Salesforce announced Koa at its annual Dreamforce conference on September 15, marking the company's first reasoning model, built on Nvidia Corp.'s open-weight Nemotron architecture and post-trained jointly by the two companies for sales, marketing, and customer-support tasks1,2,3.
Koa is designed to help AI agents reason through complex, multistep work inside Agentforce, the platform Salesforce customers use to build agents that handle tasks such as answering customer service questions or scheduling appointments. Before Koa, long-running or multistep tasks in Agentforce were routed to frontier models like Claude or ChatGPT through Salesforce's AI gateway.
How Koa was trained
Salesforce and Nvidia Corp. did not use any actual Salesforce customer data to post-train the model. Instead, the two companies crafted synthetic data that mimicked customers' patterns. In one example, they simulated a customer service environment with irate customers and a sales professional.
Salesforce says Koa has not ingested any actual customer data and therefore cannot leak it to others. The model follows all of a customer's data requirements and security embedded within Salesforce.
Positioning against frontier labs
Salesforce is offering Koa as an open-weight alternative to closed frontier models, trained for specific work tasks rather than general-purpose benchmarks. The company says Koa helps reduce AI spending because it uses fewer tokens to do the same work. Koa can be automatically routed through an AI gateway depending on the need, sitting alongside other models already available in Agentforce.
Separately, Salesforce announced a partnership with Anthropic called ClaudeForce, which allows companies to use Claude as their AI interface while their data remains in Salesforce's system of records, secured by its infrastructure.
ANALYSIS The dual announcement positions Salesforce as both a model builder and a model-agnostic platform: Koa gives it a first-party reasoning option for CRM-specific workloads, while ClaudeForce preserves access to Anthropic's frontier capability under Salesforce's data governance layer. The gateway routing mechanism means customers need not choose one model exclusively.
Training exclusively on synthetic data rather than live customer records differentiates Koa's provenance from models post-trained on user-generated content, a point Salesforce emphasizes alongside its claim that Koa cannot leak customer data.