Salesforce just showed the AI labs what enterprise software wants, and it did not involve uploading your code into a black box. At its Dreamforce conference, the company unveiled Koa, its first reasoning model, built on Nvidia’s open-weight Nemotron foundation. The announcement is a direct answer to the question of whether big business can trust models that see their data.
What Koa Is
Koa is a reasoning model, meaning it handles complex, step-by-step decision-making rather than simple retrieval. Salesforce and Nvidia trained it together for sales, marketing, and customer support tasks. The model is an alternative to the frontier models that dominate the market — the ones that charge by the token and require companies to upload everything they own.
The key difference is what Salesforce did not do. It did not train Koa on actual customer data. Instead, it used synthetic data — simulated scenarios that mimic real-world sales and service interactions.
The Synthetic Data Problem
Training a model on real customer data carries risk. Data leaks happen. Models absorb information they should not have. Salesforce chose a different path.
Jayesh Govindarajan, EVP of Salesforce AI, explained the approach to TechCrunch. “We actually simulated a customer service environment with a persona customer service professional, including irate customers that call into the customer service center, all the way to a sales professional who’s trying to close a deal,” he said.
The simulation approach means Koa learns patterns without ever seeing the underlying records. That is a deliberate design choice, not an accident.
Why Nemotron Matters
Nemotron is the foundation that made Koa possible. Before it existed, Govindarajan said, there was no sovereign American pre-trained model that met three criteria:
- It had to be state-of-the-art
- It had to have clear data provenance
- It had to avoid the unknown training data of models like Qwen, the popular Chinese open-weight model from Alibaba
“We have no idea what Qwen trains on,” Govindarajan said, explaining why Salesforce waited for a domestic option.
The result is a model that fits into Salesforce’s existing Agentforce platform. Agentforce lets customers build agents for rote tasks like answering service questions or scheduling appointments. Koa will sit alongside other models in that lineup, giving customers a reasoning option that does not raise the same privacy concerns as the frontier alternatives.
The Token Efficiency Argument
Koa is also cheaper to run. The model uses fewer tokens to perform the same work as Claude or ChatGPT, according to Nvidia’s VP of Generative AI Software for Enterprise, Kari Ann Briski.
“With Nemotron, we have a unique architecture for inference to be token efficient,” Briski told TechCrunch. “It’s kind of the trifecta of things that you need to have: sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all.”
That efficiency matters for customers who run agents on a large scale.
The Gateway Route
Koa integrates with Agentforce’s AI gateway, the system that decides which model handles each request. Before Koa, long-running or multi-step tasks were routed to frontier models like Claude or ChatGPT.
Now the route stays inside the enterprise. If a prompt needs reasoning, Koa handles it directly. The gateway routes requests based on the need, so simpler queries go to smaller models and complex ones go to Koa.
What Salesforce Keeps
Salesforce is not abandoning Anthropic or OpenAI entirely. It just announced a partnership with Anthropic called ClaudeForce. That arrangement lets companies use Claude as their AI interface while keeping their data in Salesforce’s system of records, secured by its infrastructure.
The announcement shows Salesforce is keeping its options open.
What This Means for the Labs
The enterprise world has different priorities than the frontier labs. The labs want businesses to upload files, code, prompts, and feedback. They want them to spend millions doing it.
Salesforce is showing that is not the only path. A model trained on synthetic data, with clear provenance, and delivered as an open-weight alternative can do the job without the risk. The model respects customer data requirements and security embedded within Salesforce.
The announcement is a demonstration of divergence. The enterprise world’s needs for AI are moving away from what the frontier labs are offering.
The Verdict on Koa
Koa is a bold move. It is a reasoning model that avoids real customer data, simulates scenarios instead, and delivers specialized performance at lower cost. The model is a proof point for the enterprise argument: AI can be proprietary, efficient, and respectful of data boundaries.
Salesforce built it with Nvidia, trained it on synthetic data, and released it as an open-weight alternative.
The AI labs should take note.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.