Salesforce and Nvidia launch Koa, a specialized reasoning model for enterprise CRM
Salesforce and Nvidia have unveiled Koa, a specialized artificial intelligence reasoning model designed to handle complex, multi-step customer relationship management (CRM) tasks. Announced at Salesforce’s annual Dreamforce conference by Salesforce CEO Marc Benioff and Nvidia CEO Jensen Huang, the model is built directly into Salesforce’s Agentforce platform.
Koa represents Salesforce’s first proprietary reasoning model, offering enterprise clients a dedicated, workflow-specific alternative to standard general-purpose frontier models.
Built on Nvidia Nemotron and 27 Years of Enterprise Intelligence
Koa was developed by post-training Nvidia’s open-weight Nemotron 3 Super foundation architecture using supervised fine-tuning and reinforcement learning tools from Nvidia's NeMo software stack. Rather than scraping public web data or training on sensitive customer databases, Salesforce trained the model on a proprietary synthetic dataset derived from 27 years of accumulated CRM expertise.
The training corpus simulates multi-step business scenarios across 14 major industries, including healthcare, financial services, manufacturing, and travel, pairing specific employee and customer personas with explicit tool calls required to complete complex workflows.
Key operational features of Koa include:
Multistep Workflow Execution: Optimized to autonomously qualify sales leads, update pipeline opportunities, route support tickets, and coordinate tasks across disparate backend software platforms.
3x Reduction in Action Errors: Internal benchmark testing indicates Koa matches or outperforms leading general-purpose LLMs on CRM-specific actions while producing three times fewer execution errors on tasks like case routing and scheduling.
Token-Efficient Inference: Engineered leveraging Nemotron’s architecture to reduce computational overhead and token consumption during multi-turn agentic reasoning, lowering operational costs for high-volume enterprise deployments.
Complete Data Governance and On-Premise Execution
A central pillar of the Koa launch is strict data governance. Because Salesforce controls the model weights and hosts post-training and inference entirely within its own secure infrastructure, customer data remains completely within the company's trust boundary. The architecture ensures sensitive corporate records are never transmitted to outside AI vendors.
Furthermore, Salesforce and Nvidia announced an expansion of the technology into Missionforce, equipping government agencies and heavily regulated industries with post-trained Nemotron models that can run in private clouds and air-gapped environments.
Deployment and Pilot Adoption
Koa is currently deployed internally across Salesforce, including within an automated Slack agent, and is undergoing customer pilots with high-profile organizations such as Formula 1, Xero, UChicago Medicine, 1-800Accountant, and Engine.
While Salesforce continues to support third-party models from partners like Anthropic and OpenAI through Agentforce, Koa provides a cost-effective, specialized base layer for routine enterprise workflows. General availability for Koa in U.S. regions is scheduled for winter 2026.
