Salesforce developed Koa with NVIDIA using synthetic training scenarios instead of customer records. The dataset was designed around the kinds of actions Agentforce agents perform, including generating leads, qualifying sales opportunities and resolving service cases.
Those simulations covered workflows across more than 14 industries, including manufacturing, financial services, healthcare and travel. Salesforce said individual scenarios combined specific personas, assignments and sequences of tool calls so the model could learn how to work through a task rather than simply generate an answer.
The company said no Salesforce customer data was used to train Koa.
Salesforce applied supervised fine-tuning and reinforcement learning during post-training, using NVIDIA tools including NeMo RL, NeMo Gym and NeMo AutoModel. The model itself is based on NVIDIA Nemotron 3 Super, with Salesforce controlling the weights and running post-training and inference within its own infrastructure.
That arrangement is central to how Salesforce is positioning Koa. Instead of building a general-purpose model intended to cover every type of reasoning problem, the company has concentrated its training on CRM work and the operational steps agents must take inside business processes.
Salesforce says that focus can also improve efficiency. On its CRM Benchmark, which includes tasks such as updating an opportunity, routing a case and scheduling a follow-up, the company said Koa matches or exceeds leading models on CRM actions while producing three times fewer errors.
Salesforce is also presenting Koa as a way to reduce the amount of model usage required for certain tasks. The company says the model can use fewer tokens on the enterprise workflows it was trained to handle than sending the same work to models such as Claude or ChatGPT.
That does not mean Salesforce is moving away from outside model providers altogether. Agentforce is designed to work with multiple models, and Salesforce has continued expanding partnerships with companies including Anthropic. Koa instead adds another option for situations where a Salesforce-trained model may be better suited to the job.
Koa is already being used internally at Salesforce, including through a Slack agent that helps employees locate information and complete everyday tasks. The company is also moving the model into customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.
Salesforce CEO Marc Benioff said the model reflects knowledge accumulated through the company’s years of working with enterprise customers.
“The most valuable thing Salesforce has built isn't our platform — it's the accumulated knowledge of how enterprise business actually works. With Koa, the knowledge is put inside the model itself,” Benioff said.
The Salesforce-NVIDIA collaboration also extends beyond Agentforce. The companies are bringing NVIDIA models and accelerated computing into Missionforce, Salesforce’s platform for government and regulated organizations.
That work is intended for environments where organizations need tighter control over their models, data and infrastructure, including private clouds, classified systems and air-gapped networks. Post-trained NVIDIA models will be used in Missionforce Operations for workflows including procurement, supplier management and logistics.
For Salesforce, Koa adds a specialized reasoning layer between smaller task-specific models and the frontier systems Agentforce has previously relied on for demanding requests. The company is betting that a model trained directly around CRM workflows can handle a meaningful share of those jobs with greater control and lower model usage, while leaving broader frontier models available when they are the better fit.
This analysis is based on reporting from Salesforce.
Image courtesy of Salesforce.
This article was generated with AI assistance and reviewed for accuracy and quality.
