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

Salesforce Koa Debuts as Enterprise-Focused Alternative to Frontier AI Models

Salesforce introduced Koa at its Dreamforce conference, marking the company's entry into AI reasoning models1

1

. Built on Nvidia Nemotron 3 Super architecture, the specialized AI model targets sales, marketing, and customer support tasks within the Agentforce platform2

2

. The announcement signals a strategic shift toward open-weight AI models that prioritize enterprise needs over general-purpose capabilities. Salesforce's first CRM reasoning model emerged from a joint technical effort with Nvidia, leveraging nearly three decades of CRM deployment experience3

3

.

Training Approach Emphasizes Data Privacy and Synthetic Datasets

Koa's development relied entirely on synthetic datasets rather than actual customer information, addressing growing enterprise concerns about data privacy1

1

. Salesforce simulated realistic enterprise workflows across 14 industries, including healthcare, finance, and manufacturing, to create training scenarios3

3

. The company used supervised fine-tuning combined with reinforcement learning techniques, specifically employing group relative policy optimization to teach multi-step tool execution3

3

. "Not a single byte of customer data was used," emphasized Rohan Kumar, Salesforce's president of platform and engineering2

2

. This approach ensures the open-weight AI model cannot leak sensitive information to other users, a critical advantage over frontier AI models that require data uploads for training1

1

.

Performance Benchmarks Show Significant Error Reduction

On Salesforce's CRM benchmark covering real-world tasks like updating opportunities, routing cases, and scheduling follow-ups, Koa matches or exceeds leading model performance with 3x fewer errors2

2

4

4

. The specialized AI model demonstrates particular strength in multi-step enterprise workflows that traditionally required human expertise3

3

. Token efficiency represents another competitive advantage, with Nvidia's architecture enabling reduced computational costs compared to sending identical tasks to Claude or ChatGPT1

1

. Kari Ann Briski, Nvidia's VP of Generative AI Software for Enterprise, described this as "the trifecta of things that you need to have: sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all"1

1

.

Jensen Huang Reinforces Open Model Philosophy at Dreamforce

Nvidia CEO Jensen Huang appeared alongside Marc Benioff at Dreamforce, declaring that "with artificial intelligence as an infrastructure layer across the planet, we can know everything and do anything"2

2

. Huang addressed AI safety concerns directly, stating "safety is an engineering problem" and rejecting calls for regulatory slowdowns5

5

. He noted that open models have grown from 30 percent adoption at the beginning of last year to approximately 70 percent currently2

2

. This shift reflects enterprise preference for solutions offering greater control and transparency compared to proprietary frontier AI models5

5

.

Enterprise Customers Gain Control and Cost Advantages

Salesforce retains complete control of model weights, conducting post-training and inference within its own trust boundary4

4

. The CRM reasoning model runs entirely on Salesforce infrastructure, ensuring customer data never leaves controlled environments3

3

. Koa integrates with Agentforce's AI gateway, which automatically routes requests to appropriate models based on task requirements1

1

. Early pilot customer 1-800Accountant highlighted practical benefits: "Koa gives our agents the reasoning to work through that complexity step by step and use the right tools along the way," said Chief Strategy Officer Ryan Teeples3

3

. The specialized AI model helps reduce AI spending through token efficiency while maintaining security protocols embedded within Salesforce1

1

.

Rollout Strategy Balances Innovation with Frontier Model Partnerships

Koa is currently running inside Salesforce, powering an employee agent in Slack, with customer pilots expanding in October 20262

2

. Pilot participants include Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine, and Xero, with general availability expected Winter 2026 in U.S. regions2

2

. Despite launching its own AI reasoning model, Salesforce maintains strategic partnerships with Anthropic through ClaudeForce, allowing companies to use Claude as their AI interface while data remains secured in Salesforce systems1

1

. Additionally, Nvidia Nemotron models will integrate into Missionforce Operations for high-security government and regulated sector deployments, enabling air-gapped AI operations in private clouds3

3

. This dual approach suggests enterprises will increasingly evaluate task-specific models against general-purpose alternatives, potentially reshaping how organizations allocate AI budgets and architect agent workflows.