Studying AI in the Wild: Reflections from the AI @Work Research Group
反思了AI@Work研究小组如何通过关系民族志方法，在真实工作场景中研究人工智能对知识工作的影响，旨在破除技术决定论迷思，为管理学者和实践者提供更贴近实际的洞察。
Long before AI entered the public spotlight, the AI@Work research group was already studying how algorithmic technologies change work. When the editors of this special issue invited me to reflect on my experiences, I saw this as a valuable opportunity to articulate our often tacit way of doing research – to make explicit the ontological commitments, methodological choices, and collaborative practices that have shaped how we study Artificial Intelligence (AI) in the wild. I see the AI@Work research group as a distinct school of thought, with a shared ontology, methodology, and epistemic culture that helps us study (and teach) AI in the wild. Since the group is more than just a collection of individual projects, I have chosen not to describe each project in detail. Instead, I use this opportunity to articulate our collective mission: to demystify popular beliefs around AI by analysing the actual changes that occur when AI systems are developed, introduced, and become embedded in everyday work practices. Management and organization scholars have paid insufficient conceptual and empirical attention to AI in the wild. Despite decades of critique, technological determinism has resurged in AI discourse, perpetuating several interrelated problems: treating AI as a pre-given force whose design remains black-boxed; employing methods that abstract work from the practices where it unfolds; and relying on snapshot analyses that miss how AI’s effects emerge over time. These limitations are not merely methodological; they reflect deeper ontological assumptions that leave us poorly equipped to understand AI’s emergent, relational character. Against these dominant approaches, the AI@Work school of thought advances a relational perspective operationalized through relational ethnography. By scrutinizing if, when, how, and why AI reconfigures work and its surrounding social structures, we seek both to advance academic understanding and to support more informed organizational decision-making, before and after adopting (or declining) AI. This essay presents examples from our research on AI development and uses them to illustrate that claim. Central to our perspective is our use of a relational ethnography that helps us reveal and theorize how AI reconfigures knowledge work within specific organizational settings while also generating insights that are transferable across multiple cases. Moreover, how we study AI is deeply intertwined with how we organize our research. In the second part of this essay, I will elaborate on the AI@Work research group’s epistemic culture, one that centres on collectivity and ultimately strives to provide value for academia, industry, and society. I want to stress already at the start of the essay that the AI@Work school of thought is grounded in a larger academic discourse on technology and knowledge work that uses a relational perspective and practice theory to study technology and organizing as mutually shaping phenomena (e.g., Anthony et al., 2023; Bailey et al., 2022; Faraj and Leonardi, 2022; Glaser et al., 2021; Lebovitz et al., 2022; Scott and Orlikowski, 2022, 2025; Suchman, 2023). None of our research that I will discuss in this essay would have been possible without the support of an extensive network of outstanding scholars from around the world – many of whom visited our group and/or hosted our researchers. To understand why technological determinism persists in AI discourse despite decades of critique, we must recognize that the current AI hype is not new – it is the latest iteration of recurring promises to automate knowledge work. Each wave follows a similar pattern: new technologies are marketed as finally capable of capturing expertise, optimizing work, and eliminating reliance on human judgement. Yet these promises consistently fail to materialize as predicted, not because technologies are ineffective, but because they rest on flawed assumptions about the nature of knowledge, work, and change. The rapid pace at which new technologies are developed and introduced fuels recurring promises, creating a ‘hype mill’ that mainly benefits industry. Many of these hypes fade quickly, but the one around AI and work has proven more persistent, most likely because of its wide applicability and the significant investments already made. As Bailey and Barley (2020) put it, ‘the train is already running’. Tracing this history reveals why fighting hypes through grounded empirical research has become central to the AI@Work school of thought. To do so, let me take you back to my early years as a junior researcher. Like a modern-day Don Quixote,[1] I set out on my journey of fighting hype-mills in the field of knowledge, work, and technologies. My first fight was against the knowledge management (KM) hype. KM systems were marketed as tools capable of converting individual expertise into organizational memory. This was based on the belief that knowledge could be codified, stored, and moved around much like information. Yet, our empirical studies revealed that such systems rarely functioned as envisioned (Huysman and De Wit, 2013). As the ethnographers at Xerox PARC in Silicon Valley had already shown, knowledge sharing often occurred informally through social interaction rather than through centralized (technical) systems. Their workplace ethnographies not only challenged dominant narratives that treated information as context-free and technology as a neutral tool for control and optimization; they also inspired our own research by demonstrating the value of close, situated study.[2] Their work contributed not only to academic debates but also shaped more grounded understandings of what Brown and Duguid (2000) called ‘the social life of information’ and what new technologies can realistically achieve, given the contextual, social, and negotiated nature of knowledge. One frequently cited illustrative case involves Xerox customer service repairmen and the introduction of a centralized database meant to standardize and improve access to information for field technicians (Brown and Duguid, 2000). The database was intended to increase efficiency by providing technicians with official documentation and repair protocols. However, technicians often ignored the centralized system, finding it too rigid and poorly matched to the realities of on-the-ground troubleshooting. Instead, technicians relied on informal storytelling and peer-to-peer knowledge exchange, such as during coffee breaks or shared rides, to solve complex machine problems. These practices were essential to maintaining technical competence. But the recognition that knowledge is socially embedded and resists codification not the for the way for the management hype I to of The of was developed by and at Xerox to how by through in the of and The its in an study of in than context-free information in the through in practice – and on knowledge was embedded in the social, and how was negotiated and how tools and were challenged the of knowledge as individual and because this why knowledge resists sharing and the was as a tool to informal knowledge Yet our research revealed that are frequently shaped by and from the had for (e.g., et al., et al., when are on to knowledge The hype around and social the for similar knowledge practices rigid and control (e.g., et al., the KM hype the belief that knowledge could be and This is by the current hype surrounding tools to support knowledge work, such as algorithmic management and that are as the in knowledge, eliminating reliance on (e.g., et al., to and tools to human and AI knowledge as mainly in the of the Xerox PARC scholars already us that knowledge is socially and in research on systems has the of expertise (e.g., The not only in the of capturing individual as many AI but more in the social nature of knowledge, which resists codification without (Brown and Duguid, practice perspective on knowledge attention to the shared and that assumptions about what as knowledge, that the This also for the AI and and which are for understanding how AI systems are and in organizational AI can us from it also on to us In the latest AI tools can automate our tacit knowledge – what as more than we can for in et al., in tools reveal to et al., and provide with by These rest on of AI. AI systems train on to – developed with reliance on that knowledge. are to to knowledge and As I will with examples from our these how work is with AI. of work must the own treating it as an whose and often et al., can and Yet, popular and academic discourse rarely see these as Instead, AI is treated as a with the specific nature of AI technology that could to tool work – or This both and we understanding or advance knowledge about AI’s distinct when our insights are our become too to for AI or this us to what AI In methods to study AI in the of work work as a collection of studies (e.g., and and and use to without how or why work changes on the the of the research studies of changes in work (e.g., et et studies and based on individual how social AI’s fail to with the and nature of AI systems – that and often second or This is by methods – and – that work into them from the everyday social and practices where they and research often the of AI as a to the This the more of how AI is the nature and organization of work. These reflect deeper ontological assumptions that and and that from technology to To such the AI@Work school of thought advances a relational that on AI systems through and because the work practices into which they are embedded are in is and often than treating this as a we use it as a it to new and To for we have a relational and and not as but as that take through In this are they through rather than As I will with examples from our this that we AI systems not as that through and organizational we work not as a collection of individual but as of practices by human and this the of AI’s and can work in that human and To these without them to we to practice In the of the Xerox PARC practice theory has proven a for analysing how technology and work become in work and and to what effects it can et al., theory work not as individual or but as and through which AI and human and knowledge work over et al., 2025; Scott and Orlikowski, In where this practice theory do collective practices of treating individual (or as the of practice theory us to work practices as the of I often use our study of et al., to the value of collective practices as the central of much of the discourse around within and on the and with the and our this is too In the is at the but with the a and to the The the of the and the of the from the how changes in practices during the collective work practices. and were to tacit of in on-the-ground new of and the most of was not about technical but about how the as a its practices in to a work the case at the also more technologies can work when they are not by the work as part of a this changes that because AI is in the not because an individual uses design and practices in to use of AI they do not use these tools the we to study changes in work a relational relational is the most our early work on technology relied on case studies and which valuable but in access to practices. to that studies of specific and that how technologies are and in However, ethnography is to how AI systems are shaped by and have and effects across multiple and it the ethnography these by and across settings and over practice changes to organizational and helps the study of rather than an to complex phenomena like where and are across and et al., 2023). relational ethnography to study AI at work helps to changes across multiple – from development through to use – rather than studying a it rather than capturing it insights across organizational settings to and it where human and technological rather than treating them as In the of this essay, I the value of our relational through empirical I how studying AI development reveals the of assumptions and into and how studying AI use over to or snapshot To understand AI’s at work, we to not only at how use a system, but also at how that is The work of a central often in shaping how AI technologies et al., make about which to train how and how are to These and into the system, of work and AI tools are that reflect the of work practices in helps why and how systems are and are not (e.g., in an study at a et al., we the practices of a One of the was that the and the for a tool in the a to with the of whom in and work as a of Their as them to machine systems that the after the and back to the could not the to use and systems that one was systems in are of machine In on examples to or that group or without and that a through on These because they to design and of work around To be more explicit during and significant human in and not only from the of the but also from In an study at a and we the an AI and the et al., while a for This over in which expertise was from the AI but as and to and both the AI and to of like and to AI not what was a they they could a AI tool that was for this specific organization without the and knowledge of for AI to train the system, were to and on current first had to the which they by with the and with current and to This development and shaped both the and to train the to and (e.g., in them to with through several the and were This case that AI’s is already in the of because the organization an and the had to (e.g., and and the over work the and its in As systems provide not it to these for in a study at a we how an AI to et al., it was to the that the and to that could the recognize with on customer and the were not to understand why a given and a that not the or and for but such as the an and a of to for and (e.g., through interaction and The an the a or and its over it that case in a and service et al., that more than AI approaches, and to the a that not to from a on to a set of that the could to This challenged to also a of on the The latest in AI that and that or on with in from human as in is capable of generating and to which fuels rather than and how is developed from because can be during for through and the and can in to studying how set and how and are and how everyday work. and is and must be as AI reconfigures work with the into which it is and Orlikowski, 2025; Suchman, that often is by the nature and by the changes over that we often This research with research and Orlikowski, of how research helps to is the case of the of the algorithmic et al., several we the of a machine to support where to intended as neutral of algorithmic had or to or However, to the of the and a of understanding both and these on more the study how moved through with and algorithmic with own informed These from a of for the and a in the over time. an in our it is often the that us to deeper for possible and et al., By embedded in for a of we are to into for and – – theory when we we and to the field to see which as practices or in an study of the introduction of an AI in a we were by by what we which us to the case in The a in an AI and to automate the of this work was by through deeply tacit in to and this expertise was as informal and to work, and a for through The only over was the that the AI not it work. To the developed new practices from to and to individual for and with tacit knowledge. they the AI’s and the of and the could not and insights back into its the case us to the and a the introduction of AI into organizational and of machine whose were and When we study how work practices change over we also study the often that and these practices the of systems. Scott and to this as the a current that reconfigures work in and often not just how work is but also how, for control and are through embedded algorithmic These often into work practices. in an early study of algorithmic in a we of of what as knowledge, the et al., with algorithmic a of and management to the of This change and it to or without a perspective to both and deeper epistemic more of that into work from our study of how to are negotiated and into et al., on we that is not a but work. on AI and have often one of the one is treated as a technical that can be and embedded in through design and et al., et al., the is as the of social and shaped by the and in practice 2022; approaches, technological determinism or the social, the of and In an study of the of AI for at the that I to illustrate during AI we to an by how was and through the of and the of AI systems. as a neutral efficiency the AI tool on a its to and to on a as that because the each to the it a of that had been to in the the as and when with the of both the tool and own in this a of what meant within the The by to the of the AI to while understandings of (e.g., by or by were was not into the AI it was through the and the algorithmic that an AI is rest on than on of that one of while and The case that AI not just but the what as or is not it is during use as and the a studying at work for a than studying AI. than studying the from development to to what the tool in practice is while are In our latest research on use work at an we and as they in et al., The work before moved through that at the and the through of and The that out are already shown, they to the as the work must with this early and the new work that and and or The is into a and the this ethnography an early of changes with it is only a and I to inspired by the studies that many will In the rest of this essay, I discuss the AI@Work research group’s shared practices and through which we knowledge. This is and from most research that value individual expertise and I this specific relational epistemic culture as to the AI@Work school of thought. The way we study AI at work is deeply to how we organize our research. Like our and methodology, I like to describe our epistemic culture, by as the set of practices and that and knowledge in a as a relational epistemic the shaped by the larger the AI@Work research group’s has around practices and such as sharing and with a on individual and As our is the on which we do let me first discuss our of access to these and how we train to study both AI in the as as AI in both AI development and its use can with the technical design as as with the in which the AI tool is studying how AI the field of more than an understanding of the knowledge practices central to a it also for knowledge of and and is to within the of most research projects, and these are than from in a specific work we have an with knowledge one of our a is research on in we to in AI This them with a understanding of technical such as the of and the to the and the of and is how we research cases. research access is a and our through with we use are or more for which the value of with industry. As the is with adopting AI the has been more and mutually the an opportunity and an from a of to that we on junior a this that we are on the for cases. to that reveal change to the introduction of provide more on how AI is work practices than studying or work practices. we to rather such as AI at the and at the our on work practices that are and with the use of AI. the of it is essential to be to the that in it for The most is the benefits of a perspective that with our relational ethnography. this not only the of AI tools but also such as and it insights that to are also more such as on such as for or for also take on to an us to within the of for that are central to us the to our work. while studying how technologies in work, our by during in a project the of a system, contributed by in the back In many research with to mutually knowledge over One is our research with one of the in the in and in practice Since the first interaction in this has hosted a research for several and many As a the research group and the a where we with AI use of technology and work, and in The is into a where and social (e.g., AI use with them in and discuss the for own practices. to how we do research is our collective of sharing and we our research as and essential to with the and nature of the world to make of it and a This is what to as an to with without for and is when in an that on empirical and grounded by and as collective where often they are in a of can These from a the of our own and and with the and Moreover, the us to how AI reconfigures work across This collective and of across for the development of the of individual to and in the of for such et al., we our both during the and after the research is This us to across cases. during our we across by such a similar AI technology in to the of work, while in it in these by the design of the technology or by specific organizational By with such we can that how AI reconfigures work. these us to and across providing essential for of this collaborative is in our on AI et al., The is the of a collective of individual projects, us to from to we several recurring that us to based on One is the to when and in the case of machine with the to and the of its and also a belief that expertise through sharing and that by one we the from during or to with and and to AI@Work we informed about each The sharing centres mainly around the of much more than about the This for in on our where you can the about As with our research practices are our use of a shared as et al., and us informed of what project are The can be about our This helps us informed of each and sharing the on the many and each the we see also organize research to the and such as the on AI@Work and our where are invited to with technologies. The of these can be when the to own is not for are often by this mainly the – such as the to expertise, the it or the of a of knowledge. most often they to but it also that leave the our culture of collectivity and sharing it also with a it that only in the while and often force us to on the the collectivity and individual and we to own research by with and our group to increase us of our assumptions and against the of an This is for by research scholars to our in and with This also the to have developed in such as and AI and and work. The AI@Work school of thought advances a relational that AI and work as through This methodological study collective not individual not through and not in a These methodological why snapshot and to miss how AI work in the wild. adopting a relational of AI and work a treating AI as the latest technology in for shaping its Instead, must during to how practices during and for effects across organizational by AI’s relational can to of work in This work, be in like the that the attention during the I like to see the epistemic culture of the which in the AI@Work research as one on In are not to rigid they across the for one and to the of the The of this in the shared understanding of the where each can and the of This how we relying on the of the to support the research where which is grounded on a collective for the of the individual and that of the AI@Work group in each of us distinct expertise, we are not into as to understand both the of the and the of our research on understanding of each work. To part of the AI@Work school of thought is not only but also deeply where the research will take but to the journey studying AI in the do this as we each to be to and new is one I have it is that research is but a it is a deeply relational for I I of the AI@Work research group for and our school of thought, in I also want to the editors of this special issue for valuable on of this