WorkWorlds: An Infrastructure for Evaluating AI Agents on Workplace Tasks
Yining Hua, Levi Lian · arXiv · 20 Sept 2026 · not peer-reviewed
Source finding
Across 192 matched evaluations, estimated evidence access was 90.4% with preselected documents and 74.5% with all role-accessible documents. Criterion pass was 79.4% versus 68.2%. Most of the difference occurred before sufficient evidence was reached.
Method and limits
Controlled evaluation of three agents, eight tasks and six roles in a simulated pharmaceutical company. These are model-estimated probabilities, not population shares or measured company savings. The preprint has not been peer reviewed.
What companies can do
Include information retrieval and actual access permissions in AI tests.