“Frontier models are excellent generalists, but it’s becoming more important to choose the right model for the job. With Pangaea, our payroll and compliance knowledge is trained directly into the model, delivering answers faster and cheaper at scale, and validated by the experts who do this work every day.”
Getting a global work question wrong can mean a missed notice period, an unlawful dismissal, or a hire who cannot legally start. EmBench Global Work is built to catch those errors in AI models. Its questions reflect the work Deel’s specialists resolve daily across hiring, payroll, and compliance for more than 40,000 customers in over 150 countries. Having run this work directly for seven years, Deel’s experts set the standard for what counts as a correct answer.
Answers are scored fact by fact. Each reference answer is broken into individual facts, marked as either necessary or supporting, so a model that writes well but leaves out key information scores lower. The 601 questions come from Deel’s internal knowledge base of 43,500 pages covering policies, procedures, and country rules.
That knowledge base changes constantly, with about 15% of its pages updated each month. Every fact is tagged as stable or volatile based on its page’s edit history, and the benchmark is updated when information goes out of date.
Through Emet, specialists from the countries they cover check each fact against the law and Deel’s practice, and they update it as regulations change. Deel finds and pays these experts through its own contractor system, which gives Emet what other evaluation labs have to buy or build from scratch: a trusted source of truth and the experts to check it.
Deel tested this on its own AI first. Pangaea K1, trained on that knowledge, rose from 19 to 70 on this benchmark. That places it within five points of the best frontier model with search, and it answers in about 7.7 seconds, roughly half the time of the fastest top model, DeepSeek V4.1 Flash, which takes 14.8 seconds.
As AI takes on more of the work that keeps companies running, Emet gives labs a way to see where their models fall short and the expert-verified knowledge to close the gap, so any answer about global work can be trusted by the person acting on it.
To see how frontier models score on payroll and compliance, read the EmBench results at EmBench Global Work. AI labs and teams that want to improve a model on a specialized domain can talk to the Emet team.
