Salesforce has never trained its own reasoning model. Until now. At Dreamforce, the company announced Koa, a reasoning model built on Nvidia’s open-weight Nemotron architecture and post-trained specifically for sales, marketing, and customer support tasks. It’s a direct shot at the business model of every frontier AI lab selling API access to enterprises.
The tension here is real. Labs like OpenAI and Anthropic want enterprises to feed their proprietary models with files, prompts, and feedback, essentially becoming dependent on closed systems that cost millions at scale. Salesforce is betting there’s a better path, one where the model is purpose-built, data-clean, and cheaper to run. Koa is designed to do exactly that.
The model will sit inside Agentforce, Salesforce’s platform for building AI agents that handle repetitive tasks like booking appointments or fielding customer service calls. Before Koa, any agent that needed multi-step reasoning would get routed to Claude or ChatGPT through Agentforce’s AI gateway. That routing now has a new option, one Salesforce actually controls.
What makes Koa different from a generic open-weight model comes down to a few specific choices:
- It was post-trained on synthetic data, not real customer records, so there’s no risk of data leakage between clients
- It uses fewer tokens to complete the same tasks, which directly reduces inference costs
- It runs within Salesforce’s security and compliance infrastructure by default
- It can be automatically routed through Agentforce’s AI gateway based on task complexity
- It’s built on Nemotron, which Salesforce describes as the first sovereign, state-of-the-art American pre-trained open-weight model with clear data provenance
That last point matters more than it might seem. Jayesh Govindarajan, EVP of Salesforce AI, told TechCrunch that the team previously had no suitable base model to build from. Chinese open-weight models like Alibaba’s Qwen were off the table because of data provenance concerns. Nemotron changed that calculation.
The synthetic training data is worth paying attention to. Salesforce and Nvidia simulated entire customer service environments, including difficult callers and sales scenarios, without touching any actual customer information. It’s a model trained on the shape of enterprise work rather than the content of it.
So does this mean Salesforce is walking away from Anthropic or OpenAI? Not quite. The company also announced ClaudeForce, a partnership that lets customers use Claude as their primary AI interface while keeping their data inside Salesforce’s infrastructure. Koa is an addition to the menu, not a replacement for it.
But the direction is clear. As enterprises get more sophisticated about AI costs and data control, the value of a purpose-built model that doesn’t require sending sensitive business data to a third-party lab becomes harder to ignore. Frontier labs should be paying attention.
