Salesforce has launched Koa, a reasoning model for customer-relationship management work, built by post-training an Nvidia Nemotron open model on synthetic business-workflow scenarios. The product gives Salesforce a model it hosts and controls for its Agentforce software, as large enterprise-software suppliers look for ways to make AI agents more reliable on narrow, repeatable tasks.
The company announced Koa on September 15 at Dreamforce in San Francisco. Salesforce says the model is already used inside the company and is moving into pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.
Koa is not a general-purpose chatbot that a business trains from scratch. Salesforce says it took Nvidia’s Nemotron 3 Super model and adapted it using a proprietary set of synthetic scenarios based on how sales, customer-service and other CRM workflows are carried out. A CRM system stores and organizes a company’s interactions with customers, prospects and service cases.
Why Salesforce built a CRM-specific model
A language model can draft an email or summarize a call, but many CRM jobs require a sequence of decisions and actions. A service agent may need to check a customer’s record, identify the relevant policy, update a case and route it to the correct team. Salesforce says Koa was tuned for work such as generating leads, qualifying sales opportunities and resolving service cases.
The commercial appeal is control over that sequence. Salesforce says it runs Koa within its own infrastructure and retains control of the model weights, the numerical parameters that determine how a model responds. The company also says it did not use customer data to train Koa. Instead, the training material simulated roles, tasks and tool calls across more than 14 industries.
That approach limits the story to Salesforce’s own claims for now. The company says Koa made three times fewer errors than leading models in its CRM Bench tests, which include tasks such as updating an opportunity, routing a case and scheduling a follow-up. The benchmark is Salesforce’s, so customers will need to judge performance in their own data, permissions and workflows.
The announcement sits alongside a wider shift in enterprise software. As we reported earlier this month, suppliers are moving beyond a simple per-employee license when agents can complete work on a customer’s behalf. An agent that carries out thousands of actions can create a different cost profile from a conventional software user, even if one employee supervises both.
Open models give vendors a starting point
Salesforce is using an open-weight Nvidia model as a starting point, then adding its own training, testing and operating controls. Open-weight means the model’s parameters are available for others to run and adapt under its license. It does not mean a customer’s records are automatically shared with the original model developer.
The structure resembles the route taken by other information businesses that want a model tailored to their own data and products. Our August coverage of Thomson Reuters’ proprietary model described a legal and tax publisher using an open-weight foundation while keeping control of its specialized system and professional content.
For Salesforce, the difference lies in the product context. The company has spent years collecting the rules, fields and workflow patterns that organizations use in its CRM products. Koa is intended to encode some of that operational structure into a model that can select a tool or next action, rather than only generate text about the task.
Availability and the test for customers
Salesforce says Koa is available now to selected pilot customers and expects general availability in US regions in winter 2026, with an open beta planned shortly afterward. The model can be selected in Agentforce, Salesforce’s platform for creating agents that can use business data and software tools.
The company lists Koa as a managed model option and says administrators can choose it across an organization or for individual agents. Those choices matter because an agent can only be as dependable as the records it can access, the policies it is given and the controls around its actions.
Salesforce’s data-security claims also need to be separated from the model’s capabilities. Hosting the model inside Salesforce’s infrastructure may address one concern for buyers who want to limit where data travels. It does not remove the need to set permissions, review agent actions and test how an automated workflow behaves when a customer record is incomplete or an exception appears.
Koa gives Salesforce a more direct stake in the model layer of enterprise AI, while Nvidia gains another large software partner building on Nemotron. The immediate question is not whether a CRM model can produce convincing answers. It is whether customers find that it completes routine work accurately enough to justify the added spending and governance.
