📰 Key Takeaways

Koa is Salesforce’s first reasoning model, unveiled this week at the Dreamforce conference. It’s built on Nvidia’s open-weight model Nemotron, with post-training done jointly by both companies, and it’s focused on tasks related to sales, marketing, and customer service.

Jayesh Govindarajan, Salesforce’s EVP of AI, said that while the company has already built a number of task-specific small language models as part of the Agentforce platform, it has relied on frontier models like Claude and ChatGPT for long or multi-step tasks that require reasoning — routing requests to them through Agentforce’s AI gateway, the system that dispatches requests to the appropriate model. The reason Salesforce hadn’t built its own enterprise-grade frontier model sooner is that the market lacked a top-tier, US-based pretrained foundation model with clear data provenance — a gap Nemotron finally filled. Govindarajan also noted that, by contrast, there’s no way to confirm what data Alibaba’s Chinese open-source model Qwen was actually trained on.

During post-training, Salesforce and Nvidia didn’t use any real customer data. Instead, they built synthetic data simulating sales and customer service scenarios — including persona simulations of furious customers calling into a support center and sales reps handling those situations — to give the model deep domain expertise in sales and customer service.

Koa is positioned as an open-weight alternative to the models already available on the Agentforce platform. Its advantages include: it has never been exposed to real customer data, so there’s no risk of leakage; it uses fewer tokens to complete the same tasks, cutting AI spend; it can be automatically routed to as needed through the AI gateway; and it fully complies with Salesforce’s built-in customer data governance and security controls.

💬 JudyAI Lab’s Take

What stands out to us is that Salesforce didn’t take the easy route of just plugging in a frontier model. Instead, it partnered with Nvidia to build its own reasoning model, Koa, on top of the open-weight Nemotron — purpose-built for multi-step reasoning tasks in sales and customer service.

This case marks a turning point in enterprise AI adoption: when tasks require long-horizon reasoning, routing through general-purpose frontier models via a gateway just isn’t enough anymore. Enterprises are starting to care a lot more about two things: traceable data provenance and never touching real customer data. Salesforce’s choice to generate synthetic data from simulated scenarios — like playing out an angry customer call or a sales rep’s response — rather than using real customer interaction logs for post-training, solves both the data leakage risk and the training data compliance problem at once. For AI builders, this is a reminder: beyond raw model capability, “where did the data come from” and “how was it trained” are becoming hard requirements for enterprise AI procurement, not nice-to-haves.

Worth asking yourself: if you had to prove “data security” to an enterprise customer for your own AI product today, do you have a comparable synthetic-data or de-identification process you could point to?

📅 Source Info

🔗 Further Reading