Deel, the payroll and HR platform serving more than 40,000 companies across 150-plus countries, is launching Emet and Pangaea for teams that hire, pay, and manage people across borders, and for model builders who want expert-labeled business tasks. Emet is an AI lab for payroll, HR, compliance, legal, and finance. Practicing specialists from Deel's own teams and its global network define what a correct task looks like. Deel finds, verifies, and pays them, then turns that work into public benchmarks and training data called EmBench. On EmBench Payroll, 295 real tasks across 8 countries built with 29 in-country specialists, the best model meets 90% of expert criteria but fully solves only 49% of the tasks. On EmBench Global Work, 601 questions on rules such as notice periods and who can legally start a job, the best model with search gets 75% of the facts right. Pangaea is the first family of models trained on that material, and it is rolling out as the main model behind AI inside Deel. Training uses Deel's knowledge base and expert-built data from Emet, and does not use personal or customer data. K1 covers cross-border employment rules and was trained on more than 43,500 pages. It scores 70 on EmBench Global Work against a leading score of 75, answering from memory in a median 7.7 seconds, up to 3x faster than models that search before they answer. P1 covers payroll and builds on Nvidia's Nemotron family, with AI21 Labs running further training on Emet examples. On the 110 hardest EmBench Payroll tasks, that training raised the share of criteria passed from 55% to 61%, at about 40x lower cost than the top model. Deel says the lower cost helps it keep AI included in existing plans, with no charge per question. Emet remains open to other model builders, and EmBench scores every model, Pangaea included, against the same expert answers. CEO and co-founder Alex Bouaziz announced the launch.