Salesforce built its own CRM reasoning model on Nvidia Nemotron

Enterprise AI is hitting a massive fork in the road. On one side, closed-model labs want every company to pipe proprietary documents and code into their cloud endpoints. On the other side, enterprises are staring at mounting token bills and wondering why an AI trained on advanced astrophysics is answering basic support tickets.

Salesforce just made its countermove. On September 15 at Dreamforce, the software giant unveiled Koa, its very first reasoning model purpose-built for CRM. Built on Nvidia’s open-weight Nemotron base (Nemotron 3 Super), Koa is designed to handle complex, multi-step business logic directly inside Salesforce's own trust boundary.

Trained on workflows, not customer data

Until now, Salesforce leaned on closed frontier giants whenever its Agentforce agents needed to reason through multi-step logic. Salesforce AI executive vice president Jayesh Govindarajan admitted that while they built dozens of tiny task-oriented models, real reasoning always got routed out to Claude or ChatGPT.

Govindarajan explained they avoided building their own frontier engine because open models with clear provenance were nonexistent. They refused to adopt Alibaba’s Qwen because of mystery training data, but Nvidia’s open Nemotron gave them a sovereign, auditable architecture with efficient token economics.

Rather than scraping actual customer records, Salesforce and Nvidia post-trained Nemotron using proprietary synthetic data. They modeled 27 years of enterprise processes across 14 industries, simulating real-life scenarios like irate callers and complex B2B contract negotiations.

To keep behavior predictable, Salesforce runs Koa at temperature zero behind a dedicated safety harness. On Salesforce's CRM Bench, Koa posted a 0.86 average score, essentially matching Claude Opus 4.8 at 0.87 and trailing just behind GPT-5.5 at 0.90. Salesforce claims the model cuts errors threefold compared to default models, boosting action precision by 11% and context recall reliability by 2.1 times.

Koa is already humming inside Salesforce’s internal Slack employee agent and piloting with organizations like Formula 1, UChicago Medicine, and Baxter Credit Union. General availability in the U.S. lands in winter 2026. Simultaneously, Salesforce widened its Anthropic bet by launching Claudeforce, bringing 37 native sales skills straight into Claude.

Why it matters

For enterprise operators, Koa offers a practical escape hatch from the open-ended token tax. If your team is deploying agents to route tickets, update records, and schedule handoffs, you probably do not need a multi-trillion parameter math genius. You need an agent that reliably triggers the right database call without leaking records outside your compliance perimeter. Once winter rolls around, IT leads should run A/B pilots comparing Koa against GPT-based setups on error rates and latency.

For the AI ecosystem, this marks a tactical retreat by enterprise software from the closed-model monopoly. Closed labs want to swallow the application layer whole, turning SaaS platforms into passive databases. By teaming with Nvidia on custom-tuned open weights, Salesforce keeps the reasoning engine inside its fortress, insulating its margins and securing negotiating leverage against frontier labs.

That said, walking a tightrope between building Koa and reselling Claude through Claudeforce looks tricky. Salesforce is simultaneously a major customer, a distribution channel, and a direct rival to Anthropic. Whether synthetic workflows can completely match frontier reasoning when edge cases strike messy, unformatted corporate records remains an open question.