Abstract
The advent, and large-scale adoption, of artificial intelligence in management learning and education poses complex questions about the discipline, its pedagogy and institutional contexts. Despite its rapid evolution and uptake, we still know little about how academics themselves experience – and shape – the transition to ‘educating with the machine’. Guided by microfoundational perspectives, this study asks how an external artificial intelligence shock travels through educators’ cognitive–emotional framings, resource landscapes and assessment practices to generate artificial intelligence–aware institutional policy. We draw on 30 semi-structured interviews with academics at all career stages in a UK business school. The findings trace a four-stage macro–micro–macro pathway starting with (1) generative artificial intelligence as a disciplinary macro shock; contributing to (2) different cognitive–emotional framings at the micro level; revealing (3) uneven, micro-level resource landscapes, horizontal interactions and artificial intelligence-enabled assessment practices; and leading to (4) norm disruption at the macro level. Those stages are supported by four linkage mechanisms: cognitive framing (L1), affective response (L2), resource-to-practice translation (L3) and norm reinforcement/disruption (L4). We conclude by outlining the study’s empirical contributions to microfoundations scholarship and to debates on artificial intelligence in management learning and education.
Introduction
If there was the universe according to me, I would just unplug it and go back to a time when students had to learn how to write and think for themselves . . . But realistically, there’s no going back. We can’t put the genie back in the bottle. It’s here, and we need to figure out how to work with it, not against it. That means rethinking how we teach, how we assess and, honestly, how we define learning. (Sandra, female, senior lecturer)
Sandra’s reflections on the challenges of educating with artificial intelligence (AI) are representative of current academic debates, where the professional impact of large-scale advancements in emerging technologies is a matter of ongoing interest. This is a topic familiar to Management Learning readers as well. In an editorial piece, Barros et al. (2023) foresee AI redefining not only all aspects of academic practice – teaching, learning, research and administrative duties – but academic identities as well. Krammer (2025) extends this by anticipating that AI will have a ‘creatively destructive’ impact on both the business school as an institution and management learning as a discipline. More recently, Molesworth et al. (2025) use science fiction movie tropes to imagine a series of benign and dystopian outcomes and call for a more nuanced understanding of AI as a future ‘coworker’.
It is, as yet, unclear what the future may look like, but an increasing number of perspectives suggest that it is likely to de-centre the human manager and move towards post-human conceptualisations of management learning (Butler and Spoelstra, 2025). Such arguments are not entirely new. Scholars of critical pedagogy will detect echoes of the work of Brazilian educator and philosopher Paulo Freire, whose pedagogical thinking has informed subsequent developments in ecopedagogy, even if Freire did not use that term himself (Jandrić and Hayes, 2022). As a systemic approach, critical (eco)pedagogy advocates going beyond human rights and anthropocentrism, anchoring teaching and learning in biocentric and ecocentric ethics (IndelliCato, 2021).
We are mindful of such important perspectives and acknowledge the environmentally embedded nature of machine learning through AI’s reliance on energy and natural resources (UN Environmental Programme, 2024). Yet, as management educators, we are faced with the need to work with AI and negotiate our professional positionality towards it, while observing our institution’s changing policy frameworks. The AI, as a case study, also offers an opportunity to address a gap in the existing literature: existing management learning and education research has documented how individual academics and students respond to generative AI in the classroom but has not yet explained the mechanisms through which those distributed, emotionally charged micro-level responses coalesce into institutional assessment policy.
To address this gap, we adopt a microfoundations (MF) approach as a way of unpacking the multiple levels at which organisational phenomena occur (Felin and Foss, 2005; Foss and Linder, 2019). The MF scholarship recognises the interplay between individual agency and institutional structures and examines how actors’ cognitive, normative and emotional frames both shape micro-level interpretations of institutional demands and inform macro-level structural changes (Barney and Felin, 2013; Felin et al., 2015). Business schools are complex professional organisations with diverse staff roles (academic, administrative and managerial). This complexity allows for a nuanced exploration of individual-level behaviours, decisions and interactions – the core concern of MF theory.
Drawing on 30 semi-structured interviews with UK business school staff members at different career stages and in different roles, the study asks how a sector-wide AI shock travels through educators’ cognitive–emotional framings, resource landscapes and assessment practices to generate AI-aware institutional policy. The study contributes primarily to emerging debates on generative AI in management learning and education, using a microfoundational (MF) lens to explain the mechanisms through which educators’ responses become institutionalised as policy. It outlines the mechanisms through which institutional AI governance goals are filtered, and sometimes distorted, by locally situated interpretations and emotional responses (Frödin, 2024; Voronov and Weber, 2020). It also demonstrates that individual interpretations are driven by both affective responses and rational–cognitive sensemaking. These dual influences reveal why AI governance in academic settings is complex and why nuanced, emotionally attuned strategies are needed to connect institutional intentions with lived experiences.
The rest of the article is structured as follows. We first provide an overview of debates in the MF literature and scholarship on AI in management learning and education, before outlining our institutional context and methodology. We then present our findings, discuss the study’s implications for debates on AI in management education and MF scholarship, and conclude with a consideration of limitations.
Literature review
AI and management learning
Since late 2022, a rapidly expanding body of work has begun to examine the implications of generative AI for higher education in general and management education in particular. Early contributions by management educators framed generative AI as a fundamentally paradoxical phenomenon – simultaneously a threat and an opportunity – and called for critical yet pragmatic engagement rather than simple prohibition (Lim et al., 2023; Ratten and Jones, 2023). Recent editorials in the Academy of Management Learning & Education and Management Learning journals have signalled that generative AI is now recognised as central to core concerns about learning, knowing and critical thinking in business schools; AI is not merely a technical or integrity issue (Izak et al., 2025; Larson et al., 2024).
Building on this normative work, a second wave of empirical studies in management education journals has begun to document how staff and students are actually integrating generative AI into their teaching, learning and assessments. Mixed-methods and case studies explore, for example, how management educators and students negotiate the meaning and use of generative AI in the classroom: Gupta et al. (2024) find that these negotiations are contested, messy and shaped by preexisting disciplinary assumptions, with neither educators nor students arriving at stable or consensual interpretations of AI’s role. Sharma (2025) shows that productive AI adoption among management educators is mediated by unlearning – educators must first dismantle prior assumptions about knowledge transfer and student effort before they can integrate AI constructively into their pedagogy, and this process is slow and uneven. Fischer et al. (2024) document how postgraduate management students mobilise AI in written assessments, finding that assessors’ expectations are themselves being reshaped in response. Complementary research on AI literacy and classroom interventions in business and marketing education further highlights the need for structured support: Beninger et al. (2025) demonstrate that targeted interventions can move students beyond purely text-generating uses of AI, but only where institutional backing is sufficient.
In parallel, work from business schools in the United Kingdom shows that generative AI unsettles long-standing assumptions about academic integrity and the protective power of ‘authentic’ assessment (Kofinas et al., 2025). Also, broader higher education analyses point to generative AI as a potential driver of curriculum reform and highlight its ability to reexamine learning outcomes (Lee and et al, 2024; Ma, 2025; O’Dea and Kremantzis, 2024).
To summarise, the emerging scholarship establishes that generative AI is not simply another educational technology; it is a catalyst for rethinking what counts as learning, evidence and academic work. However, the studies reviewed above document what is happening – in classrooms, assessments and individual practices – without fully explaining how those distributed, emotionally charged micro-level responses aggregate into institutional policy. That is the explanatory step this study takes. By tracing the MF of AI-related policy change in a UK business school, we explain the mechanisms by which academics’ situated responses to generative AI coalesce into AI-aware assessment practices and formal institutional policy – a question the existing literature raises but does not yet answer.
MF and the micro–macro link
Our use of an MF lens builds on a long-standing concern in sociology and organisation theory with the micro–macro link: how macro-level structures both shape and are reshaped by individuals’ situated actions. The MF debates in strategy and organisation studies explicitly draw on James Coleman’s macro–micro–macro bathtub model and the tradition of analytical sociology. Scholars argue that macro phenomena, such as organisational capabilities or institutional arrangements, should be explained by specifying how they influence individual-level situations, perceptions and resources and how those actions subsequently reaggregate into macro patterns (Barney and Felin, 2013; Coleman, 1990; Felin et al., 2015; Felin and Foss, 2005; Foss, 2016).
In institutional theory, the MF turn emerges in a field already heavily shaped by Giddens’ structuration theory and the associated structure–agency debate (Barley and Tolbert, 1997; Giddens, 1984). Barley and Tolbert’s seminal work on institutionalisation and structuration conceptualises institutions as scripts that are enacted and modified over time through situated action, emphasising the recursive links between action and institution. Recent scholarship on the MF of institutions builds on this structuration-inspired tradition, combining it with Coleman’s multilevel schema. It calls for models that articulate how institutional logics, rules and symbols are enacted, contested and reproduced through everyday cognition, emotion and interaction and how these micro processes feed back into institutional structures (Haack et al., 2019; Harmon et al., 2019; Zilber, 2020).
Our study follows this line of work. We adopt a Coleman-style macro–micro–macro framing because it offers a clear process architecture for tracing how a sector-wide technological shock (generative AI) is interpreted and enacted by individual educators and how their micro-level experiments, emotions and interactions crystallise into AI-aware assessment practices and institutional policy (Coleman, 1990; Felin et al., 2015; Felin and Foss, 2005). At the same time, we treat these MF as fundamentally structuration-compatible: educators are not passive carriers of structure but knowledgeable agents whose actions recursively reproduce, and sometimes disrupt, institutional arrangements (Barley and Tolbert, 1997; Giddens, 1984; Harmon et al., 2019; Zilber, 2020).
During the past two decades, this MF turn has gradually emerged in the fields of strategic management and organisation studies, seeking to ‘open the black box’ of aggregate, macro and meso organisational constructs and expose the human activity that underpins them (Barney and Felin, 2013; Felin and Foss, 2005). Whereas traditional approaches often treat organisations and institutions as self-contained entities, MF approaches suggest that institutional structures emerge from individual cognition, motivation and interaction, which in turn shape, by either supporting or disrupting, the overarching institutional contexts (Felin et al., 2015). Coleman’s (1990) bathtub model (see Figure 1) is frequently used to visualise these multilevel processes. Seen as a time-sequenced path diagram (Abell and Engel, 2018), the model juxtaposes two macro nodes with two micro nodes and connects them by four causal arrows. In this way, it depicts how sector-wide institutions, governance regimes and technological shocks constitute a macro starting point that exerts downward pressure on individuals by reconfiguring what is possible, desirable and thinkable (Kozlowski and Klein, 2000).
A representation of Coleman’s bathtub model.
In previous studies adopting this approach, scholars have typically pursued two primary analytical starting points. Most of the early work equates micro with individuals (Bridoux et al., 2017; Felin and Hesterly, 2007; Yao and Chang, 2017). The second stream of work follows what Felin et al. (2015) call the ‘micro-foundations-as-levels’ lens, decoupling higher-level constructs into successively lower ones. These later studies consider interorganisational networks as a layering of firm traits (Kim et al., 2016) or link team processes to organisational outcomes (Baer et al., 2013; Wilden et al., 2016). This heterogeneity extends to theoretical assumptions. Some authors adopt rational-choice premises (Felin and Hesterly, 2007), others foreground bounded rationality (Pentland et al., 2012) or cognitive heuristics (Bitektine and Haack, 2015; Helfat and Peteraf, 2015). Accordingly, MF operates less as a single theory and more as a heuristic toolbox compatible with economics, psychology and sociology (Foss and Linder, 2019).
The MF literature also recognises that organisational phenomena are multilevel: individuals nested in groups, groups in organisations and organisations in institutional fields (Molloy et al., 2011). This multilevel flexibility is precisely what makes MF useful for understanding sector-specific challenges and, specifically, why ostensibly similar digital reforms in higher education produce starkly different results across departments and institutions (Feng et al., 2025).
Debates, linkage mechanisms and gaps
Despite its explanatory utility, MF critics argue that it adds little new theoretical insight, merely rebadging concepts long familiar to organisational behaviour and psychology (Devinney, 2013; Hodgson, 2012). By foregrounding agency, MF theory seems to neglect the structuring force of social context, thereby risking reductionism and methodological individualism (Barney and Felin, 2013; Jepperson and Meyer, 2011). Echoing some of these reservations, Foss and Linder (2019) note that the theory’s focus on individuals invites unfavourable comparisons with established microfields and multilevel research. In response, they and others move beyond a simple focus on individuals to emphasise the scope for micro-to-macro translation as three interlocking logics – vertical (micro acts that cumulate upwards), horizontal (peer-level collaboration or resistance) and feedback (new macro structures reshaping subsequent individual actions) (Foss and Linder, 2019). Untangling these intertwined pathways has allowed researchers to layer multilevel models with longitudinal ethnography and agent-based simulation to observe how micro events first coalesce and then, at times, rebound along the Coleman bathtub (Kozlowski and Klein, 2000).
The advent of generative AI opens a notable gap in this literature. Studies reviewed in the above section ‘AI and management learning’ establish that generative AI is reshaping assessment, academic integrity and learning practices and that both cognitive and emotional responses shape how educators engage with it (Gupta et al., 2024; Kofinas et al., 2025; Sharma, 2025). However, those studies document what is happening at the classroom and departmental level without explaining the mechanisms through which those distributed, emotionally charged micro-level responses aggregate into institutional assessment policy. That explanatory step requires a theoretical language that can connect individual cognition, emotion and resource use to macro-level institutional outcomes across multiple levels simultaneously. MF theory provides precisely that language. The four linkage mechanisms we introduce – cognitive framing (L1), affective response (L2), resource-to-practice translation (L3) and norm reinforcement/disruption (L4) – map directly onto the explanatory gap the AI and management learning literature leaves open. They specify how the contested, emotionally intense and resource-uneven responses documented by existing studies eventually crystallise into the assessment routines and formal policies that existing studies observe as outcomes but do not explain as processes. By applying this MF lens to a UK business school case, the present study empirically explores this gap; we move to discuss our methodological approach next.
Methodology
Research design
This study employs an inductive qualitative design to examine how a sector-wide AI shock travels through educators’ cognitive–emotional framings, resource landscapes and assessment practices to generate AI-aware institutional policy. Qualitative inquiry is well-suited to unpacking complex social phenomena and capturing the lived experience of participants (Creswell and Poth, 2018); we used semi-structured interviews to gather multiple perspectives and generate rich, contextual insight (Braun and Clarke, 2006).
Participants and sampling
The study involved business school educators as the primary participant group. The sample comprised 30 staff members who participated in individual interviews. None of the participants had received formal – institutional or otherwise – training in large language models (LLMs) at the time of the research, but they had various levels of experience through individual experimentation. Furthermore, although we included participants at different career stages and different levels of seniority, we were unable to secure interviews with members of the senior leadership team or staff in governance roles linked to AI policy.
A purposive sampling strategy was employed to ensure representation across different roles, experience levels and organisational contexts (Patton, 2015). Purposive sampling is particularly appropriate for qualitative research where the goal is to identify information-rich cases that illuminate the research questions rather than achieve statistical representativeness (Miles et al., 2020). Teaching staff participants were selected to represent diversity in academic disciplines, career stages and institutional positions. A participant profile is presented in Table 1.
Interview participants.
All the participants were employed in a UK business school at the time of data collection, with roles spanning education-focused and research-focused contracts, leadership responsibilities and early- to late-career positions.
Data collection methods
Participant interviews were conducted between January 2025 and July 2025. Semi-structured interviews were chosen to allow for in-depth exploration of individual experiences while maintaining consistency across interviews through the use of an interview guide (Brinkmann and Kvale, 2015). This approach provided flexibility to probe deeper into areas of interest while ensuring coverage of key topics across all interviews (King and Horrocks, 2010).
The interview guide invited the participants to reflect on their initial encounters with generative AI; how they made sense of its implications for learning, assessment and academic integrity; and the resources and constraints they experienced (time, tools and guidance). The interviews lasted between 30 and 60 minutes, were conducted via video conference, were audio-recorded (with participant consent) and were transcribed verbatim.
Data analysis
The data were analysed using the Gioia method’s three-stage procedure, which is designed to move systematically from the participants’ own words to theoretically meaningful concepts (Gioia et al., 2013). In accordance with the first stage, every transcript was subjected to open, informant-centric coding, producing a broad set of first-order codes that mirrored the language used by the interviewees. During the second stage, these first-order codes were compared, contrasted and clustered into researcher-centric, second-order themes that captured deeper patterns and relationships across cases. Finally, in the third stage, related second-order themes were distilled into a smaller number of aggregate dimensions, yielding a data structure that linked concrete observations to abstract concepts.
This iterative procedure allowed the analysis to remain grounded in the participants’ lived experiences while building a coherent theoretical explanation of how generative AI is reshaping assessment practices and academic identities and how these micro-level processes connect to emergent institutional policy. The initial coding was conducted independently by two researchers in the team, with codes then compared and discussed to ensure consistency and interpretive rigour (Nowell et al., 2017).
In the resultant data structure, some second-order themes are connected to more than one aggregate dimension. We made this choice deliberately to reflect the multifaceted character of the practices and framings we observed and to acknowledge that the change process itself is inherently messy, non-linear and overlapping. Forcing each theme into a single, discrete dimension would have implied a clarity and ‘sequentiality’ that was not present in the participants’ accounts, so we opted instead to preserve this empirical ‘messiness’ in the representation of our findings. A sample of participant quotes and respective first-order codes is presented in Table 2.
A sample of first-order codes.
Following our initial analysis, first-order codes were grouped into potential themes, which were then reviewed and refined through an iterative process. The analysis moved from descriptive to interpretive levels, seeking to understand not only what the participants said but also the underlying meanings and implications of their responses (Patton, 2015; Saldaña, 2016). The resultant Gioia table connecting first-order codes to second-order themes and aggregate dimensions is provided in Figure 2.
Data coding structure.
Use of an LLM in reflexive checking
After completing our initial round of Gioia-style coding – including the development of first-order codes, second-order themes and aggregate dimensions – we conducted an additional reflexive check using an LLM (ChatGPT). We undertook this additional step to enhance reflexivity and consistency checking in a context where rapid AI-led change is itself part of the phenomenon under study. However, the LLM was specifically used as an auxiliary heuristic rather than a substitute for the researcher’s reasoning. Consequently, by the time we engaged with the LLM, the coding was already stable enough to be represented in a draft data structure table. We exported this table into a text-based format and provided it to ChatGPT, together with a short description of the study’s aim and the logic of the Gioia method.
We employed prompts that focused specifically on the relationships between codes and themes rather than on generating codes from raw data. For example, Here is a data-structure table with first-order codes, second-order themes and aggregate dimensions from a qualitative study of academics’ responses to generative AI. Please review the links between second-order themes and aggregate dimensions. Where do the links appear redundant, overlapping or misaligned with the theme labels? Looking only at the labels, are there first-order codes that could sit more clearly under a different second-order theme without changing their meaning? Please explain your reasoning.
The outputs from ChatGPT consisted of narrative comments on potential redundancies (e.g. two aggregate dimensions with similar scope), suggestions that certain themes might be more clearly distinguished, and questions about ambiguous labels (e.g. where a second-order theme seemed to mix beliefs and practices). Crucially, the model did not have access to the underlying interview transcripts at this stage; it operated only on the abstracted data structure that we had already produced.
Evaluating and acting on LLM feedback
The research team then evaluated ChatGPT’s suggestions in a separate round of discussion. We treated the LLM’s comments as prompts for our own reflexive scrutiny rather than as recommendations to be followed automatically. For each suggested change, we went back to the underlying first-order codes and illustrative quotations and asked (1) whether the proposed relabelling or regrouping would be consistent with informant language and our interpretive notes and (2) whether the suggested change would improve the clarity or parsimony of the structure.
In several cases, the exercise confirmed our existing coding decisions. In a small number of instances, we adopted minor changes, such as: (1) renaming a second-order theme to better reflect the underlying first-order codes, (2) splitting a broad second-order theme into two more precise themes, and (3) moving a single first-order code between neighbouring themes when this alignment better matched the quotations.
We did not act on suggestions that would have relocated codes in ways that would conflict with interviewees’ wording or our prior analytic memos. We therefore treated ChatGPT as a supplementary tool for reflexive checking, not as a substitute for human coding or interpretation. All of the codes, themes and dimensions were initially developed, and ultimately confirmed, by the research team, with the LLM used only to surface potential inconsistencies in the already constructed data structure.
Findings
In line with Coleman’s (1990) macro–micro–macro model, we organise our findings into four stages and four linkage mechanisms.
Stage 1: Generative AI as disciplinary shock (macro conditions)
Consistent with the broader field of management education, the business school in which our participants were based experienced a sudden and significant disruption with the public release of ChatGPT and similar LLM tools in late 2022. Senior academics spoke of a ‘policy vacuum’, while external bodies issued only tentative guidance. The result was a liminal period in which individual academics had to improvise responses without clear institutional steer: As soon as those conversations started, it was clear that it was going to take months before we’d even have a policy. (Sandra, female, senior lecturer) I just thought it was . . . moving so rapidly . . . I certainly didn’t know where this is [sic] going to end in education. (Andrew, male, senior lecturer)
The macro shock manifested locally in three main ways. First, professional body accreditation teams began to ask how the school would ‘safeguard assessment integrity’ (Sandra, female, senior lecturer; Michael, male, professor). Second, employer advisory boards pressed for graduates ‘fluent in prompt engineering’ (Michael, male, professor). Third, league-table metrics for student satisfaction spurred concern that visible AI bans might provoke backlash (Andrew, male, senior lecturer; Sandra, female, senior lecturer).
Taken together, these forces reshaped the disciplinary landscape and created a sense of crisis: academics were simultaneously expected to protect assessment integrity, demonstrate AI-readiness to employers and reassure students. This external shock, experienced under conditions of institutional ambiguity, set the stage for educators to frame AI in divergent ways and to experience intense emotional responses – the first two linkage mechanisms (cognitive framing and affective response).
Stage 2: Cognitive–emotional framing of AI (micro sensemaking)
L1: Cognitive framing
To make sense of the shifting institutional terrain, the study’s participants drew on their professional values, academic integrity, focus on student development and individual disciplinary tradition to frame AI along an ethical–pedagogic axis. At one end, AI was characterised as an existential threat: My initial response was that it’s just cheating. (Richard, male, professor)
At the other end, it appeared as an amplifier of critical thinking: As an educator, my role is more about teaching what it is to be human rather than what it is to use these machines . . . I think that’s quite exciting. (Thomas, male, senior lecturer)
L2: Affective response
Emotions operated as a second linkage mechanism (L2), sometimes reinforcing and sometimes destabilising cognitive frames. A mix of curiosity, professional pride and anxiety surfaced as educators grappled with the possibility that AI tools could enable students to bypass learning and submit unoriginal work undetected: You feel like it’s a bit overwhelming . . . things are changing too quickly without us reflecting on actual value. (Helen, female, senior lecturer)
For some, these feelings tipped towards excitement and intellectual stimulation; for others, towards frustration and resignation: If there was the universe according to me, I would just unplug it . . . But realistically, there’s no going back . . . (Sandra, female, senior lecturer)
Affective intensity, rather than cognitive rationality alone, proved consequential. Highly anxious staff moved quickly to defensive manoeuvres such as reverting to handwritten or tightly invigilated exams (Lucia, female, lecturer; Michael, male, professor). Those who experienced curiosity and cautious optimism (Daniel, male, senior lecturer; Celine, female, professor) channelled their energy into experimentation.
Together, these accounts illustrate how educators’ cognitive framings (L1) and affective responses (L2) mediated the impact of the AI shock, shaping who saw AI as an urgent threat to be policed, a pedagogical partner to be explored or both simultaneously. These framings and feelings provided the motivational backdrop against which resource constraints and opportunities were interpreted in the next stage.
Stage 3: Resource landscapes, horizontal interaction and AI-enabled assessment (micro practices)
Stage 3 examined how uneven material and social resources, coupled with horizontal interactions, translated the framings and emotions from the earlier stages into concrete assessment practices – the resource-to-practice translation mechanism (L3).
L3: Resource-to-practice translation
Access to time, AI subscriptions and guidance varied sharply across campuses and departments. Some educators described being ‘around the edges’ of AI because of workload pressures and limited institutional support: I would like more guidance . . . we’re very time-poor as academics . . . I’m interested in it, but I think I’m around the edges of using it. (Aria, female, senior lecturer)
Some gained early advantage through informal peer networks, noting the absence of structured support and calling for practice-sharing spaces: I’d love to have something interactive where we can see how a colleague has been using it in their own context. (Javier, male, senior lecturer)
Several lecturers identified the students themselves as a critical, if underrecognised, resource: There’s a kind of digital illiteracy . . . we have a responsibility to train our students to better use it. (Celine, female, professor)
The resource-to-practice translation mechanism was evident in how these disparities materialised as different action sets. Where suitable antiplagiarism software was unavailable, staff translated the constraint into in-person oral exams (Andrew, male, senior lecturer). Where small research grants were offered, they explored pilot projects in AI-assisted feedback (Claire, female, postdoc). In some cases, lack of time or technical support meant that potentially innovative ideas – such as developing cross-module AI literacy pathways or embedding prompt-engineering workshops – remained aspirational and were not pursued beyond informal discussion. Thus, resources were not neutral inputs – they were fields of negotiation shaped by workload models, peer culture and institutional rhetoric.
While official policy lagged, educators turned to one another. Informal lunchtime workshops (Megan, female, lecturer), corridor conversations (Olivia, female, senior lecturer) and online message boards (Patricia, female, professor) became spaces for sharing prompts, assessment ideas and cautionary tales. Horizontal interaction often determined whether an innovation stalled or spread: We go over the AI responses together and make a critical assessment, so we use it to generate a dialogue. (Javier, male, senior lecturer)
Mentoring relationships mattered. Less confident staff were paired with early adopters to codesign assignments (Daniel, male, senior lecturer), thereby lowering perceived risk. Resistance was also collective; in one department, a critical mass agreed to pause AI integration until safeguards were clearer, effectively vetoing a proposed top-down pilot (Martin, male, professor). Thus, horizontal ties both amplified innovation and, at times, consolidated cautious nonadoption.
Against this resource and interaction backdrop, academics began to experiment, redesign and sometimes subvert traditional approaches to assessment. In some cases, this resulted in framing AI as a ‘thinking partner’ and giving students free access to tools; the aim was to enable practices that reinforced higher-order learning goals: Students are asked to submit ChatGPT’s answer and then critically assess it . . . I had students email me to say, ‘that was the best assessment I did because I could prove I knew more than the AI’. (Celine, female, professor)
Other initiatives focused on making student thinking more visible. Richard (male, professor) redesigned coursework to include short video reflections: Yes, the student might use AI to write a script, but it would take more effort to do that than just telling me what they learned.
Defensive tactics also emerged, such as pulling exams back into classrooms and abandoning online multiple-choice tests perceived as ‘a mockery now’ (Andrew, male, senior lecturer). Some of these defensive moves (e.g. handwritten exams) remained local and temporary, often described by staff as stopgaps until better alternatives were found. In contrast, practices that combined AI use with explicit critical evaluation – such as AI-critique assignments, in-person oral examinations and reflective video logs – were more likely to be shared, refined and later proposed for wider adoption.
These examples show how uneven material and social resources, together with horizontal interactions, worked to translate cognitive–emotional framings into specific AI-enabled assessment practices; effectively, the core of the resource-to-practice translation mechanism (L3).
Stage 4: Norm disruption and emergent AI-aware routines and policy (macro outcomes)
Stage 4 traces how the micro-level experiments described in the previous section fed back into disciplinary norms and institutional policy through the mechanism of norm reinforcement/disruption (L4).
L4: Norm reinforcement/disruption
Small-scale assessment trials quickly became catalysts for collective sensemaking. When educators tested their existing assessments against AI, many were struck by how easily tools could succeed: I fed an old exam into ChatGPT and it sailed through. You suddenly see how brittle the system is. (Thomas, male, senior lecturer)
Recognising assessment precarity in terms of AI hijacking prompted reevaluation of what counted as valid evidence of learning. Through teaching awaydays, committee discussions and informal networks, ‘maverick’ experiments such as in-person oral exams, reflective video logs and AI-critique assignments were reframed as promising templates rather than idiosyncratic workarounds. As Michael (male, professor) put it: It’s gone from maverick to mandatory in under a year. (Michael, male, professor)
At the same time, not all proposals travelled upwards. Ideas that required substantial technical infrastructure (e.g. integrated AI-feedback platforms) or cross-school coordination (e.g. a shared AI literacy curriculum) often stalled at the senior leadership stage and were described as ‘good ideas but not feasible this year’. The same applied to repeated proposals for institutionally funded AI licences for staff and students. Several academics argued that ‘if we’re expected to teach with this, we need proper access, not the free tier’ and suggested school-wide staff licences, while others pushed for basic licences for all students to reduce inequity in access. These suggestions were acknowledged by senior leadership but ultimately deferred, with the senior leadership team indicating that licence options would be ‘explored in the future’ rather than adopted in the immediate wave of reforms. In contrast, practices that could be implemented within existing systems and assessment regulations – such as requiring students to document AI use on cover sheets or shifting part of the assessment to in-person formats – were more readily institutionalised.
Experiences of lessons learnt and institutional barriers culminated in emergent AI-aware routines and policies. The senior leadership team, drawing on staff who had piloted new designs, created a three-level scale to guide acceptable student AI use in relation to skills-building and learning outcomes: Level 1 – no AI use; Level 2 – AI-assisted, allowing students to use AI for ideation and feedback without submitting AI outputs as their own work; and Level 3 – AI-integrated, where AI use is included in the learning outcomes. Later, this was expanded into a four-tier approach for clarity: AI-prohibited, AI-minimal (spelling and grammar only), AI-assisted and AI-integrated, with references required for AI-assisted and AI-integrated work. Students now confirm compliance with the policy via a checkbox on submission of assessments, replacing the previous AI declaration. Patricia (female, professor) reflected on the process: We create spaces for staff and students to explore . . . but at the moment, we have a very narrow focus on assessments and plagiarism. We need to move beyond that.
The policy changes described above demonstrate the impact of recursive feedback: individual innovations generated experiential data and local evidence, which fed into policy discussions and nudged formal guidelines and policies. This process closes the current Coleman cycle; we move to consider the implications of our findings next.
Discussion
Institutionalising AI in management learning: An MF pathway
In this study, we asked how a sector-wide AI shock travels through educators’ cognitive–emotional framings, resource landscapes and assessment practices to generate AI-aware institutional policy. Adopting an MF approach as the explanatory lens, we examined the multilevel pathways through which the introduction of generative AI in management learning and education shapes individual practices and, in turn, becomes embedded in organisational routines and policies. By combining Coleman’s (1990) macro–micro–macro architecture with recent advances in the MF of institutions (Felin et al., 2015; Foss and Linder, 2019), we trace a four-stage macro–micro–macro sequence.
In Figure 3, we interpret each stage and position our results within existing scholarship, then outline our contributions and implications.
Extending Coleman’s bathtub: a four-stage macro–micro–macro pathway from AI shock to AI-aware policy.
Stage 1 highlights how generative AI entered UK higher education as a high-velocity disciplinary shock that disrupted taken-for-granted assumptions about what student assessments can validly measure. Consistent with recent scholarly contributions on the disruptive impact of AI (Barros et al., 2023; Krammer, 2025), we found that the absence of clear guidance produced what participants described as a ‘policy vacuum’. Professional bodies demanded reassurance about academic integrity, employer boards expected AI-fluent graduates, and senior leaders worried about student satisfaction scores if AI was visibly banned. Our data show that such macro turbulence does not flow mechanically into practice. Instead, it creates conditions for divergent framings and intense emotional responses.
Stage 2 shows that academics did not simply comply with or resist this shock. Rather, they engaged in cognitive framing (L1), variously construing AI as a cheating device, amplifier of higher-order thinking or ambivalent hybrid. These framings were not dispassionate. Fear, anxiety, curiosity and excitement were woven through participants’ accounts, functioning as affective MF (L2). For some, AI was an existential threat to academic integrity that had to be tightly policed; for others, it was a provocative ‘thinking partner’ that might help recentre the human in management education. These cognitive–emotional framings (L1 and L2 in Figure 3) shaped whether AI was seen primarily as a threat to be controlled, a partner to be explored or both.
Stage 3 demonstrates that such sensemaking did not translate into practice in a frictionless way. Instead, resource-to-practice translation (L3) was patterned by uneven access to time, tools, institutional guidance and student know-how, as well as by horizontal interaction and mentoring relationships. Availability of time, AI subscriptions and institutional guidance varied significantly across departments – a pattern mirrored in studies of resource heterogeneity in technology adoption (Leonardi et al., 2012). Our contribution is to show how these disparities became divergent action paths. Where plagiarism-detection updates were delayed, staff converted the constraint into oral or viva voce assessments; where small teaching-innovation grants were available, educators channelled funds into AI-assisted feedback pilots. Proposals for institutionally funded AI licences for staff and students, by contrast, did not move beyond discussion: leadership acknowledged their importance but deferred them to future budget cycles. This selective uptake of proposed policy changes underscores that resource-to-practice translation is as interpretive as it is material – academics did not simply use what they had; they reframed both constraints and opportunities as design cues.
Students also surfaced as a nontraditional resource. Educators described them both as ‘digitally illiterate’ and as ‘unexpected codevelopers’ of AI-enabled assessment designs. This duality resonates with principles of critical pedagogy and ecopedagogy (Jandrić and Hayes, 2022) and suggests that MF scholarship should widen its resource lens beyond conventional artefacts and capabilities to include learner competences and barriers to learning.
With formal policy in flux, horizontal interaction became the decisive filter between resources and practice. Informal ‘sandbox’ lunches, corridor conversations and self-organised online channels acted as low-stakes spaces for testing and diffusing ideas. Such peer-to-peer ties echo Reagans and McEvily’s (2003) finding that dense, diverse networks facilitate knowledge transfer; however, our data add nuance: horizontal exchanges did not inevitably accelerate adoption. In one department, a collective decision to pause AI pilots effectively vetoed a top-down initiative, illustrating that lateral ties can impede, as well as enable, diffusion. Mentoring relationships further shaped who experimented and how. Pairing risk-averse lecturers with early adopters lowered the perceived cost of experimentation, echoing research on psychological safety (Edmondson, 1999).
Stage 4 traces how micro-level experiments feed back into macro norms and policy through norm reinforcement/disruption (L4). Small-scale trials destabilised previously taken-for-granted assessment formats. The realisation that ChatGPT could sail through existing exams made written coursework suddenly suspect, while in-person orals – an older form of assessment – gained renewed legitimacy. Some experiments reinforced long-standing values, such as authenticity and critical thinking, but in new socio-technical forms (e.g. AI-critique assignments, reflective videos and dialogical uses of AI in class). Others disrupted assumptions about individual authorship and the centrality of written work.
Importantly, these emergent routines did not remain informal; more resource-intensive proposals – such as school-wide AI licences for staff and students or a formal cross-programme AI literacy pathway – were recognised but postponed. In Coleman’s terms, these developments illustrate how innumerable micro decisions cumulate into selective macro outcomes: some practices become codified, others remain local workarounds or future aspirations. New policies then loop back to reshape lecturers’ resource evaluations and assessment options for the next academic year, opening a new cycle of cognitive framing, experimentation and norm negotiation.
Three empirical insights on AI adoption in management learning
In this section, we reframe the discussion around three empirically grounded insights that integrate management learning and MF perspectives. Rather than treating these contributions in parallel, we develop a single, cumulative argument in which each insight first speaks to AI adoption in management learning and education, and then extends to MF and institutional theory. The three insights focus on (1) affect as a primary driver of AI adoption patterns, (2) the interpretation of resource landscapes, and (3) students as micro-level agents of change. Together, they illuminate how AI-related disruption is lived, interpreted and enacted in practice and how these micro-level dynamics aggregate into institutional change. We situate these insights within a macro–micro–macro process, showing how a sector-wide AI shock reconfigures educators’ local situations, cognitive–emotional framings and perceived constraints and how their situated responses feed back into the emergence of AI-aware institutional policy. We now discuss each insight in turn.
Our first insight shows how affect operates as a primary driver of AI adoption patterns in management learning, shaping how educators interpret and respond to generative AI. In the AI and management learning literature, educator emotions – anxiety, curiosity and ambivalence – are widely documented (Gupta et al., 2024; Sharma, 2025) but are typically treated as attitudes or individual differences that moderate adoption behaviour rather than as mechanisms in their own right. Our data extend this picture: educators’ emotional responses to generative AI – fear, anxiety, resentment, curiosity and excitement – were not epiphenomenal but were primary determinants of which responses were actually enacted. They intensified or dampened experimentation, shaped whether AI was framed as an existential threat or a potential partner and influenced who stepped into ‘translator’ or ‘protector’ roles. This extends Gupta et al.’s (2024) finding on socially negotiated meanings of AI by explaining why those negotiations produce such uneven outcomes: affective intensity, rather than cognitive appraisal alone, determines which framings prevail. For MF theory, this finding challenges models that privilege cognition and interests when explaining how macro conditions shape micro actions (Felin et al., 2015; Harmon et al., 2019). Our contribution is to elevate affect from contextual backdrop or moderator to primary linkage mechanism – a theoretical move supported by institutional theory work on emotion (Tracey and Phillips, 2017; Voronov and Weber, 2020) but not yet incorporated into MF models of institutional change.
Our second insight demonstrates how resource constraints are not simply encountered but constructed, with educators’ responses shaped by socially and emotionally mediated understandings of available possibilities. The AI and management learning literature on assessment redesign (Beninger et al., 2025; Kofinas et al., 2025) has shown that educators who redesign assessments in response to AI tend to produce more educationally valuable outcomes than those who simply lock assessments down. Our data add a causal mechanism: the reason some educators reach creative redesign while others retreat to defensive lockdown is not simply a matter of attitude or competence but of how resource constraints are interpreted in the context of affective framings and peer interaction. Where plagiarism-detection updates were delayed, staff who had emotionally framed AI as a ‘thinking partner’ converted the constraint into an opportunity for oral assessment; those who had framed AI as an existential threat to integrity experienced the same constraint as confirmation of their anxiety. This extends Sharma’s (2025) work by specifying the affective and resource conditions under which unlearning does or does not lead to productive adaptation. For MF theory, this finding challenges treatments of resources as fixed inputs (Leonardi et al., 2012) and underscores the relational and interpretive nature of resource constraints and opportunities.
Our final insight highlights how students emerge as micro-level agents of institutional change, and their competencies, practices and interactions actively shape the direction of AI-related experimentation and policy formation. The AI and management learning literature has called for more nuanced accounts of students in the AI transition: Izak et al. (2025) argue that learners must be treated as active agents rather than passive recipients of AI policy, and Lee et al. (2024) highlight the need for approaches that develop learner agency alongside AI literacy. Our empirical data show what that agency looks like in practice: students were described by educators simultaneously as ‘digitally illiterate’ and as ‘unexpected co-developers’ of AI-enabled assessment designs. This duality is not a contradiction but an empirical finding – the same students whose uneven AI literacy created constraints also provided the impetus and practical knowledge for collaborative experimentation that eventually fed into institutional policy. This challenges simple portrayals of students as passive learners, vulnerable recipients of AI’s harms or savvy digital natives and grounds Izak et al.’s theoretical call in lived practice. For MF theory, this finding invites institutional MF research to widen its resource lens beyond staff competences and organisational artefacts to include learner competences, vulnerabilities and peer interactions as part of the micro-level material through which institutional change is made and unmade.
Our findings also speak directly to emerging work on generative AI in management education. Early contributions urge educators to move beyond simple ‘ban or embrace’ narratives and to engage critically with AI as a paradoxical resource that both threatens and enables learning (Lim et al., 2023; Ratten and Jones, 2023). Recent editorials in AMLE and Management Learning position generative AI as a test case for what it means to think critically and to learn in business schools (Izak et al., 2025; Larson et al., 2024). Empirical studies are only beginning to show how these debates play out in practice, documenting student and staff adoption, AI literacy and assessment redesign (Beninger et al., 2025; Gupta et al., 2024; Kofinas et al., 2025; Sharma, 2025).
By tracing the MF of AI-related policy change, our study complements this work in three ways. First, we show how the academics’ cognitive–emotional framings of AI shape which of the many proposed responses are actually pursued – from defensive lockdown strategies to more generative reimaginations of assessment. Second, we highlight that AI-related initiatives travel through organisational pipelines: resource constraints, peer networks and governance structures that selectively amplify some ideas (e.g. AI-critique tasks and AI-aware assessments) while deferring others (e.g. universal AI licences and cross-programme AI literacy curricula). Third, we make visible the recursive link between classroom experimentation and institutional rules, showing how AI-aware assessment becomes both a site and a driver of broader debates about learning, evidence and academic work in business schools.
Conclusion
Our findings have several implications for Management Learning. First, generative AI functions as a pedagogical disjuncture for academics themselves. The AI shock exposed the precarity of long-standing assumptions about individual authorship, written work and authentic assessment. For some educators, this prompted defensive moves to lockdown assessments; for others, it opened double-loop questioning of what counts as learning and how critical thinking might be evidenced in an AI-rich environment. The AI, in this sense, is not only a topic to be taught but also a catalyst for the educators’ own learning and reflexive practice.
Second, the study extends debates about expertise and coproduction in management education. It not only challenges simplistic student typologies but also suggests that generative AI brings staff and students into new forms of joint experimentation that blur teacher–learner boundaries and foreground learning as a socio-technical and relational achievement rather than the accumulation of stable, individual competences.
Third, if learning is transitioning into a socio-technical process, then scholars and practitioners should treat assessment as a socio-technical practice rather than a neutral measurement tool. Redesigning assessment in response to AI is not just a matter of changing question formats. It involves reconfiguring relationships between human and machine agency, renegotiating notions of authorship and originality, and rearticulating the values that assessment is meant to enact. By tracing how micro-level experiments with AI-informed assessment fed into policy change, we show that AI-aware assessment is a key arena in which institutional norms about learning and knowing are actively worked out and contested.
Our study also offers methodological reflections on using AI tools in qualitative research on AI. We employed an LLM (ChatGPT) in a carefully constrained way to interrogate our emerging coding structure, treating it as a reflexive aid rather than an autonomous analyst. This revealed opportunities – such as surfacing overlaps between codes and prompting clearer justifications – alongside risks, including nonreplicability and the potential reinforcement of dominant interpretive frames. Further methodological work is needed to examine how AI tools can be incorporated into interpretive management research without undermining core commitments to reflexivity, transparency and contextualised understanding.
Collectively, these insights invite a reframing of AI strategies in universities. Durable change is produced not only through formal policies, but at the nexus of emotional response, peer interaction and resource use. To harness AI’s pedagogical promise, ‘educating with the machine’ must remain human-centric, and this requires working with – not around – learners and academics.
Limitations and future research
This study is not without limitations. First, it was conducted in a single, albeit multicampus, UK business school. The findings may therefore not extend to other educational levels, national systems or disciplinary contexts. Replicating the study across multiple institutions and subjects would allow researchers to test external validity and identify context-specific constraints and enablers in the institutionalisation of generative AI.
Second, our design is cross-sectional and captures one iteration of AI adoption and policy formation. While we were able to reconstruct trajectories and feedback loops retrospectively, longitudinal designs – including digital-trace studies of assessment changes and policy revisions over time – could reveal how cognitive–emotional framings, resource landscapes and norms evolve across successive adoption cycles.
Third, our analysis of peer influence focuses primarily on reported dyadic and small-group interactions. We did not systematically map the full network structure of relationships within the school. Future research could combine qualitative data with social network analysis or agent-based simulations to examine how different network topologies (e.g. small-world vs centralised structures) shape the diffusion, stalling or reversal of AI-enabled assessment practices.
Finally, our use of an LLM in analysis introduces specific limitations. The outputs of a proprietary LLM such as ChatGPT are nonreplicable over time: subsequent versions of the model, or different prompts, may produce different suggestions even when supplied with the same data structure. Moreover, the model embodies normative assumptions and training data that are not transparent to the researcher, raising the possibility that it could subtly reinforce dominant framings of what counts as a ‘good’ code or theme. To mitigate these risks, we constrained the LLM’s role to commenting on an already developed coding table, did not provide raw transcripts, and implemented only those changes that were clearly supported by the underlying data and our analytic memos. Our approach aligns with emerging methodological work that treats LLMs as potential aids to theorisation and coding, while emphasising the need for strong human oversight and epistemic caution (e.g. Garcia Quevedo et al., 2026). We encourage future research to explore systematically both the opportunities and pitfalls of integrating LLMs into qualitative analysis.
Footnotes
Funding
The authors disclosed receipt of the following financial support for the research, authorship and/or publication of this article: This study was supported by a University of Exeter Education Incubator Grant of £5000 under the Generative AI and Data Powered Learning category.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.