Microsoft Research: AI Models Corrupt Documents in Delegated Workflows
Microsoft researchers tested 19 large language models on complex multi-step document editing tasks and found they introduce substantial errors, with frontier models losing an average 25% of document content over 20 interactions. The study highlights reliability concerns when delegating document workflows to current AI models.
What happened
Microsoft researchers Philippe Laban, Tobias Schnabel and Jennifer Neville tested 19 large language models on their ability to complete multi-step editing tasks using a benchmark called DELEGATE-52, which simulated workflows with 310 work environments across 52 professional domains containing real documents and five to ten complex editing tasks each.
The research found that current LLMs introduce substantial errors when editing work documents, with frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4) losing an average 25% of document content over 20 delegated interactions, and an average degradation across all models of 50%.
Reported impact
The errors are described as sparse but severe, silently corrupting documents and compounding over long interactions, raising concerns about the reliability of delegating document editing tasks to AI models in professional workflows.
Index analysis
The benchmark tested LLMs across diverse professional domains including coding, crystallography, genealogy and music sheet notation, finding that models struggle with long-running tasks and more interactions with less structure significantly reduce benchmark performance.
Governance controls
A researcher noted that in enterprise environments where accurate output is crucial, automation workflows would need to be designed with stronger governance to account for the models' limitations rather than relying on them to complete all work independently.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.