Salesforce just announced Koa, a custom AI model built specifically for CRM tasks. They created it by fine-tuning Nvidia's open Nemotron 3 model on three decades of synthetic business data, and the results look solid—it matches or beats leading models on CRM actions while making 3x fewer errors. The company also rolled out three other AI tools at Dreamforce: AIforce (a conversational interface for querying your Salesforce data), Claudeforce (which puts Claude's capabilities directly into Salesforce with 37 pre-built sales skills), and Headless 360 (which lets you access all your Salesforce data through APIs and plugins without touching Salesforce's own interface). Why build your own model instead of just using ChatGPT or Claude? Control and security. When intelligence becomes your competitive advantage, you can't afford to hand your proprietary business data and decision-making processes to a third party—especially one that might be working with your competitors. That logic applies to Crowdstrike and Thomson Reuters too, which have launched their own domain-specific models. The smart move is using open models as a foundation and then customizing them through post-training rather than training from scratch, which saves time and money while avoiding the complexity of becoming a frontier AI lab. The real question now is whether Salesforce will help its own customers build custom models the same way. As a platform company, they're positioned to do this, and it would let enterprises capture more value from their AI investments. That's where the next wave of enterprise AI seems headed—companies owning their own intelligence rather than renting it from someone else.