At its Dreamforce conference, Salesforce unveiled Koa, one of the company’s biggest announcements. Koa is Salesforce’s first reasoning-focused language model, built on Nvidia’s open-weight Nemotron and post-trained by Salesforce and Nvidia to specialize in sales, marketing, and customer-support tasks.
Why this matters for enterprises
Koa highlights how enterprise requirements for AI are diverging from what so-called frontier labs typically offer. Some proprietary AI providers encourage customers to upload files, code, prompts, and feedback directly into their models and agents, often at substantial cost. Koa is positioned as an alternative that aims to meet enterprise priorities for control and data provenance.
Salesforce lists several advantages of Koa for enterprise customers:
- an open-weight alternative to closed frontier models
- a model post-trained on specific work tasks rather than general academic challenges
- no ingestion of actual customer data, reducing the risk of data leakage
- greater token efficiency, which can lower AI spending
- the ability to be routed automatically via an AI “gateway” depending on the request
- adherence to customer data and security requirements embedded within Salesforce
How Koa was developed
The model was created by post-training Nvidia’s Nemotron base model to be proficient in sales and customer support scenarios. Importantly, Salesforce and Nvidia did not use real Salesforce customer data for this post-training. Instead, they generated synthetic datasets designed to simulate customer patterns and interactions.
Jayesh Govindarajan, Executive Vice President of Salesforce AI, explained that the training included simulated customer-service environments with role-played personas ranging from irate callers to sales professionals closing deals. Prior to Koa, long-running or multi-step reasoning tasks were routed from Salesforce’s Agentforce platform to external frontier models such as Claude or ChatGPT via Agentforce’s AI gateway.
Govindarajan also noted that the availability of Nemotron was a turning point: before Nemotron, Salesforce lacked a state-of-the-art, sovereign American pre-trained base model with clear data provenance that could serve as the foundation for an enterprise-grade reasoning model.
Token efficiency and deployment
Token efficiency is a core part of Koa’s appeal. Kari Ann Briski, Nvidia’s Vice President of Generative AI Software for Enterprise, said Nemotron’s architecture enables token-efficient inference, delivering faster time-to-first-token and efficient reasoning—attributes that matter for the tokenomics of real-world deployments.
Koa will be offered as an option within Salesforce’s Agentforce platform, where customers build agents that handle routine tasks such as answering support queries or scheduling appointments. The model is intended to perform those tasks more effectively and at lower token cost than routing to external frontier models.
Continuing support for third-party models
Salesforce is not abandoning support for Anthropic or OpenAI. The company announced a partnership with Anthropic called Claudeforce, which lets customers use Claude as their AI interface while keeping their data inside Salesforce’s system of records and secured by Salesforce infrastructure.
In sum, Koa aims to provide enterprise customers with an open-weight, data-conscious, and token-efficient reasoning model tailored to sales and support workflows, while Salesforce maintains interoperability with external AI providers where customers require it.
