Salesforce Announced Koa Reasoning Model for CRM Tasks
The new model, built on NVIDIA's Nemotron 3, aims to reduce CRM errors by 66% in multi-step workflows by early 2026.
Updated on Sept. 23, 2026 in Artificial Intelligence
Live Poll
Do you trust AI-driven automation to handle complex tasks in your professional dealings?
Salesforce has announced Koa, a reasoning model designed to execute complex customer relationship management actions. The model is currently used for Slack tasks and is scheduled for general availability in Winter 2026.
Why it matters
Koa was developed to improve how agents handle multi-step workflows by leveraging synthetic data derived from 27 years of CRM deployments. This approach aims to minimize hallucination and operational errors in enterprise automation.
Koa demonstrated a 3x reduction in errors compared to leading alternatives during CRM tasks. The model utilizes synthetic datasets spanning 14 industries, trained without using customer data.
The players
Salesforce
An enterprise software company that provides a cloud-based stack for customer relationship management and business process automation.
NVIDIA
A semiconductor company that designs high-performance computing hardware and AI software platforms for enterprise applications.
The details
Salesforce utilized Supervised Fine-Tuning, a process of training models on labeled datasets to improve accuracy, alongside Group Relative Policy Optimization, a reinforcement learning method that adjusts model behavior based on comparative policy scoring. The model is built on NVIDIA's Nemotron 3 Super architecture and performs all inference within Salesforce trust boundaries to maintain data privacy. These technical measures enable the model to simulate workflows across 14 industries using synthetic scenarios.
Timeline
September 16, 2026: Salesforce and NVIDIA announced the Koa reasoning model.
October 2026: Post-trained NVIDIA models reach select Missionforce customers.
Winter 2026: Koa enters general availability for U.S. regions.
The Tech Race
Koa builds upon the NVIDIA Nemotron 3 Super model to address the specific performance bottlenecks in CRM automation. It follows a growing industry trend of using synthetic data to improve model reasoning while maintaining secure enterprise boundaries.
The technology is currently integrated into Slack tasks, with broader availability for U.S.-based organizations expected by Winter 2026. Users and pilot customers such as 1-800Accountant, BCU, and Xero will be the first to implement these workflows for CRM automation.
The takeaway
Salesforce aims to improve the reliability of autonomous CRM agents by training them on synthetic data rather than raw customer inputs. Industry observers should watch for the general release in Winter 2026 to see if the 3x error reduction holds true in diverse, real-world deployments.
What happens next
Select Missionforce customers will gain access to new post-trained NVIDIA models in October 2026, followed by the general release of Koa in Winter 2026.
Further reading
For more on how new models are changing enterprise software, visit the Artificial Intelligence section.
Live Poll
Do you trust AI-driven automation to handle complex tasks in your professional dealings?
