TL;DR
Research Warning: Microsoft Research says AI agents still corrupt work documents during long delegated editing chains. Benchmark Result: Frontier models can lose 25 percent of document content, while average degradation across all models reaches up to 50 percent. Buyer Impact: Companies may expand AI automation spending without cutting the human review, audit, and approval work these systems still need.
Microsoft Research has published a preprint warning that current AI agents still corrupt or degrade work documents in long-running delegated workflows. The paper authors laid out that warning around repeated handoffs instead of one-off prompts.
DELEGATE-52, the benchmark used by the Microsoft researchers, tracked what happened after models kept revising the same work product across 20 delegated interactions. Frontier systems posted 25 percent document loss by the end of those chains, while average degradation across all models reached 50 percent. Enterprise buyers are not just asking whether a model can produce a polished first draft. They need to know whether a system can preserve intent, structure, and factual detail through a long revision chain.
Repeated edits are where the commercial promise starts to break. Contracts, policy notes, technical drafts, and research summaries can all look polished before a reviewer notices that an instruction, number, or qualification has shifted. In document-heavy workflows, AI still behaves more like a fast drafting layer than a trustworthy delegate that can be left alone.
The DELEGATE-52 test shows why long tasks break down
Microsoft Research devised the DELEGATE-52 benchmark to test how large language models handle long-running knowledge-work tasks. The research paper describes delegated document editing across 52 professional domains including coding, crystallography, and music notation. Each workflow keeps the same file in play, so the model has to preserve earlier choices while continuing to revise later sections.
To be considered ready for a given work domain, the researchers set the bar at 98 percent or higher after 20 interactions. Real office work involves drafts, revisions, side files, reference material, and distractor documents, so the benchmark measures the persistence problem that can turn a strong demo into a fragile production workflow. One accurate answer does not prove a system can survive a long editing chain.
Coverage across many task types is part of the test design. Coding, notation, and other specialized assignments force a model to preserve structure as well as language, while document-heavy business work adds references, revision history, and version drift. The study also spans 19 LLMs, which makes the result less about one model family’s quirks and more about a broader limit in current delegation behavior.
The timeline also helps explain why the paper landed quickly in enterprise AI debates. It was submitted to arXiv on April 17, 2026, then surfaced in a May 11 news cycle as companies kept pushing AI deeper into document work. Back in February 2026, DeepMind had already framed delegation around limits and oversight methods. DELEGATE-52 moves that argument from theory into a measurable document workflow.
Where the models failed in practice
Results at the top of the field still looked narrow. Python programming was the only domain considered ready after 20 interactions, and Gemini 3.1 Pro led the table with 11 of 52 domains meeting that bar. Even the strongest system still left the majority of the benchmark below the paper’s own usability threshold.
Failure type mattered almost as much as failure rate. Weaker models often drifted into content deletion, while frontier systems more often produced corruption that left a file looking superficially coherent after meaning had shifted. Single-round failures could cost 10 to 30 benchmark points in one round-trip interaction instead of accumulating slowly enough for a reviewer to intervene early.
In the Microsoft Research publication, the authors condensed that broad failure pattern into one line.
A per-domain breakdown of end-of-simulation scores found that models were not ready for delegated workflows across the vast majority of domains.
Microsoft Research paper authors
Severe failures were not confined to one corner of the test. The study found catastrophic corruption, defined as a benchmark score of 80 percent or less, in more than 80 percent of model-domain combinations. Frontier systems also finished long workflows with substantial content loss, which makes the issue look operational rather than like a collection of edge cases.
Giving models more power did not fix the weakness. A basic file-and-code harness still left the four tested systems incurring an average additional degradation of 6 percent by the end of the simulation. Larger documents, longer interactions, and distractor files all worsened the outcome, suggesting that extra tool access can create more chances to overwrite context, misread state, or carry a bad edit into the next step.
Enterprise agents are usually sold with file access, document rewriting, retrieval hooks, and code execution so they can work across a process instead of answering one prompt. If the same tool stack increases degradation inside the benchmark, buyers have to treat more capability as a new reliability variable rather than an automatic improvement.
Silent corruption is harder to manage than obvious failure. Missing text can trigger a quick human check, but a polished paragraph that quietly changes a clause, a date, or a number can survive long enough to shape later revisions. Legal, finance, engineering, and policy teams are exactly the groups that pay for that kind of delayed detection.
Why the findings matter for enterprise AI rollouts
Budgets are moving before long-workflow reliability is settled. Deloitte has put AI automation at 36 percent of digital budgets. If that spending continues while document delegation remains unreliable, organizations may accelerate deployment without reducing the supervision burden that keeps those projects expensive.
Progress inside the model families is real, but the paper suggests it still falls short of dependable delegation across varied document tasks. OpenAI’s GPT model family showed benchmark gains from 14.7 percent to 71.5 percent over 16 months. Even so, the authors treated only some domains, including Python coding, as close to usable while warning that users still need close oversight during delegated work.
Compliance and labor planning are where that distinction becomes expensive. If an agent cannot preserve intent through repeated edits, organizations still need people to verify structure, factual accuracy, and silent changes before a file reaches a customer, regulator, or internal decision-maker. Teams may save time on first-pass drafting, but they do not eliminate review checkpoints, escalation rules, approval chains, and sampling audits.
Procurement teams also have to price the hidden labor around these systems. Faster drafting matters less if the savings disappear in downstream checking, correction, and rework. Once file permissions, revision loops, and business-important context enter the workflow, the relevant metric is whether supervision hours fall after the AI has touched the document several times.
Governance teams then inherit a second problem: proving where responsibility sits after a chain of AI edits. Audit trails, approval logs, and reviewer signoffs become more important when the final file may be fluent but not faithful to the instruction trail. That keeps deployments narrower: a company may still use AI for first-pass summarization, coding help, or constrained document cleanup, but regulated or customer-facing files still need stronger approval gates if the model can silently alter meaning midstream. Human oversight remains part of the product cost, not a temporary bridge until the models improve.