Each of these organizations deals with complex, multi-step processes, and early feedback suggests that Koa is helping teams manage that complexity more effectively while freeing up time for higher-value work.
Availability and What Comes Next
For now, Koa is available to a select group of pilot customers through Agentforce, with general availability expected in winter 2026 across U.S. regions. Missionforce Operations, the government-focused product, is already generally available in the U.S., and post-trained NVIDIA models for that platform are expected to become available to select customers starting in October 2026.
This phased rollout suggests that Salesforce wants to test Koa carefully with real customers before making it widely available, which is a reasonable approach for a model meant to handle sensitive business workflows.
Conclusion
Salesforce Koa represents a meaningful step forward in how AI agents handle real enterprise work. By combining NVIDIA’s Nemotron technology with decades of Salesforce’s own CRM knowledge, Koa is built to reason through complex tasks accurately, securely, and specifically for business needs. Its early results, along with a growing list of pilot customers across different industries, point to a model that could reshape how companies use AI within their CRM systems. As Koa moves toward general availability in winter 2026, businesses using Agentforce will have a lot to watch for in the months ahead.
