‘Now We Can Know Everything and Do Anything,’ Jensen Huang Says at Dreamforce
Jensen Huang announced Salesforce’s first CRM reasoning model, Koa, which is a post‑trained NVIDIA Nemotron 3 Super fine‑tuned on a proprietary synthetic dataset spanning nearly three decades of enterprise CRM deployments. The model runs entirely on Salesforce’s own infrastructure with zero customer data used during training or inference, illustrating a fully in‑house, open‑model deployment. This demonstrates that large language models can be safely customized for enterprise workloads without exposing proprietary data, enabling engineers to build “agentic” AI services at scale.
⚡ Key Takeaways
- Koa is built by post‑training NVIDIA Nemotron 3 Super on a synthetic CRM dataset covering almost 30 years of enterprise deployments.
- The model is fine‑tuned and served entirely within Salesforce’s own infrastructure, keeping all customer data out of the training and inference pipelines.
- Open‑model adoption rose from 30 % to roughly 70 % of deployments since the start of last year, indicating rapid industry shift toward custom AI.
- Salesforce is piloting Agentforce for customer‑support workflows, showing a concrete agent‑style use case built on Koa.
- Safety is treated as an engineering problem; the team will pause releases if functionality, capability, or safety cannot be guaranteed.
Engineers can now see a real‑world blueprint for deploying a fine‑tuned LLM on private infrastructure, achieving enterprise‑grade data isolation while leveraging open‑model capabilities for high‑impact CRM automation.
✅ Practical Steps
- Deploy NVIDIA Nemotron 3 Super, fine‑tune it on a synthetic dataset that mirrors your domain (as Salesforce did with three decades of CRM data).
- Host the fine‑tuned model on your own compute environment, ensuring no customer data is ingested during training or inference.
- Integrate the model into workflow automation (e.g., using Salesforce’s Agentforce pattern) to enable agentic enterprise applications.
Want the full story? Read the original article.
Read on NVIDIA Blog ↗