New CRM AI Model
Salesforce and Nvidia Launch Koa Reasoning Model to Power Complex CRM Workflows
Anonymous
Key Points
1
Salesforce said Koa matched or exceeded multiple leading general-purpose models on CRM tasks while producing three times fewer errors, according to the company’s benchmark tests.
2
Koa is derived from NVIDIA’s Nemotron-3-Super-120B model, which has a 1-million-token context window and a Mixture-of-Experts architecture that activates only part of its parameters for each task.
3
Salesforce said Koa had already been deployed internally, including in a Slack-based agent intended to help employees with work tasks.
4
The Salesforce-NVIDIA collaboration behind Koa was announced in March 2026, when the companies began integrating NVIDIA’s Nemotron models into Agentforce.
Salesforce and NVIDIA unveiled Koa, a specialized AI reasoning model built to tackle complex sales and customer-service workflows. The Rundown reports that Koa is based on NVIDIA's open Nemotron-3 model and is now entering customer pilots. The model was trained entirely on synthetic data drawn from three decades of Salesforce CRM deployments, avoiding the use of real customer data.
Newsytes says major companies including UChicago Medicine and Formula 1 are already testing Koa. According to Salesforce benchmarks, the model matches or beats leading general-purpose AI systems on CRM tasks while producing three times fewer errors. Koa gives Salesforce full control over the model's weights and runs on Salesforce's own infrastructure, addressing data security concerns.
Built on NVIDIA's Nemotron Technology
Koa is derived from NVIDIA's Nemotron-3-Super-120B model, CXO Today reports. The base model includes a 1-million-token context window and uses a Mixture-of-Experts design. This architecture activates only the necessary parameters for each task, making the system more efficient. Salesforce fine-tuned Nemotron using supervised learning and reinforcement techniques, specifically Group Relative Policy Optimization.
Trained on Synthetic CRM Scenarios, Not Customer Data
Livemint reports that Koa was trained using synthetic scenarios modeled on nearly 30 years of real-world CRM deployments across healthcare, finance, manufacturing, and travel. Salesforce avoided using actual customer data. This approach protects privacy while allowing the model to learn from diverse industry workflows. The company launched Koa at its Dreamforce conference.
Newsytes notes that Koa has already been deployed internally, including in a Slack-based agent to help Salesforce employees complete work tasks. This internal testing provided real-world validation before the broader pilot rollout to customers.
Handles Complex Multi-Step Workflows
CXO Today explains that Koa is designed to navigate complex, multi-step business processes. The model handles lead generation, sales qualification, marketing tasks, and customer-service cases. It selects the right tool for each step and reasons through workflows that span multiple stages.
Verdict reports that the partnership between Salesforce and NVIDIA began in March 2026, when the companies started integrating Nemotron models into Agentforce. Koa represents the first major output of that collaboration. The model performs competitively with leading systems on CRM and tool-use tasks, though it does not match the strongest frontier models across all capabilities.
Data Security and Hosted Deployment Options
A key advantage of Koa is that Salesforce controls the model's weights and runs all training and inference on its own servers. The Rundown notes this gives customers a Salesforce-hosted option. For organizations with strict data governance requirements, running the model in-house eliminates concerns about data leaving their control. The model's Mixture-of-Experts design also reduces computational overhead compared to using larger general-purpose models.
Sentiment
Mostly PositiveNEGATIVENEUTRALPOSITIVE
Publishers
17
Articles
26
Reach
43
A
AnonymousAuthor
