AI watched 115 hours of professional work—and still missed rejection
Fifty professionals used agents on real projects, but frontier models struggled to predict when people would modify or reject the output. The preprint tests understanding of collaboration, not productivity or job loss.
The short version
- What happened
- Kolab records 115 hours from 50 professionals in 40 occupations, covering 100 problems, 508 workflow steps and more than 90 software tools.
- Who it affects
- Workers may have to repair or reject AI output outside the chat interface, where an agent that assumes success can add coordination and checking work.
- What remains unproven
- Unreviewed preprint; 50 paid, mostly AI-familiar participants; one main agent interface; self-chosen tasks; no no-AI comparator or independent productivity outcome.
Keep in mind
Unreviewed preprint; 50 paid, mostly AI-familiar participants; one main agent interface; self-chosen tasks; no no-AI comparator or independent productivity outcome.
Source analysis · 1 linked source. Primary analysed source date: 15 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Professionals often changed what left the chat
A new workplace study recorded what happened after AI produced an answer—not just what appeared in the chat window. Fifty professionals from 40 occupations completed two hour-long sessions on projects they had chosen from their own fields. Across 115 hours, researchers mapped 100 problems into 508 workflow steps and labelled 692 agent contributions. Professionals adopted 65.9% unchanged, modified 19.1% and rejected 15.0%. The finding matters because an apparently fluent answer can become useful, altered or discarded only after it meets the rest of a person's work.
The occupations spanned ten broad sectors and 21 countries, including architects, lawyers, journalists, engineers, educators and clinicians. Participants mainly worked with Claude but could use other agents, and their tasks had to require domain expertise, at least five interdependent steps and a concrete deliverable. This is closer to professional practice than a one-shot benchmark, although the sample was small and unusually AI-confident: 37 of 50 rated their generative-AI familiarity between eight and ten out of ten.[1]
Frontier models confused acceptance with understanding
Researchers then asked multimodal models to recognise or forecast whether a contribution would be adopted, modified or rejected. On the full video-only set, Claude Opus 5.5 reached macro-F1 of 0.676 when the human response was visible and 0.492 when it had to forecast that response. GPT-6 Astra was tested on fewer cases because of cost, so its headline score is not directly comparable with the full-set results. On the shared subset, Claude led recognition and GPT-6 led forecasting, but both still found modification and rejection harder than simple adoption.
The error pattern is the useful result. Gemini 3.1 Pro labelled 52% of true modifications and 71% of true rejections as adoption. Chat logs alone were especially weak at detecting rejection, because the decisive action often happened later in another program. Adding longer context did not reliably solve the problem: most video and workflow-graph changes stayed below 0.07 macro-F1, and none of the video gains was statistically significant. More recorded history was not automatically better understanding.[1]
The distinction that matters
What happened to 692 agent contributions
The study followed outputs beyond the chat window and recorded the professional's next action.
Adopt
Used substantially as given
65.9% (456)
Modify
Changed before use
19.1% (132)
Reject
Discarded or replaced
15.0% (104)
The benchmark cannot tell employers whether agents help
Participants rated collaboration and output quality highly, with roughly 85% of sessions scoring four or five on both measures. Yet every tested model had worse mean absolute error than simply predicting the most common rating. Rank agreement with people was also weak: the highest reported Spearman correlations were 0.379 for collaboration satisfaction and 0.352 for output quality. A system that routinely overpredicts approval could miss when a worker needs clarification, revision or control returned.
The preprint does not compare work with and without AI, measure time saved, audit deliverables independently or estimate employment effects. Recruitment through word of mouth, Prolific and Upwork, payment of $100 plus bonuses, heavy representation of advanced users and one default agent all limit generalisation. The study had ethics approval, removed personal identifiers and was partly supported by a University of Minnesota fellowship. Confidence would rise with larger occupational samples, independent quality assessment, low-familiarity users and controlled comparisons of error, time and workload. For employers, the message is practical: keeping a person in the loop is not enough if the system cannot recognise when that person has rejected its work.[1]