
Best AI Assistant Fails Most Corporate Rule-Following Tests
Surge AI's HANDBOOK.md benchmark, accepted to a 2026 AI research workshop, tested AI assistants on following 20–124-page company instruction manuals across 65 tasks in simulated offices. The best of thirty setups passed 36.2% of trials; most leading models stayed below 25%. Code is public at github.com/surge-ai/handbook.
Published