Microsoft Research benchmark finds frontier LLMs corrupt 25% of document content over 20 delegated interactions across 52 professional domains

A Microsoft Research paper introduces DELEGATE-52, a benchmark simulating long-horizon delegated workflows across 52 professional domains with 19 LLMs tested. Even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) degrade document content by an average of 25% over 20 interactions, with the average across all 19 models reaching 50% degradation; Python is the only domain where a majority of models meet a readiness threshold of 98% reconstruction fidelity. Critically, agentic tool use does not improve performance, degradation compounds with document size and interaction length, and roughly 80% of total loss stems from sparse catastrophic single-step failures rather than gradual drift. For financial institutions evaluating AI deployment in knowledge-work workflows such as document drafting, compliance review, or report editing, this provides a concrete, cross-domain reliability floor for current frontier models under realistic multi-step delegation conditions.