Deel’s Emet lab is publishing benchmarks for workplace AI, while its Pangaea models target payroll and cross-border questions.
Deel is building its own AI models around the work its business already handles: payroll and employment across borders. The company announced Emet, a lab for evaluating and training models, and Pangaea, a family of models designed for global work. The move puts a practical question at the center of workplace AI: can a system complete a consequential task accurately, rather than simply produce a convincing answer?
That distinction matters for small firms and growing companies. A payroll error can affect a paycheck; an incorrect answer about hiring across borders can shape a decision about when someone can start work. Deel says general-purpose models still fall short on these kinds of tasks, even when they can search for information. Its answer is to build evaluations from specialist knowledge and train models for the situations its customers encounter.
Benchmarks built around real tasks
Deel introduced two EmBench evaluations. EmBench Payroll covers 295 real payroll tasks across eight countries. The company says 29 in-country payroll specialists helped create the benchmark. The best-performing model met 90% of the specialists’ criteria, but fully solved only 49% of the tasks.
That gap between meeting most criteria and completing a task is the telling measure. A model may get many parts of an answer right and still miss a detail that changes the outcome. For business users, a polished response can be harder to assess than an obvious failure, so task-level evaluation offers a more useful signal than a general claim that a system performs well.
EmBench Global Work tests a different set of questions: 601 in total, covering issues such as notice periods and who can start work in a particular country. Answers are scored fact by fact. Deel says the best model, when equipped with search, gets 75% of the facts right. The company’s point is that access to more information does not, on its own, guarantee dependable answers in a specialized field.
Deel says EmBench is open to every model builder, and that models are scored against answers from experts. Making the benchmarks available gives other developers a way to assess their systems against the same work tasks. It also gives operators a more concrete basis for asking where an AI tool performs well and where a person should check its output.
Pangaea targets payroll and global work
Pangaea is Deel’s family of models for payroll and cross-border employment questions. The company says the models use open-weight models and its own knowledge base, rather than personal or customer data. Deel plans for Pangaea to serve as the main model behind AI features in its platform.
The first model, Pangaea K1, focuses on global work. It draws on a Deel knowledge base of more than 43,500 pages covering subjects such as leave rules, notice periods and hiring in a new country. Deel says K1 answers from that information without searching at the time of the question. On EmBench Global Work, the company says K1 comes within five points of the top model with search, while responding from memory at up to three times the speed.
Pangaea P1 is trained for payroll. Deel says Emet’s in-country specialists created the payroll data used in its training. On the 110 hardest EmBench Payroll tasks, the training increased P1’s share of criteria passed from 55% to 61%. Deel says that result came at about 40 times lower cost than the top model.
Those figures describe different measures: K1’s comparison concerns fact-based questions about global work, while P1’s result tracks criteria passed on a subset of difficult payroll tasks. They offer an early look at how Deel is assessing its models, but the company’s results do not establish how every system will perform in a particular firm’s day-to-day workflow.
What business teams can take from the launch
Emet’s approach reflects a useful operating principle for companies adopting AI: define the task, identify what a correct result requires, then measure performance against specialist review. A model’s general fluency is a poor substitute for that process when the work depends on country-specific details or calculations with direct effects on employees.
Deel’s next test will be whether its benchmarks and models hold up across the variety of cases that payroll and cross-border teams face. For businesses considering AI in HR or payroll, the practical takeaway is to examine performance on the work they actually do, including what happens when an answer is incomplete. Deel says more information is available from Emet and its Pangaea page.
