TL;DR — Key Takeaways
- Salesforce’s Koa is a specialized reasoning model built on NVIDIA’s open-weight Nemotron architecture for CRM and Agentforce workflows.
- Salesforce trained Koa using synthetic data derived from decades of CRM expertise rather than customer information.
- Salesforce says Koa can match or outperform leading general-purpose models on CRM tasks while generating fewer errors.
Salesforce has unveiled Koa, a reasoning model built with NVIDIA technology to handle CRM workflows, giving enterprises a specialized alternative to the general-purpose AI models commonly used to power AI agents.
Announced at Salesforce’s Dreamforce conference, Koa is based on NVIDIA’s open-weight Nemotron architecture and will operate within Salesforce’s Agentforce platform. The model is built to reason through multistep business processes like qualifying sales leads, updating opportunities, routing service cases and coordinating tasks across multiple systems.
Koa marks an expansion of Salesforce’s AI strategy. The company has previously built smaller models for specific Agentforce functions, but has depended on outside frontier models when an agent needed more advanced reasoning.
Salesforce and NVIDIA trained Koa for CRM work without using Salesforce customer information. Instead, Salesforce created synthetic data based on knowledge accumulated over 27 years of CRM deployments. The training scenarios reproduce business processes across 14 industries, including healthcare, financial services and manufacturing.
The synthetic environments include different employee and customer personas and require the model to determine which actions and software tools are needed to complete a task.
Salesforce touts this specialized approach as providing a significant performance benefit. The company said its benchmark testing showed Koa matching or surpassing several leading general-purpose models on CRM tasks while producing three times fewer errors.
“For a SaaS company, renting the model means renting the most valuable part of the product. Owning the weights gives Salesforce control of inference cost, the data path, and how fast it can tune the model to its own workflows,” said Mitch Ashley, VP and Practice Lead at The Futurum Group.
“The differentiating asset in Koa is Salesforce’s own record of how CRM workflows run, encoded as training data. NVIDIA supplies the open base model. That record is the part that competitors cannot copy. Expect this to become standard.”
Additionally, he noted, “Open base models are good enough that any vendor with decades of domain data will post-train its own, and buyers should ask every SaaS vendor who owns the model underlying their agents.”
The Value of Specialized Models
Complex reasoning models can consume large numbers of tokens as they work through a problem, driving up inference expenses. NVIDIA designed Nemotron for greater token efficiency, potentially allowing Koa to perform CRM reasoning with lower compute costs than larger general-purpose models.
On the security front, Salesforce controls Koa’s model weights and runs the system inside its own infrastructure. So enterprise data remains within Salesforce’s security environment when Agentforce uses the model.
This combination of specialization, cost control and data governance reflects a major enterprise shift. The largest frontier models remain valuable for broad knowledge and advanced reasoning, but businesses do not necessarily need the largest available model for every task. A model trained around a defined set of workflows can offer a better fit for applications where accuracy and data control matter as much as raw model capability.
Still, Salesforce continues to work with outside AI providers. Agentforce can route workloads among different models, and Salesforce recently announced Claudeforce, an alliance with Anthropic that enables enterprises to use Claude while keeping their business data within Salesforce systems.
Koa is currently available to selected customers through an expanded Agentforce pilot, with general availability planned for winter.
