The Hidden Work of Making AI Useful

by

Author: Denis Avetisyan

New research reveals that successful integration of generative AI in higher education relies on significant, often unseen, effort from users adapting the technology to complex organizational realities.

This study demonstrates that user-driven workarounds are not failures of implementation, but integral acts of sociotechnical integration and ‘invisible labor’ necessary for functional AI within complex contexts.

Despite growing enthusiasm for Generative AI in higher education, successful implementation often relies on unseen labour beyond strategic goals. This paper, ‘Making AI Work: An Autoethnography of a Workaround in Higher Education’, examines a stalled GenAI project, revealing how user-driven workarounds become essential acts of sociotechnical integration when institutional aims clash with technical limitations and organizational politics. Through an autoethnographic lens, we demonstrate that these workarounds aren’t deviations from intended use, but rather crucial-and often invisible- labour that makes malleable AI functional in practice. How can institutions better recognize and value this hidden work to ensure sustainable and equitable GenAI adoption?

Navigating Scale: The Limits of Data and Model Capacity

Despite the considerable promise of generative AI to revolutionize data analysis and insight generation, a fundamental challenge lies in the sheer magnitude of contemporary datasets. Real-world information, unlike curated examples, often extends to terabytes or even petabytes, far exceeding the capacity of current models to process effectively. This limitation isn’t simply a matter of storage; it impacts computational efficiency, processing time, and ultimately, the ability to extract meaningful patterns. While GenAI excels at identifying relationships within manageable datasets, its performance degrades significantly when confronted with the complexity and volume characteristic of many real-world applications, necessitating innovative approaches to data handling and model architecture to unlock its full potential.

Large Language Models (LLMs), despite their impressive capabilities, operate with a defined ‘context window’ – a limitation on the amount of text they can process at any given time. This poses a considerable challenge when analyzing comprehensive datasets, such as the Staff Development Fund (SDF) Data, which often exceed these boundaries. Essentially, LLMs can only ‘remember’ and effectively utilize information within this window, meaning crucial details from earlier parts of a large document may be lost when processing later sections. Consequently, extracting holistic insights from such datasets requires innovative approaches to either condense information to fit within the window, or to architecturally modify the LLM to expand its contextual awareness – a task that presents significant computational and algorithmic hurdles.

Analyzing expansive datasets presents a considerable challenge to conventional analytical techniques. Traditional methods, designed for more manageable data volumes, often falter when confronted with the sheer size and complexity of modern information streams, requiring substantial computational resources and time. Extracting actionable intelligence frequently necessitates extensive pre-processing – cleaning, filtering, and restructuring data – which can be both labor-intensive and introduce potential biases. Furthermore, these methods may demand significant architectural modifications to existing systems, including the implementation of distributed computing frameworks or the development of specialized algorithms, to effectively handle the scale without sacrificing accuracy or speed. Consequently, unlocking the full potential of large datasets often requires a move beyond established practices and the adoption of innovative approaches to data processing and analysis.

Bridging the Gap: Adapting Data for Practical Application

To address limitations in accessing and interpreting Source Data Format (SDF) data, users implemented a ‘Workaround’ consisting of a customized data summarization process. This process utilizes a Python script to pre-process the SDF data, reducing its complexity and enabling more efficient analysis. The script functions by extracting key data points and compiling them into a condensed format suitable for downstream applications. This bespoke solution was developed in response to inherent constraints within the existing data infrastructure and the need for readily available summarized information.

The process of adapting Structured Data Files (SDF) for use with Microsoft Copilot necessitated a pre-processing stage. Raw SDF data, in its original format, was not directly compatible with Copilot’s querying capabilities. This pre-processing involved transforming the data into a format Copilot could interpret, specifically focusing on structuring the information to facilitate effective question answering. This transformation included data cleaning, normalization, and the application of metadata tags to enhance semantic understanding during the query process, ultimately enabling Copilot to access and utilize the information contained within the SDF files.

Successful implementation of the data summarization workaround extended beyond the technical creation of the Python script and data pre-processing. Significant ‘Articulation Work’ was necessary to integrate the solution into established workflows and address organizational requirements; this involved translating the technical capabilities of the adapted system into terms understandable by stakeholders, negotiating its use within existing data governance policies, and demonstrating its value in relation to current business objectives. This process ensured the workaround wasn’t simply a functional tool, but a usable and accepted component of the organization’s data infrastructure, requiring ongoing communication and adaptation to maintain alignment with evolving needs.

The Price of Pragmatism: Unseen Labor and Shadow Systems

The implementation and ongoing support of the workaround necessitated substantial ‘Invisible Labor,’ defined as the unpaid and often unrecorded effort expended by individuals beyond their formally defined roles. This labor included time spent documenting the workaround for colleagues, providing ad-hoc training, troubleshooting issues arising from its use, and adapting it to changing circumstances or data formats. Crucially, this effort was not reflected in performance reviews, project budgets, or formal help desk requests, remaining largely unacknowledged by the organization’s formal systems of recognition and reward. The cumulative effect of this uncompensated effort constituted a significant, though hidden, cost associated with maintaining the workaround.

The implementation of the workaround necessitated navigating established organizational structures and, consequently, fostered a parallel IT ecosystem known as Shadow IT. Because the solution bypassed formal IT request and approval processes, its development and ongoing support relied on informal networks and individual initiative. This circumvention wasn’t simply a technical matter; it was a political one, requiring those involved to manage perceptions and avoid conflicts with departments responsible for official IT governance. The resulting Shadow IT infrastructure, while addressing immediate needs, operated outside of standard security protocols, documentation, and budgetary oversight, creating potential risks alongside its benefits.

The application of autoethnographic methods demonstrated that implementing the workaround necessitated ongoing negotiation and adaptation to pre-existing organizational power structures. Researchers observed that successful integration wasn’t solely a technical matter; it required individuals to understand and respond to the formal and informal hierarchies within the company. This involved continuous assessment of stakeholder influence, strategic communication to manage expectations and minimize resistance, and a willingness to modify the workaround based on feedback from, or circumvention of, official channels. The process revealed that the workaround’s sustainability depended not on its inherent technical merit, but on the ability of its proponents to navigate and leverage these power dynamics effectively.

Technology as a Social Artifact: Reciprocal Shaping in Practice

The concept of technology’s duality suggests that technological development isn’t a one-way street of innovation dictating social change, but rather a continuous interplay between the two. This is clearly demonstrated by instances where users creatively adapt systems to bypass limitations, a process which isn’t merely a ‘fix’ to a technical issue. Instead, these workarounds actively reshape both the technology itself – through unanticipated usage – and the social context surrounding it by altering established workflows and power dynamics. The technology, in turn, influences subsequent social practices and further technological development, creating a reciprocal loop where each element continuously molds the other. This perspective moves beyond viewing technology as a neutral tool and acknowledges its inherent social constitution, highlighting how meaning and functionality are jointly produced by both technical systems and the people who employ them.

Alter’s Process Theory illuminates the development of this workaround not as a sudden innovation, but as a series of interconnected stages. Initially, a perceived need – a disconnect between the intended technological function and practical organizational requirements – sparked the process. This need prompted a period of diagnosis, where users actively analyzed the situation and identified the core problem. Following diagnosis, a solution began to take shape through iterative experimentation and modification, with users constantly refining their approach based on observed results. This wasn’t a linear progression; rather, it involved continuous feedback loops where execution informed further diagnosis and refinement. Ultimately, the fully realized workaround represents the culmination of this iterative process, demonstrating how technology is not simply applied but actively constructed through ongoing interaction and adaptation within a specific context.

The implementation of new technologies rarely unfolds as initially planned, and this case study illustrates how users proactively bridge the gap between a system’s design and the realities of their working environment. Rather than passively accepting limitations imposed by organizational constraints, individuals actively ‘make’ technology functional through resourceful adaptation and workaround development. This process isn’t simply about fixing technical glitches; it’s a demonstration of human agency, revealing how people reshape tools to align with existing workflows, social norms, and practical needs. Consequently, a comprehensive understanding of the human element – user motivations, collaborative problem-solving, and informal knowledge sharing – is critical for successful technology integration and a realistic assessment of its true impact within any organization.

The study illuminates how users proactively bridge gaps within sociotechnical systems, revealing that apparent ‘workarounds’ are, in fact, essential acts of integration. This process isn’t simply about adapting to technology; it’s about performing the often-unseen labor required to make it function within the constraints of an organization. As Carl Friedrich Gauss observed, “If I were to wish for anything, I should wish for more time.” This sentiment resonates deeply with the findings; the time expended on these workarounds, this ‘invisible labor’, highlights the gap between a system’s theoretical design and its practical application. The research underscores that anticipating these points of friction – understanding where the system will demand more from the user – is crucial for building truly robust and effective tools.

Beyond the Fix

The demonstrated reliance on user-driven workarounds suggests a fundamental miscalibration in how generative AI is often conceptualized and deployed within established institutions. To treat these adaptations as mere deviations from ‘intended use’ obscures the reality: these are the very acts of integration, the quiet calibrations that bridge the gap between technological potential and organizational inertia. The challenge, then, is not to eliminate workarounds – a futile proposition – but to recognize and account for them. This necessitates a shift in perspective, moving from a focus on system features to an analysis of the complete sociotechnical ecosystem.

Further investigation should move beyond documenting these ‘fixes’ to understanding their systemic effects. What are the cumulative costs of this invisible labor? How do these adaptations subtly reshape organizational structures and power dynamics? Are certain workarounds systematically favored or suppressed, and with what consequences? The study of generative AI cannot remain solely within the domain of computer science; it requires a sustained engagement with the humanities and social sciences to fully grasp the intricate web of trade-offs inherent in its implementation.

Ultimately, the persistence of workaround solutions reveals a deeper truth: elegance in complex systems rarely arises from radical innovation, but from incremental adaptation. The pursuit of seamless integration should not prioritize technological sophistication, but rather acknowledge the messy, contingent nature of practice. The future of AI within organizations may well depend not on building ‘smarter’ tools, but on cultivating a greater understanding of how humans, inevitably, make them work.

Original article: https://arxiv.org/pdf/2512.21055.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-26 16:15