Salesforce and NVIDIA Launch Koa as Jensen Huang Casts Enterprise AI Agents as a New Infrastructure Layer

Salesforce and NVIDIA unveiled Koa, Salesforce’s first CRM reasoning model for Agentforce, at Dreamforce in San Francisco. Koa was built by post-training NVIDIA Nemotron 3 Super and is designed to handle multistep sales, marketing, and customer-support work inside Salesforce systems.

The model is trained with a proprietary synthetic dataset modeled on nearly 3 decades of Salesforce CRM deployments. Salesforce said customer data was excluded from training, while the model’s weights, post-training, and inference remain inside Salesforce’s trust boundary. The company positions that arrangement as a way to give enterprises stronger control over sensitive operational data.

Koa is aimed at the practical decisions that CRM agents must make before taking action, including updating opportunities, routing cases, and scheduling follow-ups. Salesforce said Koa matches or exceeds leading models on its CRM benchmark while producing 3 times fewer errors on CRM actions.

That emphasis separates Koa from general-purpose chatbots. The model is tuned around Salesforce’s accumulated knowledge of enterprise processes, while Agentforce supplies the surrounding tools and business context. The result is a system intended to complete work across connected workflows rather than simply generate text. Independent performance comparisons were not provided in the event coverage, so the benchmark claims remain Salesforce’s own evaluation.

NVIDIA CEO Jensen Huang used the Dreamforce stage to frame AI as a new infrastructure layer with reach beyond software assistance. Electricity enabled broad access to power, the internet enabled broad access to information, and artificial intelligence, he said, enables people and organizations to know more and perform more tasks.

Huang described the opportunity for AI agents as “gigantic” and said development had entered a “flywheel zone,” where faster advances reinforce further adoption. Salesforce CEO Marc Benioff joined Huang for the presentation, which combined the Koa launch with a wider argument that enterprise agents will become embedded in core business operations.

Koa’s foundation on NVIDIA’s open-weight Nemotron gives Salesforce a model layer it can adapt to CRM-specific work. That arrangement reflects a growing enterprise preference for control over model behavior, deployment, and data handling, rather than sending every workflow and customer record to a general-purpose external service.

The partnership also extends to Missionforce, Salesforce’s platform for government and regulated organizations. Salesforce and NVIDIA said the collaboration brings NVIDIA open models and accelerated computing to deployments where organizations need tighter control over the model, data, and operating environment. TechCrunch described Koa as evidence that enterprise AI priorities are diverging from the incentives of frontier model companies, especially around customization and data governance.

Koa puts Salesforce’s large installed base at the center of the reasoning-model race. General-purpose labs compete on broad capability, while Salesforce can focus its model on a narrower but commercially important domain: customer records, sales pipelines, service queues, marketing operations, and the actions that connect them.

That strategy creates a different measure of progress. A model that makes fewer mistakes on a defined business workflow can be more valuable to an enterprise than a stronger open-ended conversational model. It also gives Salesforce a way to keep Agentforce differentiated as companies assess whether agents can reliably perform work without constant employee supervision.

Salesforce must now turn Koa’s benchmark claims into dependable customer outcomes across varied CRM environments. Accuracy in a controlled evaluation is only one part of enterprise deployment; permissions, audit trails, exception handling, integration quality, and the cost of running inference will shape adoption.

The Missionforce announcement points to regulated industries as an important test. If Salesforce can offer domain-specific reasoning while keeping data and model operations within customer-controlled boundaries, other enterprise software vendors will face pressure to build or tailor their own models. Dreamforce’s agent focus, including customer examples such as Adecco Group’s effort to integrate agents into core workflows, shows where the contest is heading: from demonstrations onstage to accountable systems performing daily business tasks.