Salesforce released Koa, a reasoning model built on Nvidia's Nemotron architecture designed specifically for CRM workflows. The model powers Agentforce to handle multistep tasks like routing support cases, updating sales opportunities, and qualifying leads. Salesforce claims Koa matches or beats general-purpose models on CRM work with three times fewer errors, but the company hasn't named which models it compared against or what benchmark it used. The training approach relied entirely on synthetic data rather than real customer records. Salesforce simulated what it calls "almost 30 years' worth of internal CRM deployment experiences" across 14 industries including healthcare, finance, and manufacturing. The training pipeline combined supervised fine-tuning with reinforcement learning and a technique called group relative policy optimization to handle multistep tool execution. Chief Platform and Engineering Officer Rohan Kumar emphasized the depth of knowledge embedded in Salesforce's own CRM history as the foundation. Koa is currently in pilot with 1-800Accountant as a launch customer, with general availability expected in winter. The move fits Salesforce's broader push into enterprise AI and continues Nvidia's strategy of embedding Nemotron into partner products across the market.