The Task Horizon Illusion: Why METR’s Exponential Doesn’t Reach the Enterprise Data Problem

Abstract

METR’s time-horizon metric shows AI agents doubling their autonomous task-completion capability every seven months, with recent acceleration to an 89-day doubling period. Claude Opus 4.5 reached a 50% time horizon of approximately 5 hours in late 2025. These findings have catalyzed extraordinary claims: Sequoia Capital has declared “2026: This Is AGI”; Bloomberg Intelligence projects 200,000 Wall Street job cuts; Citi estimates 54% of banking jobs are “highly automatable.” We reproduce the METR trendline from 16,598 raw evaluation runs across 228 tasks and 14 frontier models, confirming the exponential. We then demonstrate that this trend, while methodologically careful, systematically misrepresents the automation frontier in Banking, Financial Services, and Insurance (BFSI). The defining challenge is not task length but data breadth: the reconciliation of structured transactional systems with unstructured corpora under real-time consistency and regulatory audit requirements. We introduce the Data Unification Horizon (DUH) as a complementary metric, identify three load-bearing capabilities absent from all current AI benchmarks, and map BFSI task families against both dimensions. Our analysis reveals a sharp partition: low-DUH tasks (boilerplate generation, single-source reconciliation, rule-based screening) are genuinely headed for obsolescence within 18–24 months, while high-DUH tasks (complex underwriting, multi-party claims, KYC remediation, model validation) will resist time-horizon expansion indefinitely absent enterprise data infrastructure investment. The 200,000 jobs at risk are not writing code. They are reconciling data across systems that don’t talk to each other, under regulatory frameworks that demand auditability. The METR trendline doesn’t reach them. What reaches them is data unification.