ATLAS Finance test reportedly puts best AI at 12%, versus 100% for humans

A September 15, 2026 post describing ATLAS Finance reports results of 100% for humans and 12% for the best AI. The author says their team gave 11 frontier models 100 expert-level tasks that each take humans 15–30 hours. Agents had access to resources including email, chat, calendars, documents and Excel. Dozens of coworker and client personas introduced ambiguity, controlled contradictions and real-time updates that agents had to resolve to produce client-ready deliverables. The author argues AI remains far from meeting the standard required of junior professionals in banking, advisory and private equity.