Salesforce and NVIDIA have announced Koa, Salesforce's first reasoning model designed specifically for customer relationship management tasks. The model will be integrated into Agentforce and is intended to help enterprise agents handle multi-step workflows, select tools, and act using business context held inside the platform.

The announcement was published on September 15. Salesforce says Koa is built on NVIDIA's Nemotron model family and trained with knowledge associated with 27 years of Salesforce CRM processes. The company did not disclose the model's size, full training-data details, or a technical report with reproducible results. Koa is currently available to selected pilot customers, with general availability expected in winter 2026 in US regions.

The stated goal is not a general conversational assistant. It is a model focused on decisions inside defined business processes, including generating and qualifying sales opportunities, resolving service cases, and coordinating actions across data and tools. In such workflows, an agent must understand an objective, retrieve authorized information, choose an action, and preserve context across several steps. Koa is intended to provide specialized reasoning for that sequence.

Salesforce says it already uses the model internally, including in a Slack agent that helps employees find information and complete routine tasks. It also named 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero as pilot participants. Those pilots show a limited operational deployment, but they do not amount to broad availability or independently validate performance across those environments.

The collaboration also brings NVIDIA models and accelerated computing to Missionforce, Salesforce's offering for government and regulated organizations. The companies describe support for private clouds, air-gapped networks, and other environments where organizations require tighter control of models, data, and infrastructure. Missionforce Operations is generally available in US regions, while post-trained NVIDIA models are expected for selected customers in October 2026.

Koa points to a broader enterprise-model strategy: rather than relying only on a general model with added instructions, Salesforce is using post-training and domain context for a defined set of processes. That specialization could reduce interpretation errors and improve tool selection, but those benefits remain vendor claims until independent evaluations or comparable metrics are available.

Governance questions also remain. A CRM agent may access customer data, change records, or initiate actions with commercial consequences. Reasoning accuracy is only one part of the problem; permissions, tool-call audit trails, reversible actions, and human oversight remain essential controls. The announcement did not provide enough detail to compare Koa with other models on those dimensions.

The clearest signal is that Koa has entered named pilots and has a release window. Public evidence is not yet sufficient to conclude that it outperforms general or competing specialized models. The real test will come when Salesforce publishes fuller evaluations and customers can measure task completion together with error rates, cost, latency, and required human intervention.

— Nova Rivera