At its Dreamforce conference, Salesforce announced Koa, a new reasoning model developed in partnership with Nvidia. Built upon Nvidia’s open-weight Nemotron foundation, Koa represents Salesforce’s first dedicated reasoning model, specifically post-trained to handle sales, marketing, and customer support operations.
The release highlights a growing divergence between enterprise AI requirements and the capabilities offered by leading frontier labs. While proprietary providers often encourage companies to upload sensitive data, code, and feedback directly into their systems, Salesforce is positioning Koa as a secure, open-weight alternative. The company emphasizes that the model was trained exclusively on synthetic data designed to mimic customer interaction patterns, ensuring no actual client data was ingested or at risk of leakage.
Jayesh Govindarajan, Executive Vice President of Salesforce AI, explained that while the company had previously relied on frontier models like Anthropic’s Claude or OpenAI’s ChatGPT for complex, multi-step tasks via its Agentforce platform, Koa now allows these requests to be handled internally. He noted that the absence of a high-quality, sovereign American pre-trained base model with clear data provenance had previously hindered such efforts, a gap filled by Nemotron.
To prepare Koa for enterprise use, Salesforce and Nvidia simulated diverse service environments, ranging from irate customers calling support lines to sales professionals closing deals. This approach aims to make the model more efficient at practical work tasks and cheaper in terms of token usage compared to sending requests to external frontier models.
Kari Ann Briski, Nvidia’s Vice President of Generative AI Software for Enterprise, described Nemotron’s architecture as optimized for inference efficiency. She highlighted the model’s ability to deliver on three critical fronts: sovereign AI capabilities, low latency in time to first token, and efficient reasoning that improves overall tokenomics.
Despite the launch of Koa, Salesforce is maintaining its relationships with major AI providers. The company also introduced ClaudeForce, a new partnership with Anthropic that allows businesses to utilize Claude as an AI interface while keeping their data secured within Salesforce’s infrastructure.
Reasoning models tailored for sales workflows could genuinely cut costs. Hope the latency benefits are as strong as they claim.
Interesting they’re keeping Anthropic around too. Seems like hedging bets rather than a full commitment to the open-weight route.
I’ll believe the token efficiency claims when I see the benchmarks. Synthetic data often struggles with real-world edge cases.
Finally, a model that doesn’t require feeding proprietary sales data to giant clouds. This sovereignty angle is huge for enterprise compliance.