Configuration Work: Four Consequences of LLMs-in-use

TL;DR

This study explores how workers adapt and configure Large Language Models (LLMs) for specific tasks, revealing four key consequences that reshape work practices and routines over time.

Contribution

It introduces the concept of configuration work to describe the labor involved in making LLMs usable for professional tasks, highlighting four intertwined consequences.

Findings

Discretization of activities for system processing

Generation of cluttering through iterative prompting and correction

Attuning practices to the system's rigidity and limits

Abstract

This article examines what it means to use Large Language Models in everyday work. Drawing on a seven-month longitudinal qualitative study, we argue that LLMs do not straightforwardly automate or augment tasks. We propose the concept of configuration work to describe the labor through which workers make a generic system usable for a specific professional task. Configuration work materializes in four intertwined consequences. First, workers must discretize their activity, breaking it into units that the system can process. Second, operating the system generates cluttering, as prompting, evaluating, and correcting responses add scattered layers of work that get in the way of existing routines. Third, users gradually attune their practices and expectations to the machine's generic rigidity, making sense of the system's limits and finding space for it within their practices. Fourth, as LLMs absorb repetitive tasks, they desaturate the texture of work, shifting activity toward logistical manipulation of outputs and away from forms of engagement that sustain a sense of accomplishment. Taken together, these consequences suggest that LLMs reshape work through the individualized labor required to configure a universal, task-agnostic system within situated professional ecologies.