Salesforce and Nvidia Launch Koa Reasoning Model to Challenge Frontier AI Labs

Salesforce has introduced Koa, its first enterprise-grade CRM reasoning model built on Nvidia's Nemotron 3 Super architecture, during its Dreamforce conference in San Francisco. Developed through a post-training process using synthetic scenarios modeled on nearly three decades of enterprise CRM data, Koa is designed to help software agents execute multi-step business workflows without exposing actual customer data to external training pipelines.

The release addresses growing enterprise anxiety regarding data governance and intellectual property risks associated with closed frontier models. Companies such as Palantir and Nvidia have recently pushed for strict data retention guarantees from major AI labs following policy adjustments on usage logs, prompting organizations to seek secure alternatives that operate entirely within private cloud environments and dedicated infrastructure.

Unlike general-purpose models designed to solve complex mathematical problems, Koa was post-trained specifically for sales, marketing, and customer support operations. Salesforce and Nvidia avoided using actual customer data during training, instead generating synthetic scenarios that simulate enterprise interactions, including handling irate customer service calls and managing sales pipelines across more than 14 industries.

Each simulation mapped the sequence of actions and tool calls an agent must make to resolve customer service inquiries, update opportunities, and schedule follow-up appointments. According to Salesforce, the model achieves three times fewer errors on its internal CRM Benchmark compared to leading general-purpose models, while leveraging Nvidia's token-efficient inference architecture to reduce operating costs for multi-step agentic workflows.

Koa is being integrated into Salesforce's Agentforce platform as a specialized alternative to third-party frontier models. When autonomous agents require advanced reasoning to navigate complex tasks, the system routes requests through an internal AI gateway, allowing enterprises to utilize Koa while maintaining strict adherence to corporate security protocols.

The model has already been deployed internally within a Slack agent designed to assist employees with everyday tasks. Customer pilots are currently underway with organizations including Formula 1, Xero, Baxter Credit Union, and UChicago Medicine, with general availability scheduled for Winter 2026 in U.S. regions.

Despite introducing its own sovereign reasoning model, Salesforce is maintaining its relationships with external AI developers. The company recently announced Claudeforce, a partnership with Anthropic that brings Claude into Salesforce workflows while ensuring enterprise data remains secured within Salesforce's internal system of records.

Salesforce expanded its enterprise footprint at Dreamforce by partnering with Siemens to integrate Agentforce with the Teamcenter Service Lifecycle Management platform. Siemens is currently deploying the platform to automate inbound lead qualification across 132 countries, processing more than 2,500 unqualified leads per month for its 18,000 sellers. The Koa release arrives as Salesforce expands its Agentforce business, with Agentforce ARR exceeding $1.5 billion in the second quarter of fiscal 2027, representing a year-over-year increase of more than 240 percent.