Salesforce Koa AI Reasoning Model Built on Nvidia Nemotron Challenges Frontier Labs at Dreamforce

Salesforce Koa Redefines Enterprise AI with CRM Reasoning Model

Salesforce and Nvidia partnership delivered Koa at the Dreamforce conference, marking a pivotal shift in how enterprises approach AI reasoning model deployment1

1

. Built on Nvidia Nemotron 3 Super, Salesforce Koa represents the company's first CRM reasoning model specifically designed for the Agentforce platform2

2

. Unlike general-purpose frontier models, this specialized AI model tackles complex multistep workflows across sales, marketing, and customer support without touching actual customer data3

3

.

Jensen Huang joined Salesforce CEO Marc Benioff onstage at Moscone Center, declaring that "with artificial intelligence as an infrastructure layer across the planet, we can know everything and do anything"2

2

. The announcement signals how enterprise needs are diverging sharply from what proprietary AI labs offer. Jayesh Govindarajan, EVP of Salesforce AI, explained that reasoning capabilities previously required routing prompts to frontier models like Claude or ChatGPT through Agentforce's AI gateway1

1

.

Synthetic Dataset Powers AI Model for CRM Without Customer Data Security Risks

Salesforce created Koa through post-training Nvidia Nemotron using a proprietary synthetic dataset replicating nearly three decades of CRM deployment knowledge across 14+ industries including manufacturing, financial services, healthcare, and travel2

2

. Rohan Kumar, Salesforce's Chief Platform and Engineering Officer, emphasized that "not a single byte of customer data was used" during training or inference2

2

. The company simulated realistic enterprise scenarios with personas ranging from irate customers calling support centers to sales professionals closing deals1

1

.

The training methodology combined supervised fine-tuning with reinforcement learning using NeMo RL, NeMo Gym, and NeMo AutoModel technologies2

2

. Kumar detailed how the team applied group relative policy optimization to teach Koa how to execute multistep tool calls with precision3

3

. This approach allowed Salesforce to leverage its accumulated CRM intelligence while maintaining absolute customer data security within its trust boundary4

4

.

CRM Benchmark Results Show 3x Fewer Errors on Enterprise Tasks

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

2

. The specialized AI model demonstrates particular strength in automating long, multistep processes that typically require human expertise, including qualifying sales leads and executing multisystem workflows3

3

.

Kari Ann Briski, Nvidia's VP of Generative AI Software for Enterprise, highlighted the "unique architecture for inference to be token efficient," describing it 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

. This efficiency translates directly to reduced AI spending for enterprises, as Koa uses fewer tokens to accomplish the same work compared to routing tasks to Claude or ChatGPT1

1

.

Open-Weight Model Architecture Enables Enterprise Control and Missionforce Expansion

The open-weight model foundation proved critical to Salesforce's strategy. Govindarajan noted that before Nvidia Nemotron, "there was no sovereign American pre-trained model that was available" with both state-of-the-art capabilities and clear data provenance1

1

. Huang observed that open models surged from 30 percent adoption at the beginning of last year to approximately 70 percent currently2

2

.

Because Nvidia Nemotron is open, Salesforce fine-tuned and runs Koa entirely within its own infrastructure. Salesforce controls the weights, and the model operates on Salesforce's systems with no customer data exposure during training or inference2

2

. The companies are extending their collaboration to Missionforce, targeting public-sector agencies and regulated industries handling sensitive data5

5

. Customers will train, fine-tune, and deploy specialized models on their own infrastructure, including air-gapped networks and private clouds for procurement, vendor management, and logistics automation3

3

.

Early Pilots and General Availability Timeline for Agentforce Platform

Koa is already running inside Salesforce, powering an employee agent in Slack, and moving into customer pilots in October as a customer-selectable model in Agentforce2

2

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

2

. Ryan Teeples, chief strategy officer at 1-800Accountant, noted that "Koa gives our agents the reasoning to work through that complexity step by step and use the right tools along the way"3

3

.

Salesforce isn't abandoning partnerships with Anthropic or OpenAI entirely. The company simultaneously announced ClaudeForce, allowing companies to use Claude as their AI interface while data remains secured within Salesforce's infrastructure1

1

. Nvidia itself runs on Salesforce across sales, service, marketing, and operations, currently piloting Agentforce for customer-support workflows2

2

. Nemotron models will roll out for Missionforce Operations in October, enabling mission-specific intelligence for regulated sectors3

3

.