Salesforce launches Koa reasoning AI model with Nvidia, expanding enterprise partnerships

Salesforce unveiled Koa, its first CRM reasoning model, at the Dreamforce conference on September 15, 2026. The model was built by post-training NVIDIA's open-weight Nemotron 3 Super on a proprietary synthetic dataset drawn from 27 years of Salesforce CRM intelligence, and it is designed to reason through the complex, multi-step workflows that sales, marketing and customer-support agents run. Koa will power Agentforce, the platform CRM customers use to build agents for tasks such as answering service questions and scheduling follow-ups.

The technical claim behind the launch is efficiency rather than raw capability. On Salesforce's own CRM Benchmark, a suite of real-world tasks like updating an opportunity or routing a case, the company says Koa matches or exceeds leading model performance on CRM actions with three times fewer errors. Salesforce holds the model weights and runs both post-training and inference inside its own trust boundary, and says no customer data was used in training: the corpus was synthetic scenarios simulating workflows across more than 14 industries, from manufacturing and financial services to healthcare and travel.

Until now, Agentforce routed reasoning-heavy prompts to outside frontier models through its AI gateway. "Reasoning has always been something that we've relied on the frontier model providers for. Until now," Jayesh Govindarajan, executive vice president of Salesforce AI, told TechCrunch, adding that the blocker had been the absence of a state-of-the-art American pre-trained base model with clear data provenance until Nemotron. For NVDA, the deal extends Nemotron's reach beyond developer adoption into a flagship enterprise software stack, and the two companies are also bringing Nemotron models and accelerated computing into Missionforce for government and regulated customers running private cloud or air-gapped deployments.

The framing that Koa replaces frontier labs deserves a qualifier that the headlines skipped. TechCrunch notes Salesforce is not abandoning Anthropic or OpenAI: it simultaneously announced Claudeforce, a partnership letting companies use Claude as their AI interface while their records stay inside Salesforce infrastructure. The practical picture is a routing decision, with Koa absorbing high-volume CRM-shaped reasoning and outside models kept for the rest.

The announcement sits alongside broader alliance news with NVIDIA, Google Cloud and AWS, and a widening third-party tooling layer: Copado launched Agentia Headless to run governed AI agents inside Salesforce development environments, and Brillio joined the Forward Deployed Engineering Partner Network to push agentic projects into production. The same coverage carries the bear case, namely that Microsoft and ServiceNow are competing for exactly this workflow layer while hyperscalers push their own agent frameworks.

Shares were roughly flat on the day, with CRM at $259.97, up 0.2%, and SentiSense sentiment on the name running positive at +0.27 against a 30-day average of +0.36 (SentiSense data, September 15, 2026). Whether Koa moves the financial needle depends on Agentforce consumption rather than the launch itself, so the metrics to watch are adoption disclosures, any shift in AI gateway routing mix toward the in-house model, and whether the claimed token efficiency shows up as margin. Salesforce's most recent signal on that front is free cash flow up 81.5% year over year from Q2 FY2026 to Q2 FY2027 (SentiSense data).