Salesforce and Nvidia Ship an Air-Gapped AI That Makes 3x Fewer Errors Than Frontier Models

A digital agent at UChicago Medicine can now resolve complex, multi-step customer service disputes without a single piece of patient data ever crossing into a third-party cloud. Yesterday, an enterprise that wanted an AI to handle intricate reasoning had to route those prompts out to general-purpose giants like OpenAI or Anthropic. Today, they can run an engine that does the exact same job entirely within their own secure network. It marks a fundamental shift in how businesses handle their most sensitive workflows, proving that specialized expertise can out-reason the broadest artificial intelligence on the market.

Beating the Generalists by 3x

The math behind the breakthrough comes down to a hyper-specific training diet. When Salesforce benchmarked its new Koa model on real-world tasks like updating sales opportunities, scheduling follow-ups, and routing complex cases, it matched the top frontier models in completion rates, but produced three times fewer errors.

This accuracy stems from a unique dataset. Instead of scraping the open internet, Salesforce distilled 27 years of proprietary enterprise deployments across 14 industries into synthetic training data. They then layered this deeply specialized knowledge onto Nvidia’s open-weight Nemotron base model. Rohan Kumar, Salesforce’s engineering chief, spearheaded the technical collaboration with Nvidia to encapsulate decades of customer management workflows into this highly focused reasoning engine.

The result is a system that understands the exact lifecycle of a service case and the specific structure of a financial deal. But building a smart model was only half the challenge; the other half was keeping it locked down.

Bringing the Brain Inside the Firewall

Operations leaders in highly regulated fields face a strict ceiling: they cannot outsource their data to external APIs. Jayesh Govindarajan, EVP of Salesforce AI, identified this exact bottleneck. Companies needed elite reasoning capabilities, but they lacked a transparent, sovereign model they could build on securely.

“Until Nemotron came along, there was no sovereign American pre-trained model that was available that had clear data provenance.”— Jayesh Govindarajan

By embedding Koa directly within its Agentforce platform, Salesforce allows organizations like 1-800Accountant, Baxter Credit Union, and government agencies to deploy autonomous agents across private clouds. The dynamic mirrors the historic shift from mainframe computing to personal computers. Just as PCs brought processing power safely in-house for maximum control, localized reasoning models bring cognitive work inside the enterprise boundary.

It fundamentally alters the economics of automation. Companies no longer have to pay a premium for a massive AI that knows how to write code or compose sonnets just to parse a tax rule. Now, the open question is exactly how these local agents will handle the unpredictable curveballs of real-world business.

What people are saying

Keith Edwards@keithedwards

“Something clearly happened with a frontier AI model that hasn't been made public and it spooked them so much that it made Elon Musk, Dario Amodei, and Sam Altman all simultaneously agree to slow down.”

28.9K4.7K3.5K3.8MSep 12

Financial Review@FinancialReview

“Anthropic, OpenAI and SpaceX have suddenly warned that the industry must slow its development of AI frontier models. But as Chanticleer writes, the surprising development could have big ramifications for global markets.”

312.9KSep 14

Ron@CodeMonkeyZ

“Altman, Amodei and Musk are likely realizing that Google did not focus their external efforts on improving consumer ai and instead went all in on unlocking ASI.”

7.9K2151.0K957.1KSep 14