Salesforce and Nvidia Launch Koa, a Purpose-Built CRM Reasoning Model

Salesforce and Nvidia announced Koa on September 15, 2026, Salesforce's first CRM reasoning model for Agentforce, purpose-built to help agents reason through complex, multistep CRM workflows rather than relying on a general-purpose foundation model.

Built on Synthetic Data, Not Customer Data

Koa was built by post-training Nvidia's Nemotron 3 Super with a proprietary synthetic dataset modeled on enterprise knowledge from nearly three decades of CRM deployments. Notably, no customer data was used to train Koa — the training data was built entirely on synthetic scenarios designed to reflect what Agentforce agents would realistically encounter.

Three Times Fewer Errors on CRM-Specific Tasks

In Salesforce's own CRM benchmark, Koa matches or exceeds leading general-purpose model performance on CRM actions with three times fewer errors — a specific, measurable claim for a purpose-built model competing against far larger general foundation models on a narrower task domain.

Real Use Cases and Early Pilots

• Use cases include generating leads, qualifying opportunities, and resolving service cases
• Available now to select pilot customers: 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero
• General availability expected in Winter 2026 in U.S. regions

Extending Into Government and Regulated Sectors

Salesforce and Nvidia are also integrating Nemotron-based models into Missionforce, which targets government and regulated organizations requiring deployment in private clouds and air-gapped networks — extending the same reasoning-model approach into environments where public cloud AI deployment isn't an option.

What This Means for Enterprise CRM Buyers

A purpose-built model outperforming general-purpose alternatives by a specific, measured margin on domain-specific tasks is a concrete signal that the next wave of enterprise AI competition may center on narrow, task-specialized models trained on synthetic domain data — not solely on which vendor has access to the largest general-purpose foundation model.