Salesforce just dropped Koa, its first proprietary reasoning model, at Dreamforce. Built on Nvidia’s open-weight Nemotron architecture, this isn’t another general-purpose chatbot trying to solve abstract math puzzles. It is a specialized engine designed explicitly for the messy, multi-step realities of sales, marketing, and customer support. This marks a sharp pivot from relying entirely on frontier labs like OpenAI or Anthropic for complex logical tasks. Instead of routing every difficult query to an external API, enterprises can now keep that reasoning power in-house.
The collaboration with Nvidia was driven by a specific gap in the market: the lack of a sovereign, state-of-the-art pre-trained base model with clear data provenance. Jayesh Govindarajan, EVP of Salesforce AI, pointed out that prior options like Alibaba’s Qwen lacked transparency regarding training data. By using Nemotron, Salesforce gained a trustworthy foundation. They then post-trained Koa using synthetic data that mimics real-world scenarios, from irate customers on the phone to sales reps closing deals. Crucially, no actual customer data was used in this training process, addressing a major security concern for enterprise buyers.
Cost efficiency is the other half of the value proposition. Koa is engineered to be token-efficient, meaning it burns fewer resources to achieve the same outcome as larger, closed models. Kari Ann Briski from Nvidia highlighted the “trifecta” of sovereign AI, fast time-to-first-token, and efficient reasoning. For CFOs watching their AI spend creep upward, this offers a tangible way to reduce operational costs while maintaining high performance. The model integrates directly into Salesforce’s Agentforce platform, allowing automatic routing based on task complexity.
This move signals that the enterprise AI market is maturing beyond simple prompt-and-response interactions. Companies are demanding tools that respect data boundaries and offer predictable pricing. While Salesforce isn’t abandoning its partners—it simultaneously launched Claudeforce to integrate Anthropic’s Claude securely—the introduction of Koa proves they are serious about owning their AI stack. The question now is whether other enterprise giants will follow suit, building their own vertical-specific reasoning layers rather than renting intelligence from Silicon Valley’s biggest labs.
