Salesforce and NVIDIA Unveil Koa: An AI Reasoning Model Built for Customer Relationship Management (CRM)

What happened

According to the NVIDIA blog, Salesforce and NVIDIA announced Koa at Salesforce's Dreamforce conference on September 15, 2026, a CRM reasoning model for Salesforce's Agentforce. CRM stands for customer relationship management, the systems businesses use to manage customer data, sales and service; a reasoning model is an AI model that works through steps before answering or taking action. NVIDIA says Koa is Salesforce's first CRM reasoning model and is built on NVIDIA Nemotron 3 Super.

NVIDIA explains that Koa was created by post-training Nemotron 3 Super on a proprietary synthetic dataset, meaning the existing model was trained further on specific data. Synthetic data is computer-generated rather than taken directly from real records; NVIDIA says this dataset draws on nearly thirty years of enterprise CRM deployments. NVIDIA also says training used two methods, supervised fine-tuning and reinforcement learning, with tools including NeMo RL, NeMo Gym and NeMo AutoModel, and covered synthetic enterprise scenarios across 14 or more industries, including manufacturing, financial services, healthcare and travel.

Data handling and performance: NVIDIA's claims

NVIDIA says that because Nemotron is an open model, Salesforce can fine-tune it (adapt it to its own use) and run it inside its own infrastructure. Salesforce controls the weights (the parameters that determine model behavior), the model runs on Salesforce's systems, and no customer data was used during training or inference (the stage where the model actually runs and produces results). As relayed in the NVIDIA blog, Rohan Kumar, Salesforce's president of platform and engineering, said: "Not a single byte of customer data was used."

NVIDIA says that on Salesforce's CRM Bench, Koa matches or exceeds leading models on CRM actions, with "3x fewer errors." CRM Bench is a benchmark for comparing model performance, covering real work tasks such as updating opportunities, assigning cases and scheduling follow-ups. This is a result as described in the NVIDIA blog; the material this article relies on contains no independent third-party verification.

How readers should weigh these claims

The performance results come from Salesforce's own CRM Bench and are reported on the NVIDIA blog; the material this article relies on contains no independent evaluation results.

For companies that use Salesforce, how data is handled is the key question. NVIDIA says the model runs inside Salesforce's own systems and was not trained on customer data; for details, defer to the official statements from Salesforce and NVIDIA.

What Jensen Huang said on stage

Per the NVIDIA blog, NVIDIA founder and CEO Jensen Huang, speaking about AI safety, said: "Safety is paramount in a lot of ways. It's job one." He said safety is an engineering problem, and that if you aren't confident in a product's functionality, capabilities or safety, you shouldn't release it. He also said the share of open models rose from 30% at the start of last year to about 70% now, and said: "Every company, every enterprise, every country would become an AI company."

NVIDIA also says it uses Salesforce itself across sales, service, marketing and operations, and is currently piloting Agentforce in its customer support workflows.

Browse the latest news in this topic

• Tech news