Salesforce just did something the big AI labs probably did not want to see: it built its own AI model instead of paying OpenAI or Anthropic for one.
The model is called Koa, and it launched at Dreamforce, Salesforce's big annual conference. It is Salesforce's first reasoning model, meaning it can work through multi-step problems, not just answer simple questions. Think of a customer service agent that needs to check a return policy, look up an order, calculate a refund, and then explain it to an upset customer, all in one flow. That kind of multi-step thinking used to require sending the request to an outside model like ChatGPT or Claude, and paying for it every single time.
Salesforce built Koa on top of a model called Nemotron, made by Nvidia and given away for free for anyone to build on. Nvidia's executive in charge of enterprise AI software described the appeal as getting efficient reasoning without burning excess computing power on every request, which lowers the cost of running it at scale.
Why did Salesforce wait until now to build this? Its own AI lead said the real blocker was not effort, it was the lack of a strong starting model made in the United States with a clear paper trail on what data trained it. He specifically contrasted this with Qwen, the popular Chinese open model from Alibaba, saying Salesforce has no way to verify what went into training it. For a company that sells trust and data security as its core product to big businesses, using a foreign model with unclear origins was never going to happen.
The training method is also worth noting. Instead of using real customer conversations, Salesforce and Nvidia built fake ones. They simulated angry customers calling into support and sales reps trying to close deals, then trained Koa on those made-up scenarios. That means no actual client data ever touched the model, which closes off a real fear many businesses have: that their private data could somehow leak out through a shared AI system used by other companies too.
This is not Salesforce cutting ties with the frontier labs. In the same conference, it expanded its partnership with Anthropic into something called Claudeforce, letting customers plug Claude directly into their Salesforce data while keeping that data inside Salesforce's own walls. So the strategy is not either-or. It is: use Koa for the bulk of routine, repetitive agent work where cost and control matter most, and reach for Claude or another frontier model when a task genuinely needs top-tier intelligence.
For any business already using or considering AI agents to handle customer service, sales follow-ups, or scheduling, this is a signal worth watching. Software vendors are starting to build their own narrow, cheaper AI models instead of just reselling access to OpenAI or Anthropic. That likely means lower prices down the line for routine AI tasks, but it also means buyers need to ask vendors exactly whose model is doing the work and what data trained it, because the answer increasingly affects both your costs and your risk.