Salesforce has launched Koa, a specialized AI model designed to improve reasoning over customer relationship management data by leveraging synthetic data and advanced training methods, aiming to automate intricate business workflows within its Agentforce platform.
- Koa uses synthetic CRM data to avoid handling real customer info
- Achieves high accuracy in complex, multi-step CRM tasks
- Available now in pilot; full launch planned for winter 2026
What happened
Salesforce announced the release of Koa at its 2026 Dreamforce conference, a new domain-specific AI model co-developed with Nvidia. Koa is designed to interpret and process customer relationship management data more effectively by reasoning through complex workflows, enabling automation of tasks traditionally requiring human expertise. The model builds on Nvidia’s Nemotron architecture but is further trained with synthetic data representing nearly three decades of Salesforce’s CRM experiences across various industries.
To maintain data privacy, Salesforce avoided using actual customer data for Koa’s training. Instead, they created a large synthetic dataset simulating realistic enterprise workflows in sectors such as healthcare, finance, and manufacturing. This approach allowed the company to fine-tune the model for CRM-specific applications while ensuring data security. Koa is being deployed initially within Salesforce’s Agentforce platform and is currently available to select customers in a pilot program.
Why it matters
Koa addresses a critical challenge in enterprise AI deployment: automating long, multi-step processes that are often complex and require domain expertise. By enhancing AI reasoning capabilities over CRM data, Koa can help companies improve operational efficiency in managing customer interactions, sales pipelines, and support requests. The model’s ability to execute workflows with fewer errors than leading general-purpose AI models signals a significant advance in verticalized AI applications.
The use of synthetic training data also sets an important precedent for preserving customer privacy while developing powerful AI tools. As organizations become more cautious about data use and compliance, Koa’s development exemplifies how AI specialists can leverage proprietary knowledge without exposing sensitive information. Furthermore, Salesforce’s plan to integrate Nvidia’s Nemotron models within the Missionforce platform highlights a commitment to enabling AI in highly regulated industries with strict data security requirements.
What to watch next
Koa is currently in an expanded pilot stage within the Agentforce platform, with general availability expected in winter 2026. Industry observers should track how broadly Salesforce can scale adoption of Koa and how effectively it automates multi-step CRM tasks in live customer environments. Feedback from pilot users, particularly in complex domains like accounting and sales operations, will be crucial for refining the model’s capabilities and proving its business impact.
Additionally, the integration of Nvidia’s Nemotron models into Salesforce’s Missionforce platform for secure, regulated environments will be an important development to monitor. The October launch of these models aims to support mission-critical workflows in areas such as government and supply chain management. How well these AI models perform on dedicated, private infrastructure without compromising security will influence wider acceptance of AI-driven automation in sensitive industries.