Algorithmic Management Across Platform and Traditional Work: Evidence from Hungary—Testing the Five Principles of Stark and Vanden Broeck

Abstract

Algorithmic management (AM) has become a key lens for analyzing the digital transformation of work, yet empirical research remains centered on platform labor in North America, Western Europe, and China. This paper applies the five-principles framework covering organizational form, object of management, ideology, modality, and accountability in an exploratory comparison of contrasting organizational settings within one country, Hungary. We define AM as the exercise of managerial functions such as direction, allocation, evaluation, discipline, and remuneration through software systems that continuously capture worker data, process them automatically, and feed the resulting decisions back into the labor process. The analysis draws on a base of 43 semi-structured interviews conducted between 2019 and 2024, all of which, drawn from the five interview-based cases (Wolt, Bolt, Upwork, Data Analytics and ConLog), constitute the coded analytical corpus, together with 21 h of participant observation and documentary evidence, contrasting four platform cases (Wolt, Bolt, Uber, and the freelance marketplace Upwork) with two traditional organizations, a Data Analytics company and a multinational logistics subsidiary. Evidence on Uber is confined to documentary sources (regulatory, parliamentary, legal and media records); the case is therefore used as an institutional and regulatory comparator only, and no worker-level generalizations are drawn from it. Findings suggest that AM is neither homogeneous nor confined to platforms. Full co-optation, including twisted accountability and dissolved organizational boundaries, appears mainly in platform work, whereas AM in traditional firms operates within hierarchies and formal employment, producing constrained co-optation, bounded ideology, and partially re-anchored accountability. We propose the partial Möbius effect as a testable hypothesis rather than as an established theoretical result: it was derived inductively from two traditional cases only (Data Analytics and ConLog) and has not yet been tested against a third, independent case. Its scope, boundary conditions, and durability are left as questions for future comparative and longitudinal research.

1. Introduction

Algorithmic management, the use of data-driven, software-mediated systems to direct, monitor, evaluate, and discipline workers, has become a key concept in discussions about the future of work. Because the term is applied in the literature to phenomena as different as a delivery app’s dispatch engine and a corporate HR analytics dashboard, Section 2.1 states the precise working definition adopted here, together with the inclusion and exclusion criteria through which it is applied in the empirical analysis. However, empirical research has developed under a twofold bias. Geographically, it has been heavily concentrated on North American, Western European, and Chinese cases (Makó et al., 2024). Sectorally, it has focused almost exclusively on platform labor, specifically ride-hailing and food delivery (Rosenblat & Stark, 2016; Möhlmann & Zalmanson, 2017; Wood et al., 2019). These biases together fostered the notion of treating AM as an inherent characteristic of the gig-economy platforms, which obscured its presence in non-platform organizational contexts and its diversity across different national institutional contexts.

A definitive theoretical intervention was Stark and Vanden Broeck’s Principles of Algorithmic Management in Organization Theory. Stark and Vanden Broeck challenge the dominant framing in earlier scholarship of ‘digital Taylorism’ (Huws, 2014; Braverman, 1974) and argue that AM is a qualitatively different organizational logic, the emblematic setting of which is the platform, but the scope of application of which is not limited to it. They suggest five structural principles that distinguish AM from both scientific management and post-bureaucratic/collaborative management: (1) organizational form, (2) object of management, (3) ideology, (4) modality, and (5) accountability. The framework makes an explicit call for comparative empirical research to investigate how the principles work in different organizational settings, a call to which the present paper responds.

We apply the five principles in both platform and traditional organizational settings in Hungary, a Central and Eastern European (CEE) country that has analytically distinctive features: fragmented platform regulation, the “Uber fiasco” of 2016, widely read as the result of a void in platform-specific regulation that was filled by the restrictive application of existing taxi rules, one of the lowest union density rates in the EU (below 10%; ETUI, 2024), and rapid digitalization of platform as well as non-platform workplaces. Hungary is a productive, not a peripheral test case: if AM principles work here as they do in Western Europe, this increases their generalizability; if they do not, institutional context should be theorized as a constitutive dimension of AM, rather than a background condition.

Our empirical base is built on six cases: four platform firms (Wolt, Bolt, Uber and Upwork) and two traditional firms (Data Analytics and ConLog), based on 43 semi-structured interviews, participant observation, and documentary analysis collected between 2019 and 2024 as part of the CrowdWork21 and InCoding European research projects. The Uber case is an exception within this evidence base: neither interview nor observational data were available for it, so it rests on documentary sources alone (Section 3.3), and the claims we draw from it are correspondingly restricted. This paper offers an exploratory empirical application of Stark and Vanden Broeck’s (2024) five principles across both platform and traditional organizations. Building on this comparative analysis, it proposes the partial Möbius effect as a candidate boundary condition for the generalization of algorithmic management theory. The empirical analysis of the six Hungarian cases further suggests that algorithmic management is not confined to platform work but assumes distinct forms depending on organizational context and institutional embeddedness, although six cases from a single country cannot establish how far this pattern generalizes.

Analytically, we make three interrelated contributions. First, we offer an empirical operationalization of the five principles of Stark and Vanden Broeck (2024) across contrasting organizational forms in a CEE institutional context; we are not aware of a prior comparative application of the framework across platform and non-platform settings, but we make no priority claim and present the operationalization as one possible reading of the framework rather than as a definitive one. Second, our cases indicate that AM is not monolithic and not platform-exclusive, but that its realization varies with organizational form: one principle (the object of management) appears to travel essentially intact, two (ideology and modality) travel in form but are institutionally attenuated, and two (organizational form and accountability) require re-specification off the platform (see Section 6.2). The Upwork case further shows that accountability is differentiated within the platform category itself: on a matching-style marketplace, responsibility is displaced onto the client relationship rather than onto the platform’s own “technology company” construct. Third, we propose the ‘partial Möbius effect’, which denotes the limited intrusion of platform-style algorithmic governance into traditional firms as a bounded hybrid configuration. Because the concept was induced from two traditional cases and has not been tested on independent material, we advance it as a testable hypothesis for subsequent research rather than as a settled theoretical contribution of this paper. We position the concept explicitly against adjacent constructs in the theoretical literature (Section 6.3), and we leave the question of its durability to future research rather than claiming a stable hybrid equilibrium (Edwards, 1979; Burawoy, 1979; Thompson, 1990; Kellogg et al., 2020).

The paper is organized as follows. In Section 2, we review the literature on AM, situate Stark and Vanden Broeck’s (2024) framework in it, and synthesize the labor process and fissured-workplace traditions necessary to theorize the partial Möbius effect. In Section 3, we describe the research design and methodology. Section 4 and Section 5 present empirical results from both platform and traditional settings. Section 6 provides a comparative analysis. We introduce the two-axis scope-condition space and develop the concept of partial Möbius effect. Section 7 considers theoretical and regulatory implications. Section 8 concludes and sets out limitations and directions for future research.

2. Literature Review

2.1. The Contested Terrain of Algorithmic Management

What is algorithmic management? Since the term algorithmic management is applied in the literature to anything from a single scheduling tool to a whole mode of production, we start by stating the working definition that guides the analysis in this paper. Algorithmic management is the management of work through software systems that (i) continuously collect digital data about workers, work, and outputs, (ii) process this data automatically to generate decisions, scores, rankings, or recommendations, and (iii) feed these outputs back into the labor process with material consequences for the worker. The three elements are cumulative: a system that records data but does not act on them, or that acts on them without material consequence for the worker, is not covered by the definition.

Three boundary conditions follow, serving as inclusion and exclusion criteria throughout the empirical sections. The criterion is functional, not technological. A system qualifies as AM when it performs a managerial function, regardless of the technical sophistication or whether it involves machine learning. Thus, a rule-based warehouse management system that sets, sequences, and monitors picking targets would qualify, but an enterprise resource planning module or payroll database that merely stores and re-ports data would not. Second, the definition is agnostic about the degree of automation: it includes decisions that are executed without human intervention and decisions where the algorithmic output is the authoritative input to a human manager. This is what allows us to compare platform and traditional settings on a common scale, and it is the reason why the automation continuum of Wood (2021) is used afterward as an analytical axis instead of a definitional boundary. Third, the definition is agnostic on employment status: AM is defined by what the system does to the labor process, not by whether the worker is an employee or a contractor. This last criterion is analytically important for the argument of the paper, because a definition based on self-employment would make the question of whether AM travels beyond the platform unanswerable by construction.

In the following analysis, the term will be used with two further specifications. First, AM is not treated as binary but as a variable: cases vary in the number of managerial functions that are algorithmically mediated, the continuity of data capture, and the extent of discretion left to human managers; it is these variations that the five principles are used to characterize. Second, AM is differentiated from algorithmic surveillance and from digitalization more broadly: monitoring technologies only become part of the analysis where their outputs feed managerial decisions, and general workplace digitalization becomes part of it only as context. Operationally, this definition is translated into empirical indicators by the coding questions set out in Section 3.3, and each of the five principles was coded only against systems satisfying criteria (i) to (iii) above. Where a case had systems that did not meet one of the criteria, for example, the attendance recording infrastructure at ConLog prior to its integration with performance targets, those systems were not coded and not treated as evidence of AM.

The term ‘algorithmic management’ (AM) was introduced by Lee et al. (2015), who defined it as the use of automated, data-driven systems to ‘enable, constrain, and evaluate’ workers at scale. Kellogg et al.’s (2020) systematic review consolidated the field around six mechanisms: restricting, recommending, recording, rating, replacing and rewarding. Invoking Edwards’s (1979) typology of simple, technical and bureaucratic control, they position algorithmic control as ‘the new contested terrain’, a recombinant fourth form that amplifies technical and bureaucratic control while adding continuous data capture, real-time feedback and opacity. Platform studies substantiated this diagnosis, showing how these mechanisms generate information asymmetries, reduce autonomy and create new forms of managerial control despite workers’ formal status as independent contractors (Rosenblat & Stark, 2016; Möhlmann & Zalmanson, 2017; Wood et al., 2019; Duggan et al., 2020).

Much of this early literature framed AM as digital Taylorism: in the lineage of Braverman (1974), digital systems decompose cognitive labor into observable, measurable units (Huws, 2014), and just-in-time labor models reproduce the Taylorist drive for predictability under a vocabulary of ‘flexibility’ (De Stefano, 2016), within wider transformations of capitalist accumulation (Standing, 2011; Srnicek, 2016). This framing has since been qualified. Vallas and Schor (2020) emphasize platform heterogeneity and worker agency, and Krzywdzinski et al. (2025) reframe AM through the politics of its contested implementation. Wood’s (2021) five-level automation continuum, from assistance to full automation of management functions, shows AM in practice as a differentiated spectrum rather than a monolithic control system. Most fundamentally, Stark and Vanden Broeck (2024) contend that the Taylorism framing mistakes AM’s ideology for its structural logic: where Taylorism theorized humans as programmable machines, AM theorizes machines as capable of learning.

Recent comparative research extends the field to regular workplaces. Fernández-Macías et al. (2023) and Gonzalez Vazquez et al. (2025) trace the ‘platformization of work’ in the EU, distinguishing industrial and office variants with different AM intensities and working-condition implications. The study by Schmid and Wiesche (2026) of a traditional automotive supplier finds that AM ‘complements existing organizational structures’ rather than dissolving them, and Dupuis (2024) shows that union power and shopfloor conflict shape AM’s deployment in manufacturing. These studies establish AM’s prevalence in traditional workplaces, its distinctiveness from platform AM and its institutional mediation, but they do not theorize the structural mechanism underlying this distinction.

2.2. The Stark and Vanden Broeck Framework

Stark and Vanden Broeck (2024) begin with a topological observation. The very form of organization within which value creation occurs has changed fundamentally. The factory and the project, emblematic forms of scientific management and post-bureaucratic management respectively, presupposed bounded organizations with clear inside/outside distinctions (Stark & Vanden Broeck, 2024). They draw on Watkins and Stark’s notion of the Möbius organizational form and Stark and Pais’s analysis of the platform economy to argue that the platform is “tendentially unbounded”; value-creating actors, assets, and activities are distributed across the firm’s boundary, making classical managerial supervision impossible and requiring a different organizational logic. They compare this logic to its predecessors on five dimensions (Watkins & Stark, 2018; Stark & Pais, 2020).

AM’s organizational form is first the platform, or the “Möbius organization”: a structure in which inside and outside are constantly folded into each other, so that consumers, workers, producers, and devices all participate in value creation without stable boundary distinctions. Second, the object of management in AM is not the supervision of labor but the co-optation of heterogeneous actors, assets, and activities wherever they sit in organizational space. Third, AM’s ideology is the exact opposite of Taylorism: where scientific management theorized humans as programmable machines, AM theorizes machines as learnable, legitimizing the claim that algorithmic decisions are objective and apolitical. Fourth, the modality or governance principle of AM is co-optation: where hierarchies command, markets contract, and networks collaborate, platforms co-opt through ratings, rankings, recommendations, nudges, and interface design. Fifth, accountability in AM is “twisted”: neither vertical/hierarchical nor lateral/heterarchical, but a triangular, asymmetric structure among platform operators, providers, and users that systematically displaces responsibility from the platform onto workers and customers (Stark & Pais, 2020; Stark & Vanden Broeck, 2024).

The framework is theoretically powerful, but two lacunae limit its direct application to empirical data. First, it was developed on the basis of analyses of paradigmatic platform cases and has not been systematically tested in traditional organizational settings where AM is increasingly present. Second, it lacks an institutional dimension specifying the role of national regulatory fields, industrial-relations systems and employment law in shaping the realization of each principle. The present paper addresses both gaps, within the limits of an exploratory, six-case design situated in a single national setting.

2.3. Labor Process Theory and the Control Lineage

To theorize AM in traditional workplaces, we return to the labor process theory (LPT) tradition that Kellogg et al.’s (2020) ‘contested terrain’ framing continues and that Stark and Vanden Broeck implicitly presuppose but do not explicitly engage. Beyond the control typology introduced above, three LPT resources do analytical work here.

First, Burawoy (1979) explains how workers come to comply with, rather than merely submit to, managerial control. In his factory, piecerate ‘games’ both generated effort and camouflaged exploitation: ‘the very act of playing the game simultaneously produces consent to its rules’. Platform AM’s gamified dashboards, leaderboards and surge incentives replicate this consent dynamic in algorithmically mediated form, now divorced from the employment relationship of factory regimes.

Second, Thompson’s (1990) ‘core’ disciplines any account of AM: it must remain grounded in the structured antagonism of capital and labor, the indeterminacy of labor power, and the control imperative of accumulation, which prevents theorizing AM as a post-human phenomenon transcending the employment relation (Thompson & Smith, 2010). Zuboff’s (1988) distinction between automating, using IT to surveil and standardize, and informating, using IT to generate knowledge workers can use, adds a further degree of freedom: the same infrastructure can support controlling or augmenting deployments, depending on strategy and institutional constraints.

Third, Weil (2017) supplies the structural mechanism linking platform and traditional-firm AM. Fissured lead firms shed direct employment through subcontracting, franchising, and platforms while retaining control through specifications and monitoring, severing control from accountability. We propose the ‘partial Möbius effect’, developed as a hypothesis in Section 6.3, as an inverse, internal version of fissuring: the firm adopts platform-style control logic without externalizing the employment relation, producing a partial rather than total dissolution of organizational boundaries.

2.4. The CEE Institutional Gap

Research on AM has been incomplete across the world. The studies are overwhelmingly based on North American, Western European, and Chinese cases, while Central and Eastern Europe is dramatically underrepresented (Makó et al., 2024). This is analytically relevant because CEE economies display specific institutional arrangements. Bohle and Greskovits describe CEE capitalism as peripherally embedded, with strong dependence on foreign direct investment, hybrid welfare regimes, and variable labor protection (Bohle & Greskovits, 2012). Szabó in particular documents a trajectory of declining collective-bargaining coverage and institutional fragmentation in Hungary, resulting in what Meardi characterizes as ‘social failures’ of EU enlargement (Szabó, 2013; Meardi, 2012). Data from ETUI confirms that Hungary has one of the lowest rates of union density in the EU (below 10%), which makes the collective voice in AM governance especially weak.

At the same time, Hungary is a revealing institutional laboratory. The 2016 Uber exit by coordinated institutional resistance from taxi lobbies, the state, and parliamentary legislation, documented by Makó et al. (2023), indicates how national regulatory fields may shape platform AM even when those fields are underdeveloped. As a single episode reconstructed from public records, however, it cannot by itself establish a general pattern. Bolt and Wolt’s subsequent market-entry strategies were explicitly calibrated to this regulatory history (Pap et al., 2021). The two traditional cases, DA and ConLog, represent, respectively, a knowledge-intensive service firm and a multinational logistics subsidiary with over 700 employees. They provide variation along the task-complexity and institutional-embeddedness dimensions that Wood identifies as key moderators (Wood, 2021).

3. Research Design and Methodology

3.1. Comparative Case Study Design

We employ a multiple case study design following Yin (2003), selected for its established suitability for ‘how’ and ‘why’ questions where context and phenomenon are deeply intertwined (Yin, 2003). The design comprises six cases, four platform firms and two traditional firms, held constant in one key dimension (national institutional context: Hungary) while varying in the primary dimension of interest: organizational form. This controlled variation allows us to observe how the five AM principles manifest differently depending on organizational form, independent of cross-national regulatory variation.

The four platform cases are (1) Wolt, the Finnish-origin food delivery platform that entered Hungary in 2016 and achieved rapid market penetration; (2) Bolt (formerly Taxify), the Estonian platform initially operating in ride-hailing that diversified into food delivery (Bolt.Food) during the COVID-19 pandemic; and (3) Uber, which entered Hungary in 2014, faced coordinated institutional and regulatory resistance, and exited in 2016 before negotiating a regulated re-entry via the Főtaxi partnership in 2024, a case included primarily as an institutional and regulatory comparator documenting the consequences of full platform AM in a hostile regulatory environment rather than as a source of contemporaneous interview data (see Section 3.3). The fourth case is that of a platform, (4) Upwork, to broaden the comparison beyond localized dispatch platforms (food delivery, ride-hailing) toward a global online labor market for knowledge and creative work. The Hungarian sub-sample is drawn from a parallel Hungarian/Serbian UpWorkers study (Pap et al., 2021), in which fourteen respondents (Upwork 1–14; Appendix A) were interviewed as part of CrowdWork21. The inclusion of Upwork tests whether the five principles, developed largely from dispatch-style platforms, are applicable to a matching-style platform in which the allocation of work is client-driven, as opposed to algorithmically dispatched, and where workers are structurally positioned as independent entrepreneurs, not dependent contractors. The two traditional cases are (5) Data Analytics (DA), a pseudonymized knowledge-intensive business services firm with approximately 65 employees using algorithmic performance management and project-allocation tools; and (6) ConLog, a pseudonymized multinational logistics subsidiary with over 700 warehouse employees using a warehouse management system (WMS) for algorithmic scheduling, routing, and productivity monitoring. The seven interviews documented for the DA case in Table 1 and Appendix A fall into two categories. Five were conducted within the organization and encompass the tiers at which its algorithmic methods are formulated and employed: the owner-manager, senior operations management, the creator of the firm’s internal tools, and two participants in data processing and client-facing positions. The last two consisted of expert interviews with experts outside of DA, one focusing on the legal governance of algorithmic management and the other on AI ethics at a global cloud service provider utilized by DA. The two expert interviews were carried out to contextualize the firm’s operations within their legal and technological frameworks; assertions regarding DA’s systems, detailed in Section 5, are based on the five internal interviews along with the firm’s dashboard documentation.

Table 1. Overview of the empirical material used in the comparative analysis.

Case	Interviews	Participant Observation	Documentary/Organizational Material	Research Project
Wolt	8	Yes (21 h shared across platform fieldwork)	Platform documents, app interfaces, terms & conditions	CrowdWork21
Bolt	6	Yes (platform fieldwork)	Platform documents and app interfaces	CrowdWork21
Uber	0 (documentary case)	No	Regulatory, legal and policy documents	CrowdWork21
Data Analytics (DA)	7	No	Internal organizational documents and dashboards	InCoding
ConLog	8	No	WMS and HR documentation	InCoding
Upwork	14	No	Platform-related interview material	CrowdWork21
Total	43	21 h	Multiple documentary sources	CrowdWork21 & InCoding

Source: Authors’ own design.

Case selection was theoretical rather than representative: the cases were chosen to maximize variation in organizational form while holding institutional context constant, and to provide variation in task complexity and skill intensity within the traditional category (high-skill/knowledge-intensive at DA, low-skill/routine at ConLog). Both selections were guided by theoretical replication logic (Yin, 2003): we anticipated that AM would manifest differently in platform versus traditional settings, and the cases were chosen to make such differences observable; the design nevertheless remains exploratory and abductive rather than hypothesis-testing.

A note on case heterogeneity is warranted here. The six cases differ not only in organizational form but also in size (from approximately 65 employees at DA to over 700 at ConLog, and the dispersed contractor workforces of the platforms), in the type of technology deployed (food-delivery dispatch, ride-hailing matching, reputation/matching algorithm, project-allocation and performance-management dashboards, and a warehouse management system), and—as set out in Section 3.3—in the evidence base available for each. This heterogeneity is not a confound to be eliminated, but a property of the sampling logic, in a theoretical rather than representative design: the cases were selected to cover the widest possible range of organizational and technological configurations, while keeping the national institutional context constant, so that the five principles could be observed under maximally different conditions. There are two aspects of the design that make this variation tractable analytically. The comparison is made, first of all, not at the level of firms, sizes, or technologies, but at the level of the five principles which provide a common analytical scale onto which each case is mapped. Secondly, the definition of AM used in Section 2.1 is explicitly functional, not technological, and agnostic to firm size and employment status: a system enters the analysis only when it carries out a managerial function that meets criteria I–III, regardless of whether it operates via a delivery app, ride-hailing engine, corporate dashboard, or WMS. Thus, technological and size heterogeneity is integrated into the definition instead of being allowed to undermine comparability.

3.2. Data Collection: Interview Distribution and Project Integration

Data were gathered as part of two successive European research projects. CrowdWork21 (2019–2021), co-funded by the European Commission, examined platform work in Germany, Hungary, Portugal, and Spain through multi-stakeholder qualitative interviews and participant observation. InCoding (2022–2024), funded under the EU EaSI program, examined algorithmic management and collective voice in Denmark, Germany, Hungary, and Spain, explicitly extending coverage to traditional workplaces. Table 1 summarizes the empirical material for each case; interviewee characteristics are reported in Appendix A.

The integration of data from two projects with different scopes and timeframes was managed through protocol alignment and abductive re-analysis. Both projects employed semi-structured interview protocols addressing daily work processes and task allocation; awareness and experience of algorithmic decisions; ratings, scoring, and performance feedback; job quality dimensions (autonomy, work intensity, participation, transparency); collective voice; and perceptions of fairness and accountability. The InCoding protocol added questions on algorithmic transparency and collective governance mechanisms that were not present in CrowdWork21. For the comparative analysis in this paper, we used only the shared thematic domains, applying the five-principle coding frame to the full corpus of 43 interviews, including the Upwork interviews conducted within the parallel UpWorkers study. Data generated specifically for country-level reporting in each project (e.g., national industrial-relations context mapping) were used as contextual background rather than primary analytical evidence.

Interviews were conducted in Hungarian with couriers, drivers, warehouse workers, managers, trade union representatives, and works council members, and in English with European management respondents. All were audio-recorded with informed consent, transcribed, and coded using Atlas.ti. In addition to interviews, data sources include: 21 h of participant observation at Wolt couriers’ work locations (January–March 2021), analysis of platform Terms and Conditions agreements, courier apps, driver dashboards, and performance-feedback interfaces for all four platform cases; internal company documents provided by DA (performance dashboard documentation) and ConLog (WMS operational manual, HR procedure documentation); and secondary analysis of the European Company Survey (ECS 2019, N = 24,942; Eurofound & Cedefop, 2020), used as contextual background only.

3.3. Analytical Strategy and the Uber Case

Analysis was guided by within-case and cross-case logic informed by an abductive strategy (Timmermans & Tavory, 2012). In each case, each of the five principles was turned into an empirical question posed to the data (see some examples in Table 2). Coding was performed in three cycles, from inductive open coding through second-order themes to the five principles as aggregate dimensions, with constant attention to both confirmatory and disconfirmatory evidence; the procedure and its output are set out in detail below and in Table 3a and Table 4a. Patterns identified within cases were then compared across the six cases to assess convergence and divergence.

Table 2. Examples of the operationalization of the five principles.

Principle	Empirical Indicator/Coding Question Used
Organizational Form	How is work contracted, allocated, and coordinated? Does the organization operate as an employment-based hierarchy, or as a platform that mediates a triangular relationship among worker, platform, and client/customer? What is the worker’s formal status (employee, freelancer, self-employed), and has the introduction of AI/AM been accompanied by structural reorganization (flattening, splitting, or merging of departments, outsourcing of processes to external units)?
Object of Management	What do algorithmic systems act upon in practice? Which domains of the labor process are subject to AI/AM, and what data are collected, from what sources (software logs, wearables, cameras, audio; workers themselves; or third parties such as customers and clients)?
Ideology	How is the deployment of algorithmic systems discursively justified and legitimized? What motives does management invoke (efficiency, labor-cost reduction, competitive imperative, pioneering innovation)? How do workers frame their own position (independence, flexibility, entrepreneurship, “being one’s own boss”), and how do these self-classifications align or clash with how the platform/company categorizes them (e.g., self-employed contractor vs. employee)?
Modality	Through what mechanisms is control exercised? Are decisions fully automated or subject to human supervision (semi-automated)? Does control operate through direct managerial oversight, or indirectly via ratings, rankings, customer reviews, and reputation scores? How continuous and opaque is the monitoring, and do workers report intensification of work or altered behavior in response to algorithmic evaluation?
Accountability	When an algorithmic decision produces a negative outcome for a worker, who bears formal and practical responsibility? Through what channels can the decision be contested, and to what degree do workers and their representatives have access to the underlying data, involvement in parameter-setting, and protection under existing regulation (GDPR, AI Act, national labor law)?

Source: Authors’ own design.

Table 3. (a) Illustrative open (first-cycle) coding: from data extracts to first-order codes. (b) The analytical path (reflexive TA phases → activities → outputs).

(a)
Illustrative Data Extract (Source)		First-Order (Open) Codes		Analytic Note Recorded During Coding
“I like to do it due to the flexible working time, and I can work whenever I want to. However, these companies provide the platform only … they are not employers. There are no benefits…the biggest risks are taken by them (taxes, costs of having the vehicles).” (BoltTaxi-3)		flexibility as compensation; ‘platform is not an employer’; worker bears cost and risk; absence of social benefits		Extract carries consent and boundary dissolution simultaneously; double-coded and retained under two themes.
“Wolt is forcing the people to be self-employed.” (WOLT-1)		imposed self-employment; contractual status as constraint		Contradicts the entrepreneurial self-description found in other extracts; kept as a disconfirming instance for the ideology theme.
“I’m 0.96 on a 0–1 scale; it measures average speed of delivery, delays, etc. Average carriers are 0.5–0.6…” (WOLT-2)		self-knowledge as a score; metric visible to the worker; comparison with peers		Basis of the ‘worker as data profile’ theme; note that the worker reproduces the platform’s own metric language.
“We can rate the riders, and they can rate the drivers. For the drivers, if rating goes under 4.65, then the driver is suspended from the platform for 24 hours automatically…” (BoltTaxi-2)		customer as evaluator; rating threshold triggers sanction; automatic suspension		Links modality and accountability; the disciplinary act has no human author, which is what the accountability coding later picks up.
“I suspect that there is a rating system… but the company says that there isn’t… I tried to contact the ‘management’ of Wolt, but they communicate through messaging only…” (WOLT-1)		opacity of the system; denial of the rating’s existence; no addressable counterpart		First appearance of ‘twisted accountability’ in the platform material; the code was generated inductively before being matched to the principle.
“I’m pulling a computer with me, and I can see my real-time performance, the percentages… If I see that I’m delayed, I try to increase my pace to catch up.” (ConLog 8)		real-time metric feedback; self-directed intensification; metric as motivation		Same first-order code as in the platform material but in a different institutional frame; retained to enable cross-case comparison of the object of management.
“…sometimes I take it easy… I’m hacking the system, as I scan different activities than what I actually do. Nobody ever noticed…” (ConLog 7)		gaming the metric; informal slack; unmonitored discretion		Only scattered evidence of this kind in the corpus; recorded as a limitation on the study’s ability to capture worker agency (Section 8).
One worker described the system as “a peer in my work” (ConLog 6), while confirming that managers use it to track performance, attendance and breaks.		system as colleague rather than overseer; simultaneous acceptance and awareness of monitoring		Disconfirming instance for a control-only reading of AM; fed into the ‘thin objectivity’ theme.
“We are informed definitely about the changes and new implementation. Once the system is live, we can give direct feedback … We are not involved in system design or parameter setting. Neither the workers’ representatives.” (ConLog 6)		ex post consultation only; no role in parameter setting; representation present but not decisive		Residual code at first cycle; became the basis for distinguishing ex ante co-determination from ex post contestation (Section 5.5 and Section 6.3).
“I could not just simply quit, due to the rating system… I had to do these really dumb tasks… I did not [solve it]. It is what is.” (Upwork 3)		rating as lock-in; client judgment unchallengeable; no appeal channel; resignation		Accountability displaced to the client rather than to the platform; did not fit the existing ‘twisted’ code and was retained as a separate theme (Section 6.2).
“I believe in algorithms, I think the whole platform was designed to [be fair].” (Upwork 9)		belief in algorithmic fairness; rating as quality assurance; entrepreneurial self-description		Contrasts with the courier accounts that read the objectivity claim as concealment; basis of the ‘internalized objectivity’ theme (Section 4.3).
(b)
Phase (Braun & Clarke, 2006)	Coding Cycle	Activity in This Study	Reflexive Checks/Decisions	Output (Where Reported)
1. Familiarization	-	Repeated reading of transcripts, field notes, and documents in Atlas.ti; re-listening to recordings	Initial reflective memos; positionality memo for the DA insider researcher	Analytic memos
2. Generating initial codes	Cycle 1 (open coding)	Line-by-line descriptive coding close to respondents’ language, framework held in abeyance	Codes generated from data, not from the five principles; extracts carrying two principles double-coded and flagged	Table 3a (first-order codes)
3. Searching for themes	Cycle 2 (grouping)	Collating first-order codes into candidate second-order themes by shared analytical content	Candidate themes kept provisional; codes appearing in more than one theme retained in both	Candidate theme set
4. Reviewing themes	Cycle 2 (consolidation)	Testing candidate themes against coded extracts and the whole corpus; merging, splitting, collapsing	Deliberate search for confirming and disconfirming cases; single-interview codes marked as illustrative; independent second-coder recode of DA, discrepancies resolved against extracts	Audit decisions (see Table 4b)
5. Defining and naming themes	Cycle 3 (aggregation)	Relating refined themes to the five principles as aggregate dimensions; naming themes; retaining a residual category	Residual themes (ex post consultation; works-council role) kept rather than forced into a principle	Table 4a (data structure)
6. Producing the report	-	Writing the within- and cross-case account, illustrated with extracts	Cross-case comparison for convergence/divergence prior to fixing structure	Section 4, Section 5 and Section 6

Source: Authors’ own design.

Table 4. (a) Data structure: from first-order codes to second-order themes and aggregate dimensions. (b) Worked pathway from open codes to concept (accountability principle).

(a)
First-Order (Open) Codes (Illustrative)		Second-Order Theme	Aggregate Dimension (AM Principle)		Cases in Which the Theme Is Present
imposed self-employment; worker-supplied capital; multimapping; no addressable employer; entrepreneur self-identity; “no intermediary needed”; own client relationships		Worker-borne capital and dissolved employer boundary	Organizational form		Wolt, Bolt, Uber, Upwork
dashboard inside the hierarchy; employment contract retained; departmental structure unchanged		Platform-style tooling within preserved firm boundaries	Organizational form (re-specified)		DA, ConLog
self-knowledge as a score; metric visibility; interchangeable node; throughput target; success score; public rating; responsiveness metric		Worker rendered as a data profile	Object of management		All six cases
customer as evaluator; client rating as performance input; passenger score		Customer or client as co-manager	Object of management		Wolt, Bolt, Uber, DA
‘the algorithm is neutral’; ‘we are a technology company’; denial of employer status		Thick objectivity claim	Ideology		Wolt, Bolt, Uber
‘the system removes bias’; manager can overrule the score; discretion retained by partners		Thin objectivity claim with managerial override	Ideology		DA, ConLog
rating thresholds; surge and bonus incentives; push messages; gamified progress display		Real-time co-optation	Modality		Wolt, Bolt, Uber
annual review cycle; ratings routed through HR; targets covered by collective agreement; no gamification		Institutionally mediated co-optation	Modality		DA, ConLog
automatic suspension; no explanation; no addressable counterpart; ‘tech company’ legal defense		Twisted accountability	Accountability		Wolt, Bolt, Uber
line manager conducts the disciplinary process; HR channel open; works council contests targets ex post		Hybrid, partially re-anchored accountability	Accountability (re-specified)		DA, ConLog
opaque evaluation; unexplained score drops; “it is what is”		Reputation-mediated twisted accountability (client-displaced)	Accountability		Upwork
informed after go-live; no parameter-setting rights; representation without co-determination		Ex post contestation without ex ante co-determination	Residual theme; motivated the re-specification of accountability and the definition of Axis 2		ConLog (HR analog at DA)
(b)
Illustrative Extract (Source)	First-Order (Open) Codes—Ph. 2	Candidate Theme—Ph. 3	Review Decision & Reflexive Check—Ph. 4	Final Second-Order Theme—Ph. 5	Aggregate Dimension
“…if rating goes under 4.65, then the driver is suspended… automatically…” (BoltTaxi-2); “I suspect there is a rating system… but the company says there isn’t… they communicate through messaging only” (WOLT-1)	automatic suspension; no explanation; no addressable counterpart; opacity; ‘tech-company’ legal defense	“no one to hold responsible”	Confirmed against all four platform cases; matched to the framework’s existing category	Twisted accountability	Accountability
Line manager conducts the disciplinary process; HR grievance channel open; works council contests targets after the fact (DA, ConLog)	human author of the decision; formal HR channel; ex post contestation by representatives	“human author + formal channel”	Did not fit the framework’s platform-derived binary; retained as distinct rather than forced → prompted re-specification of the principle	Hybrid, partially re-anchored accountability	Accountability (re-specified)
“We are informed… once the system is live we can give feedback… we are not involved in system design or parameter setting. Neither the workers’ representatives.” (ConLog 6)	informed after go-live; no parameter-setting rights; representation without co-determination	“consulted but not co-deciding”	Mapped onto none of the five principles; retained as residual → motivated the joint-regulation axis (Section 6.3)	Ex post contestation without ex ante co-determination (residual)	Residual → definition of Axis 2

Source: Authors’ own design.

The analytical process was based on reflexive thematic analysis (Braun & Clarke, 2006) in an abductive way, in combination with the five-principle framework and a data structure of the type outlined by Gioia et al. (2013). It was performed in three coding cycles. In the first cycle, open coding, transcripts, observation notes and documentary material were read line by line in Atlas.ti and segments were labeled with short descriptive codes that were close to the respondents’ own language (e.g., ‘rating decides which orders I get’, ‘I supply the bike’, ‘the manager can overrule the score’). Coding at this stage was intentionally inductive, not guided by any framework, as codes were generated from the material, which allowed practices that Stark and Vanden Broeck (2024) did not anticipate emerging, and the coding instrument did not merely confirm the framework. This step is shown in Table 3a, which shows representative data extracts, open codes derived from the data extracts, and the analytic note recorded at the time of coding.

In the second cycle, the open codes were compared, merged, and grouped into second-order themes on the basis of shared analytical content. The codes ‘own vehicle and phone’, ‘partner contract’, ‘no cafeteria or sick pay’ and ‘works for four apps simultaneously’, for instance, were grouped into the theme ‘capital borne by workers and dissolved employer boundary’. Codes appearing in more than one theme were retained in both and flagged, since several extracts carry evidence for two principles at once. The codes on rating thresholds, for example, bear on modality and on accountability. In the third cycle, the second-order themes were related to the aggregate dimensions supplied by the framework, that is, to the five principles, while a residual category was kept for themes that did not map onto any principle. These residual themes, notably the mediating role of works councils and HR procedure, and the distinction between being consulted after a system goes live and having a say in its design, are what subsequently motivated the re-specification of the accountability principle in Section 6.2 and the definition of the second axis in Section 6.3. Table 4a presents the resulting data structure.

Themes were then compared across the whole corpus to find confirming and disconfirming cases prior to fixing structure. Cases in which a code was present in only one interview were marked as such and are presented in the text as illustrative rather than patterned. The DA material was re-coded independently by a second coder with no previous connection to the firm (see below), and where there were discrepancies, these were resolved by referring back to the extracts rather than averaging judgments. The coding structure is reported in Table 3a,b and Table 4a,b, and the full codebook is available from the authors on request.

Validity was ensured through methodological triangulation across data sources (interviews, observation, documents, platform artifacts) and analytical triangulation supported by member-checking through a stakeholder validation seminar (June 2024, Budapest). Positionality note: one researcher previously worked at DA in a role connected to the case, prior to the fieldwork. This insider knowledge facilitated access and contextual understanding. It was managed through explicit reflective memos during coding, cross-checking of all DA interpretations with a second coder who had no prior connection to the firm, and deliberate inclusion of disconfirmatory evidence in the DA case write-up. We do not claim that the concept of ‘partial Möbius’ has been validated in this work on its own. ConLog cannot be used as an external test of the concept, because it was developed abductively from the DA and ConLog material and it would be circular to present it as such. The InCoding partner-country material from Denmark, Germany, and Spain would provide a more credible external comparison, but it was used here for contextual member-checking at the June 2024 Budapest stakeholder seminar rather than as a systematic cross-national test, and we therefore do not report it as confirmatory evidence. What we report are the procedural safeguards noted above (reflective memos, second-coder cross-checking of all DA interpretations, deliberate search for disconfirming evidence). Thus, the partial Möbius effect is treated in this paper as an analytically useful interpretation of the observed cases, rather than as an independently confirmed result, and the conditions for its testing are given in Section 6.3, Section 6.4 and Section 8.

The Uber case is included as an institutional and regulatory comparator. Its evidentiary status is, by design, different from the other platform cases. Uber ride-hailing was active in Hungary between 2014 and 2016, when coordinated resistance from taxi associations, the state and the legislature forced Uber to exit. Uber came back only in 2024 via a regulated Főtaxi partnership, a configuration that came after and is different from the standalone ride-hailing operation that is the object of our analysis. Thus interviews with a contemporary Hungarian Uber workforce were not only unavailable to us but structurally impossible for most of the period under study, as no such workforce operated under the organizational form we are analyzing. Instead, our evidence on Uber is documentary material, regulatory and parliamentary proceedings, legal submissions, and contemporaneous media coverage. It is the only record we have of what makes Uber analytically valuable here: a publicly documented instance of the collision between platform AM and a national regulatory field. We do not claim that this material is equivalent to the interview evidence available for the other cases. It records how the conflict was argued and adjudicated, not how work was organized or experienced, and the inferences we draw from it are limited accordingly. We accordingly calibrate our claims. Uber is used to illustrate the organizational form and ‘twisted accountability’ principles as they emerge in the regulatory record—responsibility displaced onto drivers, with no bounded employer to hold to account—and to record the institutional dynamics of the Hungarian context. We do not infer worker-level claims about lived modality or ideology from the Uber case, as these would require the interview evidence available for Wolt and Bolt. Where the comparative tables report Uber, they do so on this restricted basis. Table 1 makes transparent the evidence type underpinning each case.

4. Findings I: Algorithmic Management in Platform Labor

The platform findings are preceded by a note on evidence. Of the four platform cases, interview data were available for Wolt, Bolt and Upwork, and participant-observation data only for the delivery-platform fieldwork (see Table 1). The Uber case relies on documentary evidence from the regulatory episode of 2014–16 (Section 3.3). Thus, statements below about observed operational practice, in particular the object of management (Section 4.2) and the modality of co-optation (Section 4.4), are made for Wolt, Bolt and Upwork only. Uber contributes to the platform findings only on the principles that its documentary record can support: organizational form (the legal construction and contestation of the platform boundary, Section 4.1), the discursive form of the ideology principle, i.e., the “technology company” claim (Section 4.3), and accountability (Section 4.5). Where the text refers to the platform cases collectively, it should be interpreted as “Wolt, Bolt and Upwork” unless Uber is specifically cited from the documentary record.

4.1. Organizational Form: The Full Möbius

All four platform cases are consistent with Stark and Vanden Broeck’s account of the platform as a Möbius organization, though on different evidence: the operational detail below is drawn from fieldwork at Wolt, Bolt and Upwork, while Uber’s consistency is established from the regulatory and legal record (Section 3.3). At Wolt and Bolt, workers’ formal “partner” status places them outside the firm as self-employed contractors, yet they are governed in real time through centralized algorithmic systems that decide routes, order priority, and waiting times. Workers supply their own capital (bicycles, vehicles, phones) and bear the associated costs, while the platform extracts coordination rents, scaling by coordination rather than asset ownership, as Watkins and Stark (2018) predict for Möbius organizations. One Bolt driver captured both sides of this arrangement:

“I like to do it due to the flexible working time, and I can work whenever I want to. However, these companies provide the platform only, compared to normal employers; in fact, they are not employers. There are no benefits, such as a cafeteria, legal protection, or language courses. Normally, the drivers are self-employed; thus, the biggest risks are taken by them (taxes, costs of having the vehicles).”

(BoltTaxi-3)

One respondent frames Upwork as an entry point to full entrepreneurship:

“It provides a good opportunity to start an entrepreneurial career… I could have my fully owned business, without the need of an intermediary”

(Upwork-2)

Other couriers described the same boundary dissolution in practical terms: a Wolt courier noted that “Wolt is forcing the people to be self-employed” (WOLT-1), while another worked simultaneously for four platforms after the KATA tax reform capped per-client income, experiencing sharply different algorithmic work rhythms on each (BoltEat-3). Orders, working hours, earnings, and the terms of the relationship were all assessed algorithmically, but no employer was legally present to be held accountable. The Uber case adds a temporal and legal dimension: on the documentary record, Hungarian regulators applied the logic of firm boundaries to an organizational form that operated across them, treating drivers as employees subject to taxi licensing law, and the 2016 market exit followed. Because this reading is based on regulatory and media sources and not evidence generated within the firm or among its drivers, we present it as a plausible interpretation of a single documented episode, rather than an established causal account.

4.2. Object of Management: Workers as Data Profiles

In the three platform cases for which we have interview data (Wolt, Bolt and Upwork; observational data only for the delivery-platform fieldwork), the object of algorithmic management is not the laboring body in a designated workspace but a data profile distributed across time and space. Couriers and drivers are governed as interchangeable nodes in a dynamic network: what matters algorithmically is real-time location, acceptance rates, completion rates, and customer ratings. Management operates as “supply and demand matching” rather than labor supervision, confirming Stark and Vanden Broeck’s (2024) argument that AM co-opts heterogeneous actors wherever in organizational space they sit. Workers are acutely aware of being reduced to a score, even when the platform does not disclose how it is computed:

“I’m 0.96 on a 0–1 scale; it measures average speed of delivery, delays, etc. Average carriers are 0.5–0.6. They can deliver in an hour 2–3 deliveries, the better ones can do 5–6.”

(WOLT-2)

Workers are rendered as a public “success score,” explicitly described by one respondent as calculated from “rating (public and private) and the earnings, responsiveness, etc.” (Upwork-4).

The most distinctive feature of platform AM in our data is the extension of the managerial object to include customers as co-managers. Through the rating interface, customers are enrolled as a distributed disciplinary infrastructure: they evaluate workers, generate data that feeds the dispatch algorithm, and, when ratings fall below a threshold, trigger automated sanctions that workers experience as managerial interventions. Drivers reported that they typically cannot learn the reason behind a low rating and therefore cannot improve (BoltTaxi-3). This triangulation of platform operator, worker, and customer is the structural basis of the “twisted” accountability described in Section 4.5. We do not extend this operational claim to Uber, whose Hungarian operation predates our fieldwork and is documented only through the regulatory record (Section 3.3).

4.3. Ideology: The ‘Thick’ Objectivity Claim

Stark and Vanden Broeck’s idea of the ideological dimension of AM was strongly present in the interview-based platform cases, Wolt and Bolt: penalties like reduced order volumes, de-prioritization in dispatch queues, or account suspension were consistently attributed to “the system” and not to managerial decisions. In the case of Upwork, the matching/ranking is dependent on the previously earned results. The same objectivity claim is documented for Uber, but at the level of corporate and legal self-presentation rather than worker experience: in the 2014–2016 proceedings Uber advanced the “technology-company” framing as a legal defense (Section 4.5; Makó et al., 2023). We thus consider the ideology principle to be evidenced for Uber only in its discursive/legal form, and not in the worker-level accounts available for Wolt, Bolt and Upwork. We call this a “thick” objectivity claim because the framing is not merely rhetorical but structurally functional in two ways. First, it underwrites the denial of employment status: without an employer making decisions, there is no employee to protect. Second, it shields decisions from contestation: if outcomes are algorithmic rather than managerial, the formal procedures available for challenging employer decisions do not apply.

Couriers in our sample widely understood this framing as a way of hiding power, but had no formal avenue to contest “system decisions”. A driver’s request for a “reason code” behind sub-five-star ratings went unanswered (BoltTaxi-3), and couriers doubted the veracity of the platform’s communication about how work is allocated and priced (BoltEat-1; BoltTaxi-2). The experience of one Wolt courier illustrates how the objectivity claim is backed by organizational unreachability:

“I suspect that there is a rating system which may be used to distribute the opportunities among carriers, but the company says that there isn’t. …the algorithm gives deadlines that the carrier cannot keep. So, he got push messages to speed up. I tried to contact the ‘management’ of Wolt, but they communicate through messaging only; it took some time to have a face-to-face discussion with one of the leaders.”

(WOLT-1)

Some of the interviewed Upworkers pointed at the rating system: “I believe in algorithms, I think the whole platform was designed to [be fair]” (Upwork-9). “I don’t feel the disadvantage of the evaluation… review and rating is a quality assurance tool” (Upwork-6)

4.4. Modality: Co-Optation Through Ratings, Pricing, and Gamification

The co-optation modality was operationalized through three mechanisms mirroring Kellogg et al.’s (2020) “Rs”. First, order allocation was made directly contingent on customer ratings (recording and rating), with sustained decline below a threshold triggering automated warnings and eventually deactivation:

“We can rate the riders, and they can rate the drivers. For the drivers, if the rating goes under 4.65, then the driver is suspended from the platform for 24 h automatically. If it stays there, further sanctions will follow.”

(BoltTaxi-2)

This outsources the disciplinary role to customers while the platform retains the prerogative to set thresholds. Second, dynamic surge pricing (rewarding) nudged couriers to work during high-demand windows without direct orders. Third, availability bonuses (recommending and restricting), such as a bonus for completing X orders in a week, deployed gamification logic, motivating continuous work through goal-framing rather than commands.

The combined effect is what Burawoy (1979) would identify as a managerially imposed game: workers are drawn into targets and competitions that generate effort in a form readable by the platform’s metrics, while the terms of value capture remain obscured. Unlike Burawoy’s factory games, which were largely worker-initiated, platform games are entirely designed and parameterized by the operator; workers play by rules they did not write and cannot renegotiate. In Wood’s (2021) terms, Wolt and Bolt sit close to “high automation” on the management-functions continuum: direction, evaluation and discipline are heavily algorithmized, with human managers stepping in mainly to respond to system-generated alerts.

Co-optation at Upwork runs through reputation rather than real-time gamification: rating thresholds gate future job access, and Upwork combines client rating with an algorithmic “success score” that determines visibility in the marketplace’s own matching/ranking system.

4.5. Accountability: Twisted Responsibility

In all four cases, accountability takes the triangular, “twisted” form theorized by Stark and Pais (2020) and Stark and Vanden Broeck (2024); the Uber documentary record points in the same direction, but at the level of legal and corporate self-presentation rather than observed practice. When a delivery goes wrong (late arrival, wrong order, an accident in transit), responsibility is systematically displaced from the platform (whose algorithm assigned the order, set the timeline, and calibrated the route) to the courier (who “chose” to accept the order as a self-employed agent) and the customer (whose rating then penalizes the courier). The platform stands formally independent of both parties.

One respondent from Upwork, asked how he resolved a dispute over an unfair rating, answered: “I could not just simply quit, due to the rating system… I had to do these really dumb tasks… I did not [solve it]. It is what is” (Upwork-3)—and separately notes that this kind of issue “could be resolved in a legal way only.” Another describes being unable to have a low rating removed despite disputing it. This is a doubly twisted accountability structure: unlike Wolt/Bolt, where the platform itself is the unaddressable party, Upwork displaces accountability to the client, while the platform’s rating algorithm enforces the client’s judgment with no internal appeal.

Within the documentary record, the Uber case is the clearest institutional example available to us. In Hungarian regulatory proceedings in 2014–2016, Uber consistently portrayed itself as a “technology company” providing software rather than a transportation company employing drivers: a legal construction of twisted accountability, not simply its organizational by-product (Makó et al., 2023). The episode suggests that twisted accountability can be actively produced and defended through legal and political strategy. Uber’s forced exit suggests that, in this instance, sufficiently organized institutional actors were able to undo it. We do not, however, characterize the episode as a natural experiment: it is a single case reconstructed from public documents, and the regulatory conditions for piercing such structures in other cases cannot be inferred from it.

5. Findings II: Algorithmic Management in Traditional Workplaces

5.1. Organizational Form: Platform Logic Inside the Firm

Neither DA nor ConLog is a platform firm; both have clear organizational boundaries, formal employment relationships, and established hierarchies. Yet both have introduced algorithmic systems that partially replicate platform-like governance within the firm’s interior, producing what we theorize in Section 6 as the ‘partial Möbius effect.’ At DA, a project-allocation dashboard assigns consultants to projects based on skill profiles, availability, and client ratings, a logic that resembles platform matching more than traditional managerial assignment. At ConLog, warehouse workers are governed by a WMS that assigns picking tasks, calculates travel routes within the facility, monitors task-completion rates in real time, and generates automated performance alerts.

Therefore, the organizational form remains the firm, but platform-like governance logics penetrate the firm’s interior. The key structural difference from the platform setting is precisely the preservation of the firm boundary: both DA and ConLog employ workers under formal contracts, operate within Hungarian employment law, and maintain organizational structures in which human managers hold formal authority. This boundary preservation has direct implications for every other principle, as we demonstrate below.

5.2. Object of Management: Employees-as-Users

In both traditional cases, the object of AM is a worker who retains formal employment status, a structurally significant difference. At DA, performance dashboards combining billable hours, client ratings, and peer feedback render the employee legible as a set of metrics, echoing the data-profile logic of platform AM; but DA employees have employment rights, can challenge assessments through HR, and have access to an internal employee-representation forum with consultative rather than co-determination competence. At ConLog, the WMS automates task assignment, continuously monitors productivity, and generates alerts when output drops below target, with fulfillment accuracy and throughput acting as customer-satisfaction proxies that partially reproduce the customer-as-quasi-manager dynamic. Yet the data profile is not the worker’s only legal identity: they remain parties to an employment contract under collective labor law.

The DA respondents articulated this duality, legibility as a metric profile plus retained employment status, in their own words:

“There are obviously many aspects to how we judge an employee. Also, it is not certain that it is good if the % in the PPS is 100%; it means that you worked a lot of overtime. Neither for the company, because then the employee will be exhausted, nor for the employee, because then he will have no private life, so this has to be balanced.”

(Data Analytics 4)

Workers’ own accounts reveal a distinctly ambivalent relationship to real-time metrification, internalized as motivation but also open to manipulation:

“I’m pulling a computer with me, and I can see my real-time performance, the percentages. I feel that it’s motivating to see my performance on a monitor during my shift. I can see how many hours I worked, how many breaks I took. If I see that I’m delayed, I try to increase my pace to catch up.”

(ConLog 8)

“…sometimes I take it easy, and do not always exactly follow the rules, e.g., I’m hacking the system, as I scan different activities than what I actually do. Nobody has ever noticed or asked me about this so far. If I need to walk a long way to find a computer to scan something and have many tasks at the same time, I’m not wasting my time walking. I just skip this step; I think it helps me to be faster in my work.”

(ConLog 7)

Others described the system in similarly non-adversarial terms: one worker called it “a peer in my work” (ConLog 6), while confirming that managers use it to track individual performance, attendance, and break discipline in real time.

5.3. Ideology: ‘Thin’ Objectivity with Organizational Friction

DA and ConLog managers both invoked the ideology of objectivity found in platform settings, presenting algorithmic decisions as based on data and, therefore, more just than subjective managerial judgment. However, in both legacy firms, this ideology faced what we term ‘organizational friction’ (Zuboff, 1988): human managers still formally controlled outcomes, were legally responsible for decisions, and had discretion over how to interpret algorithmic outputs. At DA, partners often “overrode” the dashboard’s project-allocation recommendations based on relational knowledge and strategic priorities. At ConLog, WMS productivity alerts were inputs to, not directives for, line managers’ own judgment.

At DA, the objectivity rationale attached to the PPS operated as a legitimating frame rather than as an operational rule. Management presented the system as a data-driven basis for decisions:

“When you need to standardize something temporarily, it is not as profitable as a regular project, but in the longer term, you can use it to make the operations more efficient. Here, the system supports the decision-making process by showing if standardization is necessary given the previously logged data.”

(Data Analytics 3)

Yet managers were explicit that the metric did not, on its own, determine assessments. The Operations Director (Data Analytics 4) stressed that raw productivity figures were routinely contextualized against experience and team composition:

“…we know that the best employee is not the one whose productivity is the best. There were also quite extreme cases of this. There have been times when someone comes in and does a basically not-so-great job, but immediately, his numbers are enormous. If a group-level result is good, then we guess that if there were three juniors and two seniors, it was obviously the seniors who caused this group’s production to be so good and not the guy who just came from university.”

(Data Analytics 4)

The same logic governed the system’s automated signals, which prompted rather than dictated action:

“It works more and more proactively so that colleagues see for themselves that the price needs to be renegotiated with the client. However, there are cases when… the manager has to intervene. But usually, the colleague sees this, looks at it, if it is OK, and then informs the client that it is necessary to intervene here. But, of course, we also follow it as a leader. If it should be done differently, then we indicate.”

(Data Analytics 4)

The data-driven framing therefore legitimated the introduction of metric-based evaluation while formal interpretive authority remained with managers—a “thin” objectivity claim in which the algorithmic output informs, but does not displace, human judgment.

We refer to this as a ‘thin’ objectivity claim: the ideology is present, but the structural functions are minimized. It provides a rationale for the introduction of the algorithmic system and reduces dissent against the metric-based evaluation, but it does not and legally cannot substitute the employer’s official responsibility for the subsequent decisions. This difference is equivalent to Zuboff’s automate/informate fork: in both traditional instances, management used AM partly in an informating manner (producing data to inform managerial decision-making) rather than solely in an automating manner (substituting managerial choices with algorithmic outputs). This deployment is consistent with the “thin” nature of the ideology.

5.4. Modality: Constrained Co-Optation

The co-optation modality, like rating, ranking, recommending, and nudging, was present in both traditional cases, but in a structurally constrained form. At DA, client ratings of consultants’ project deliverables fed into performance assessments, partially replicating the customer-as-quasi-manager dynamic. However, these ratings were mediated by HR processes, subject to employee contestation, and filtered through annual performance review cycles rather than operating in real time. The gamification elements typical of platform AM, like badges, streak bonuses, and surge incentives, were absent: co-optation operated through professional-identity framing rather than game mechanics.

At ConLog, the WMS’s productivity metrics functioned as ranking and recording mechanisms, and automated alerts constituted a ‘recommending’ function for line supervisors. The presence of collective agreements governing performance standards significantly constrained the platform AM pattern in which the operator unilaterally sets the rules of the game. The scope of this channel should not be overstated, or confused with participation in the design of the system. Our ConLog material indicates that the works council is ex post, not ex ante: it does not participate in the specification of the WMS, nor in the setting of its parameters, but it can object to the productivity standards that result from those parameters once the system is operating, through the information and consultation rights that apply to works councils under the Hungarian Labor Code and the collective agreement in force at the site. The episode reported to us, where contested targets during the fieldwork period were brought to the works council and then revised, is compatible with, rather than contradicts, the statement in ConLog 6 (cited in Section 5.5) that neither operators nor their representatives are involved in system design or parameter setting: the two accounts refer to different stages of the same process. Ex post contestation of outcomes is not co-determination of the algorithmic system itself.

This channel of Reynaud’s (1989) “régulation conjointe”, joint regulation, has no equivalent in the platform cases, where platform Terms and Conditions are set unilaterally and are not subject to collective negotiation (Reynaud, 1989).

5.5. Accountability: Hybrid and Partially Re-Anchored

The sharpest contrast with the platform setting appears on the accountability dimension. In both traditional cases, algorithmic outputs inform decisions, but formal managerial accountability is preserved rather than displaced. At ConLog, when an automated alert triggered a disciplinary process, it was the line manager, not the algorithm, who held the conversation, and the firm, not the software vendor, that bore legal responsibility for any sanction. At DA, partners who disagreed with the dashboard’s allocation routinely overrode it, retaining personal accountability for major resourcing decisions.

Accountability is thus “partially re-anchored”: not fully twisted as in the platform cases, but not the transparent vertical accountability of classical bureaucracy either. The degree of re-anchoring is institutionally conditioned. Hungary’s low union density means it operates primarily through managerial discretion and HR procedure rather than robust collective bargaining, distinguishing the Hungarian cases from more coordinated Western European comparators and confirming Bohle and Greskovits’s account of CEE capitalism and the ETUI’s (2024) documentation of Hungary’s weak collective-voice infrastructure. The limits of worker voice are visible in workers’ own description of their role in these systems:

“We are definitely informed about the changes and new implementation. Once the system is live, we can give direct feedback to the management and to the team in charge of deployment. We are not involved in system design or parameter setting. Neither the workers’ representatives.”

(ConLog 6)

This account, together with Section 5.4, precisely determines the axis of joint regulation and resolves a seeming contradiction between the two sections. At ConLog, worker representation is absent at the configuration stage—that is, the specification of the WMS and the setting of its parameters—and present, in a weak consultative form, at the stage at which the outputs of those parameters become binding performance expectations. The first stage is covered by ConLog 6. The target dispute described in Section 5.4 falls into the second stage. The re-anchoring of accountability that we describe, therefore, operates downstream of the algorithm rather than over it and should not be read as co-determination.

Formal authority for algorithmically informed decisions at DA still rested in the hands of managers, rather than being displaced onto the system. “There are cases when… the manager has to intervene… we also follow it as a leader,” said the Operations Leader (Data Analytics 4), adding that the PPS was suggestive rather than directive. “If it should be done differently, then we indicate,” and characterized the metrics as inputs to managerial judgment: “these numbers are good because managers can draw conclusions from them and then take the appropriate steps.” As there was no works council or union at DA, this re-anchoring of accountability happened not through a formal collective contestation channel, but through managerial discretion within a transparency-based, supportive review culture whose stated purpose was “not to punish, but to detect problems early.”

6. Comparative Analysis: The Partial Möbius Effect and Its Scope Conditions

6.1. Cross-Case Comparison Matrix

Table 5 presents the cross-case comparison of the five principles across all six cases. Three patterns are immediately apparent from the matrix.

Table 5. Cross-case comparison of Stark and Vanden Broeck’s (2024) five AM principles.

Principle	Wolt/Bolt/Upwork	Uber (HU)	DA (Data Analytics)	ConLog (Logistics)	Partial Möbius?/Does the Principle Travel?
Org. form	Full platform (Möbius): no boundary; contractor status	Full platform; boundary-circumvention logic exposed and contested in the 2016 regulatory exit (documentary/legal evidence)	Firm boundary preserved; platform-style dashboard inside	Firm boundary preserved; WMS inside hierarchy	Requires re-specification—no Möbius analog inside the firm; partial Möbius effect instead (DA & ConLog)
Object of mgt	Workers as data profiles/interchangeable nodes; customers as co-managers	Not evidenced in this study—documentary case; observed operational practice would require interview/observational data (see Section 3.3)	Consultants managed as profiles, but employment rights retained	Warehouse workers as throughput metrics; limited contestation	Travels intact—employees as users, constrained by employment law
Ideology	‘Thick’: algorithm = neutral, apolitical; denies employer status	‘Thick’: ‘technology company’ framing central to Uber’s legal/regulatory self-presentation (documentary); worker-level framing not evidenced	‘Thin’: objectivity claim present but overridden by managerial discretion	‘Thin’: ‘system removes bias,’ but line managers retain formal authority	Partial—travels in form; functional weight attenuated
Modality	Full co-optation: ratings, surge pricing, gamification, real-time nudges	Not evidenced in this study—documentary case; co-optation mechanisms are not observable from the regulatory record (see Section 3.3)	Constrained co-optation: client ratings mediated by HR, annual cycles	Constrained co-optation: WMS metrics, no gamification; collective agreements present	Partial—travels in form; co-optation present but institutionally bounded
Accountability	Fully ‘twisted’: responsibility displaced to courier/customer; platform unaccountable Upwork: displaces accountability to the client, while the platform’s rating algorithm enforces the client’s judgment with no internal appeal	Fully ‘twisted’: the ‘tech-company’ argument constructs a legal displacement of responsibility (documentary/legal evidence)	Hybrid: algorithm informs; partner retains formal accountability; HR channel open	Hybrid: line manager conducts disciplinary process; works council can contest targets ex post, but has no role in system design or parameter setting	Requires re-specification—accountability partially re-anchored in the employment relation

Note: AM = algorithmic management; DA = Data Analytics firm (pseudonymized); ConLog = logistics subsidiary (pseudonymized); WMS = warehouse management system. Uber is a documentary/regulatory comparator (see Section 3.3); its cells are based on legal, regulatory and media records, not interview or observational data, and it is marked ‘not evidenced’ for principles requiring observed operational practice.

The last column classifies each principle at three levels: travels intact (the principle is valid in traditional settings without re-specification), partial (the principle exists in form but its functional weight is institutionally attenuated), and requires re-specification (the principle as stated for platforms has no direct analog off the platform and must be reformulated). This classification is used in Section 1, Section 6.2 and Section 8.

6.2. Which Principles Travel and Which Require Qualification

The only principle that remains substantially intact regardless of the organizational context is the object of management. Two other principles move in form, but not in functional weight, and are classified as partial in Table 5 for this reason. Thus, we distinguish between the presence of a principle and its structural efficacy, and it is this distinction and not a mere number of principles that the analysis below sets out. In the five cases for which we have interview evidence, we find a strong presence of the ideology of algorithmic objectivity, the claim that algorithmic decisions are neutral and apolitical. It is most pronounced among the Upwork respondents, who largely endorse rather than contest it (Section 4.3). In the Uber case, the same claim is visible only in documented corporate and legal discourse, so we treat it as present in discursive form, and do not claim that it was operative at the level of worker experience. It serves as a legitimation device irrespective of the organizational form, although its structural weight depends on the presence or absence of formal managerial accountability. The modality of co-optation is also present across the interview-based platform cases (Wolt, Bolt and Upwork) and, in attenuated form, the two traditional firms, but its mechanisms and reach differ: in platform settings, co-optation works in real time, without the constraints of an employment relationship; in traditional firms, it is mediated by HR processes, collective agreements, and managerial discretion. The object-of-management principle operates in both cases; AM targets workers as data profiles rather than whole labor subjects, but the legal architecture of the employment relationship constrains the extent to which this reduction is achieved. The principle is there in both cases, but it works less than it does on the platform. Ideology and modality are therefore present in both settings but work less off the platform than on it, which is why we classify them as partial rather than as traveling intact. The object of management, however, remains unchanged in the sense that the principle itself does not require reformulation: workers are reduced to data profiles in each case, and employment law limits the consequences of that reduction without changing the principle.

Two principles require outright re-specification off the platform, rather than the attenuation described above. The principle of organizational form requires the strongest: the Möbius form has no direct analog in traditional firms, and what we observe at DA and ConLog is better captured by the concept of a partial Möbius effect, developed in Section 6.3. The accountability principle likewise needs to be re-specified: the “twisted” triangular structure is confirmed in the platform context but becomes a hybrid structure in traditional firms, where formal legal accountability cannot be fully substituted, and channels of joint regulation remain open.

Upwork data indicate that the twisted structure is not monolithic even within the platform category. At Wolt, Bolt and Uber, responsibility is twisted away from the platform toward an unaddressable “tech company” construct: the courier or driver has, at least nominally, a target to direct a complaint at before it is displaced into legal argument. Toward the platform’s own “tech company” construct: the courier or driver can at least direct a complaint at the platform, even if it goes unanswered and is ultimately displaced into legal argument. At Upwork, evaluative authority is delegated to the client from the outset, and a disputed rating is enforced by the marketplace’s own visibility algorithm with no internal channel of appeal. One respondent, unable to have a contested rating removed, described continuing to complete assignments he considered unreasonable “because I could not just simply quit, due to the rating system,” concluding “I did not [solve it]. It is what is.” We term this variant client-displaced accountability: responsibility is twisted not toward the platform’s own legal fiction, as at Wolt, Bolt and Uber, but toward the client relationship, with the platform’s algorithm supplying enforcement rather than the addressee of complaint. Client-displaced accountability sits alongside twisted and hybrid, partially re-anchored accountability as a third empirical variant, indicating that the accountability principle differentiates not only along the platform/traditional axis but also by the structural position of the accountable party within the platform relationship itself.

6.3. The Partial Möbius Effect: Concept, Genealogy, and Scope Conditions

We introduce the concept of the “partial Möbius effect” to name the bounded penetration of platform-like algorithmic governance into organizations that retain bureaucratic boundaries, employment contracts, and institutional channels of collective voice. Its epistemic status should be stated in the beginning. The concept was derived inductively from two traditional cases, DA and ConLog, and has not been tested against a third, independent case which did not contribute to its formulation. It is therefore articulated not as a validated theoretical result but as a testable hypothesis about the conditions under which platform-style algorithmic governance is imported into bounded organizations. The genealogy, demarcations and scope conditions set out below should be understood in terms of what such a hypothesis says and how it might be affirmed, revised or rejected, rather than as reporting an accomplished theoretical result.

The concept’s genealogy derives directly from Watkins and Stark (2018) and Stark and Pais (2020). The Möbius organizational form describes firms that co-opt assets, resources, and activities lying outside the firm, producing governance with “rules but not bureaucracy, rankings but not ranks, and accounts but no accountability.” The “partial” qualifier marks the importation of these governance properties, co-optation, rankings without ranks, accounts without accountability, into a setting from which their original boundary-dissolving context is absent. Client-displaced accountability, introduced in Section 6.2, is a differentiation within the platform category itself and is analytically distinct from the partial Möbius effect developed here, which concerns the bounded importation of platform-like governance into organizations that retain a firm boundary; the two are not alternative labels for the same phenomenon.

Within the LPT tradition, the partial Möbius effect can be understood as a recombinant hybrid control form: it overlays AM’s rating, ranking, and real-time feedback mechanisms onto the bureaucratic control structure that already governs the firm (Edwards, 1979). Kellogg et al. (2020) anticipated this dynamic in noting that AM mechanisms produce “more positive outcomes for workers” in some organizational configurations but did not specify which configurations. The partial Möbius effect is offered as one candidate specification rather than as the specification. In Wood’s (2021) continuum of automation, the partial Möbius zone maps to conditional automation: algorithmic systems can direct and assess work but require human managers to act on outputs, with a legal floor that bars the complete displacement of managerial judgment. That floor’s contents must be specifically stated. There is no general antecedent requirement of meaningful human involvement in algorithmic decision-making in Article 22 of Regulation (EU) 2016/679 (2016) (GDPR). Under Article 22(1), the data subject has the qualified right ‘not to be subject to a decision based solely on automated processing, including profiling’ which produces legal effects concerning him or her or similarly significantly affects him or her, a right which is qualified by the exceptions in Article 22(2). Where such processing is nevertheless permitted, Article 22(3) requires the controller to provide appropriate safeguards for the data subject, including at least a right to obtain human intervention on the part of the controller, to express a point of view and to contest the decision. The floor is thus a right of intervention and contestation after a decision has been made, rather than a duty of human involvement built into the decision beforehand. Specifically, in the case of platform work, Article 10(5) of Directive (EU) 2024/2831 (2024) introduces a prior requirement according to which any decision to restrict, suspend or terminate the contractual relationship or the account of a person performing platform work must be taken by a human being.

The concept is also structurally distinct from Weil’s (2017) fissured workplace. Weil describes control retained via specifications while work is externalized beyond firm boundaries; the partial Möbius effect is, in one sense, the inverse, importing control tools while retaining accountability obligations. What the two share is the partial separation of the exercise of algorithmic control from the full bearing of accountability, producing what might be called a “partial fissure” within the firm interior.

Hybrid accounts of algorithmic control are already manifold, so it is necessary to specify what the partial Möbius effect adds to the existing theoretical literature and what it inherits from it. Table 6 links the concept to the adjacent constructs it relies on, and the argument is supported by three points of demarcation. First, when considering the accounts of recombinant control in the labor process tradition (Edwards, 1979; Kellogg et al., 2020), the concept is not merely the observation that algorithmic control exists on top of bureaucratic control, which those accounts already establish. It indicates a direction of travel, in that governance properties developed in a boundaryless setting are imported into a bounded one. It indicates which property fails to make the journey: the displacement of accountability, because the employment contract continues to supply an addressee for it. Secondly, the concept is intended as an explanation rather than a description: it seeks the limit of that diffusion in organizational topology rather than in the intensity or maturity of the technology, unlike the literature on the platformization of work (Fernández-Macías et al., 2023; Gonzalez Vazquez et al., 2025; Schmid & Wiesche, 2026) that demonstrates how platform-like practices are permeating standard employment. Third, the relation is one of inversion rather than extension, as set out above, relative to the fissuring described by Weil (2017).

Table 6. Positioning the partial Möbius effect in relation to adjacent theoretical constructs.

Construct (Source)	Core Claim	Relation to the Partial Möbius Effect	What the Partial Möbius Effect Adds, or Where It Differs
Möbius organizational form; platform AM (Watkins & Stark, 2018; Stark & Pais, 2020; Stark & Vanden Broeck, 2024)	Value creation is distributed across a dissolved firm boundary; governance operates through rules without bureaucracy, rankings without ranks, accounts without accountability.	Direct parent concept; the partial Möbius effect is defined as its bounded, intra-organizational counterpart.	Specifies what happens when the governance properties are imported without the boundary dissolution that generated them, and identifies accountability as the property that does not travel.
Recombinant and hybrid control (Edwards, 1979; Kellogg et al., 2020)	Algorithmic control is a fourth control form that layers onto, rather than replaces, technical and bureaucratic control.	Closest neighbor: the partial Möbius effect is a species of recombinant control.	Adds directionality and asymmetry: it states where the imported logic originates and which of its components is blocked, whereas hybrid-control accounts describe coexistence without specifying the blockage.
Manufacturing consent (Burawoy, 1979)	Consent is produced at the point of production through games that generate effort while obscuring exploitation.	Supplies the micro-mechanism through which co-optation operates inside the firm.	Situates the consent mechanism institutionally: in traditional firms, the game is played within an employment relation that also supplies a route of appeal, which alters its stakes.
Fissured workplace (Weil, 2017)	Lead firms retain control through specifications while shedding the employment relation, severing control from accountability.	Structural inverse: control tools are imported while the employment relation is retained.	Describes the internal rather than external case; what the two share is a partial separation of control from accountability, which we term a partial fissure within the firm interior.
Platformization of work (Fernández-Macías et al., 2023; Gonzalez Vazquez et al., 2025; Schmid & Wiesche, 2026)	Platform-like monitoring and management practices are diffusing into standard employment, with variants of differing intensity.	Empirical precursor; these studies establish the phenomenon the concept seeks to explain.	Proposes a structural reason for the observed ceiling on diffusion, rather than treating intensity as a continuum without a limiting mechanism.
Politics of AM implementation (Krzywdzinski et al., 2025; Dupuis, 2024)	AM outcomes are shaped by contestation among actor coalitions within the firm and by union power on the shop floor.	Complementary rather than competing; supplies the agency that our largely structural account underplays.	Adds the cross-form comparison these single-setting studies do not make; conversely, it lacks their resolution on process, which we record as a limitation (Section 8).
Automating versus informing (Zuboff, 1988)	The same infrastructure can be deployed to surveil and standardize or to generate knowledge that workers can use.	Cross-cutting distinction rather than a rival concept.	The partial Möbius configuration is compatible with either deployment; DA leans toward informing and ConLog toward automating, which is one reason the two cases are not identical.
Automation continuum (Wood, 2021)	Management automation is a graded continuum from assistance to full automation, not a binary.	Supplies the first axis of the conceptual space in Section 6.3.	Uses the continuum as an ordering device for organizational forms rather than for tools; no measurement is claimed on either side.

Source: Authors’ own design.

Equally important is what the concept does not assert. It does not claim that DA and ConLog are becoming platforms, nor does it propose a new type of organization. It is a configuration of control and accountability. Its warrant at this stage is analytical rather than confirmatory: it is induced from, but not yet tested against, the empirical material. Nor does it displace the accounts it draws on: the characterization of the platform as a permissive potentate by Vallas and Schor (2020) and the analysis of the micro politics of implementation by Krzywdzinski et al. (2025) describe dynamics that our concept does not capture, and the distinction between automating and information made by Zuboff (1988) cuts across it, since a partial Möbius configuration can be tilted toward either pole. The lineage of the concept would not be mistaken if the reader were to interpret it as a redescription of recombinant hybrid control in a Möbius vocabulary. Our claim is more limited: this vocabulary renders visible a particular asymmetry, namely the transfer of co-optation without the concomitant transfer of accountability displacement, which the current hybridity accounts leave unspecified, and it is to name the asymmetry that it becomes possible to state the scope conditions in the next subsection.

We suggest a two-axis conceptual space in which the partial Möbius effect can be situated relative to full Möbius (platform) governance on one side and traditional bureaucratic AM on the other. We emphasize from the beginning that this space is a heuristic ordering device and not a measurement device. Neither axis is operationalized as a scale in this paper: no indicators were scored, no index was constructed, and the placement of our cases within the space is an interpretive summary of the qualitative evidence presented in Section 4 and Section 5, not a measured position. For this reason, we do not define cut-off points or thresholds. Threshold language would require a metric that we do not have and that we could not derive from six qualitative cases, so the framework is presented as conceptual work, the calibration of which is left to future research (Section 8).

Axis 1: Degree of management automation (Wood, 2021). This runs from assistance (algorithmic tools support human managers) through partial and conditional automation to high and full automation (human managers respond to, or are replaced by, algorithmic outputs). In these terms, our two traditional cases are best characterized as sitting at conditional automation: AM handles direction and evaluation, but discipline requires human managerial activation. This is a qualitative characterization arrived at by reading the case material against Wood’s categories, not a scored position on a calibrated scale.

Axis 2: Degree of surviving joint regulation (Reynaud, 1989; Wood, 2021). This captures the strength of institutional channels through which workers and their representatives can contest, negotiate, and co-produce the rules governing algorithmic decisions, including employment-contract protections, GDPR rights, works council consultative rights, and collective-bargaining coverage. In the two traditional cases, institutional embeddedness is sufficient to preserve formal accountability and some capacity for contestation, while co-optation remains present enough to distinguish the setting from purely bureaucratic management. The axis should be understood as combining two analytically distinct capacities: ex ante co-determination over the design and parameterization of algorithmic systems and ex post contestation of the performance standards that those systems produce. Our two Hungarian cases exhibit only the second capacity, which is one reason we describe them as weakly rather than strongly jointly regulated (see Section 5.4 and Section 5.5).

The space is bounded along both axes, and the two boundaries can be described as directions of movement, rather than as measured thresholds. In cases where management automation is low, the co-optation dynamic is insufficient to differentiate AM from ordinary bureaucratic supervision. In cases where management automation is high and joint regulation is weak, there is little to prevent the circumvention of the employment relation and convergence on platform logic in the configuration. These are conjectures about direction, not predictions about transition points. Proving that such transition points exist, and finding them, would require indicators for each axis, an explicit measurement strategy, and variation across a far larger number of organizations than we study here. Within these limits, the two traditional cases can be compared: DA, with its higher task complexity, stronger relational governance, and more active use of the informing mode (Zuboff, 1988), is further from platform logic than ConLog, whose routine tasks, thin union presence, and WMS-driven productivity targets place it closer to it. This is what we mean by a relative ordering of two cases, and we make no inference about where either would fall on a measured scale.

The platform cases lie outside the partial Möbius zone, but they are not identical within it. Upwork combines high automation of evaluation with client-controlled allocation, and joint regulation is virtually absent: there is no collective agreement or works council, and contracts are concluded across borders, largely with clients abroad. It therefore sits at the weak end of Axis 2 even more clearly than Wolt and Bolt, whose operation in the Hungarian market exposes them to national regulatory attention (Section 6.5). As with the traditional cases, this is an interpretive placement, not a measured one.

6.4. The Durability of the Configuration: A Question Left to Future Research

An obvious further question is whether the configuration described above is durable or transitional. Wood et al. raise the possibility that AM adoption in traditional firms may accelerate the logic of fissuring as the algorithmic infrastructure matures, the Amazon fulfillment-center pattern being one example of this trajectory (Wood et al., 2019). We raise the question here because it follows naturally from the concept, but we do not attempt to answer it, and we make no claim that the partial Möbius effect constitutes a stable hybrid equilibrium.

Our design does not give an answer to this question, and we think it would be premature to turn the question into a proposition. The six cases were studied by a cross-sectional design. Apart from the ex post revision of contested ConLog targets during the fieldwork (Section 5.4), no within-case evidence points to structural change over time in either traditional firm. Both traditional cases are drawn from a single national context where the configuration of industrial relations itself is distinctive and may be doing much of the work that would otherwise be attributed to the configuration as such. Any claim about durability based on this evidence base would be a claim about two Hungarian firms seen once, not a claim about a general organizational form. However, the material does support a description of the conditions that exist at present in these two firms. It is worth recording these as they are what a longitudinal study would need to track. First, we find that the observed automation levels are at the conditional level but not at the high level, which is consistent with the observation by Wood (2021) that full automation is an “ideal type that is still a long way from being achieved”. Second, there is a legal floor in the form of GDPR Article 22, the EU Platform Work Directive (Directive (EU) 2024/2831, 2024) and national employment law, which did not exist in platform labor before regulation and which limits the removal of human managerial responsibility. Third, the consent infrastructure described by Burawoy (1979) continues to work through the employment relation, which probably produces organizational path dependencies. We cannot tell from our evidence whether these conditions are stable features of the configuration, whether they decay over time, or whether they are simply overtaken as algorithmic infrastructures mature.

Thus, we leave the question open for future work and indicate what would settle it. Longitudinal within-case designs tracking the same firms over several years would reveal if the algorithmic infrastructure in traditional firms is used as a staging ground for workforce externalization, i.e., if systematic movement toward higher automation and diminished joint regulation is taking place, as the fissuring scenario raised by Wood et al. (2019) foresees. A cross-national comparison would test whether what we are seeing is a reflection of the configuration or the Hungarian institutional setting. The InCoding partner-country material from Denmark, Germany and Spain, where works councils and bargaining coverage are much stronger, would be a natural starting point, as would sectors other than knowledge services and logistics. Until such evidence is forthcoming, the durability of the partial Möbius configuration must be regarded as an open question, not as a finding, proposition or conditional expectation of this paper.

6.5. Institutional Mediation: The Hungarian Case

In all six cases, the national institutional context mediates AM outcomes in ways not fully captured by the framework of Stark and Vanden Broeck or the LPT tradition. Weak union density and fragmented industrial relations in Hungary lead to a limited collective capacity of workers to challenge AM governance within platform and traditional workplaces (ETUI, 2024; Szabó, 2013). The re-anchoring of accountability in traditional firms is real but thin: it is based on managerial discretion and HR procedure rather than on strong collective bargaining. The works council at ConLog, by contrast, does not have bargaining power as it does in Germany or Scandinavia, nor, as noted in Section 5.4 and Section 5.5, does it have authority over the configuration of the algorithmic system itself; its powers are limited to the performance standards that are derived from the system once it is operational. Upwork is a partial exception: as a global marketplace connecting Hungarian freelancers with mostly foreign clients, its governance is set largely outside Hungarian institutions. Its relevance for the argument is precisely that it shows what platform AM looks like where national mediation is weakest. Whether the transposition of the Platform Work Directive will reach such cross-border relationships remains an open question.

The regulatory vacuum that led to the Uber debacle and the following regulatory learning that informed the market strategies of Bolt and Wolt, on the other hand, show that platform AM is not institutionally unanchored but rather constructed by national regulatory fields, albeit underdeveloped ones. Because the Uber portion of this argument hinges on the documentary record of a single regulatory episode, we offer it as an interpretation to be tested in other national settings, not as a general finding. This resonates with Bohle and Greskovits’s account of CEE capitalism and Meardi’s emphasis on the institutional conditioning of labor outcomes (Bohle & Greskovits, 2012; Meardi, 2012). Future AM theorizations should, in the spirit of Reynaud, include an explicit institutional dimension as a constitutive, rather than background, variable specifying how the balance between control regulation and autonomous regulation at the national and workplace levels determines the partial or full realization of Möbius co-optation (Reynaud, 1989).

7. Discussion: Theoretical and Regulatory Implications

7.1. Theoretical Contributions

Rather than restating the empirical patterns reported in Section 4, Section 5 and Section 6, this section considers what they imply for how algorithmic management (AM) should be theorized. The overarching implication is that AM is best analyzed as an organizational logic whose empirical manifestations depend on organizational form and institutional embeddedness, rather than as a phenomenon unique to platform firms. Three more specific contributions follow from this.

First, the paper offers a cross-form empirical operationalization of the five principles proposed by Stark and Vanden Broeck (2024), specifying where each principle travels intact, where it travels in form but is institutionally attenuated, and where it requires re-specification (see Table 5 and Section 6.2). We are not aware of a prior comparative application of the framework across platform and non-platform settings, but we make no priority claim and present the operationalization as one possible reading of the framework rather than as a definitive one. In doing so, it moves beyond documenting that AM varies across organizational settings, a point already established in the emerging cross-organizational literature (Wood, 2021; Fernández-Macías et al., 2023; Gonzalez Vazquez et al., 2025), by proposing a candidate mechanism, the partial Möbius effect, that may account structurally for why such variation arises. Whether the mechanism holds beyond these cases is not something the present design can show.

Second, the paper develops the partial Möbius effect as a testable hypothesis about AM in traditional organizations rather than as a finished theoretical account, since it was induced from two cases and still awaits testing on independent material. The concept synthesizes three lineages: the organizational-topology tradition (Watkins & Stark, 2018; Stark & Pais, 2020), the labor process theory account of control (Edwards, 1979; Burawoy, 1979; Thompson, 1990; Kellogg et al., 2020), and the fissured-workplace tradition (Weil, 2017). Its scope conditions are defined along two axes, namely the degree of management automation and the extent of surviving joint regulation. It offers a candidate framework for systematic cross-case and cross-national comparison. Whether it extends beyond the Hungarian cases analyzed here is itself an empirical question that the hypothesis invites future research to settle.

Third, by situating the analysis in Hungary, the paper suggests that institutional sensitivity should be treated as a constitutive rather than a background concern in AM research. Identical tools of algorithmic governance, such as WMS productivity monitoring and customer-rating systems, were associated with divergent outcomes in accountability and voice across our cases, depending on the institutional embeddedness of the employment relationship. This extends the comparative capitalisms literature (Bohle & Greskovits, 2012; Meardi, 2012) into the AM domain and offers an initial CEE reference point against which future cross-national comparisons can be calibrated.

These contributions intersect with two ongoing debates. The first concerns the politics of AM implementation. Krzywdzinski et al. (2025) foreground the micro-politics of actor coalitions within a single traditional firm; the present study broadens the comparative horizon across organizational forms and varies the institutional context. The two studies are complementary: Krzywdzinski et al. (2025) show that AM is politically contested even within firms, while our analysis shows that the terms and stakes of that contest vary both structurally, depending on whether a formal employment relationship exists, and institutionally, depending on the national collective-voice infrastructure.

The second debate concerns “pro-worker AI” (Acemoglu et al., 2026). Whereas platform-based AM operates largely through automation, behavioral steering and the standardization of work processes, the DA case points to a more augmentative use of algorithmic systems, in which data-analytic tools support employee decision-making, learning and problem-solving. AM should therefore not be analyzed exclusively as a control technology: its consequences depend on whether algorithmic systems are deployed primarily to automate human judgment or to complement and enhance worker capabilities. The contrast between platform and traditional settings thus mirrors the broader distinction between automation-oriented and augmentation-oriented trajectories of digital transformation. Relatedly, the partial Möbius effect carries implications beyond managerial control and accountability: unlike platform settings, where AM often operates in the absence of formal channels of worker representation and workplace democracy (Makó et al., 2026), the traditional firms examined here retain elements of employee voice through employment contracts, managerial hierarchies and institutionalized participation mechanisms.

7.2. Regulatory Implications

The findings are relevant to the EU Platform Work Directive (Directive (EU) 2024/2831, 2024) and the EU Artificial Intelligence Act (Regulation (EU) 2024/1689, 2024). The rebuttable presumption of employment in the Directive (Article 5) and its rules on algorithmic management in Chapter III align well with the accountability problems that arise in platform settings. These duties are spread across adjacent articles, rather than being concentrated in one provision, and it is worth noting which obligation sits where. Article 8 considers the processing of personal data by automated monitoring or decision-making systems as processing likely to result in a high risk within the meaning of Article 35(1) GDPR. It requires the platform to consult persons performing platform work and their representatives when carrying out the data-protection impact assessment. Article 9 provides for transparency: platforms shall inform persons performing platform work, workers’ representatives and, upon request, national competent authorities on the use of such systems, on the categories of data monitored and on the main parameters and their relative importance in automated decision-making. Article 10 establishes oversight by humans, requiring platforms to oversee those systems and, with the involvement of workers’ representatives and in any event every two years, to assess the impact of individual decisions taken or supported by them. Where that oversight identifies a high risk of discrimination at work or finds that individual decisions have infringed rights, the platform must take the steps necessary, including, if appropriate, modification or discontinuation of the system. Article 10(5) requires that any decision to restrict, suspend or terminate the contractual relationship or the account of a person performing platform work must be taken by a human being. Article 11 provides for the right to an explanation, the right to a substantiated response to a request for review within two weeks and the right to rectification. The duty to properly assess and mitigate risks is set out in Article 12(1), which requires platforms to “assess the risks of automated monitoring and automated decision-making systems” to workers’ safety and health, and in particular the risks of work-related accidents and psychosocial and ergonomic risks, to assess whether the safeguards of those systems are appropriate to the risks identified and to introduce appropriate preventive and protective measures. Finally, Article 13 extends information and consultation of workers’ representatives to decisions likely to lead to the introduction of, or substantial changes in, such systems, which is a weaker entitlement than a right to representation in system design. It is the whole system of transparency, oversight, review and risk-evaluation responsibilities, not any single article, which addresses the ‘twisted accountability’ structure outlined above.

However, our analysis of traditional workplaces suggests that AM governance challenges are not unique to platform firms and the Directive’s platform-specific scope is a limitation. In this regard, the AI Act’s identification of employment-related AI systems as high-risk is better targeted (applying regardless of organizational form), but its implementation will have to contend with the partial Möbius effect: an automated WMS in a logistics firm, or a project-allocating dashboard in a consulting firm, may fall within the high-risk identification without most firms realizing it.

The Hungarian case also illustrates the opportunities of regulatory learning and the dangers of regulatory vacuum, with the caveat that its Uber component is documented only through public regulatory, parliamentary, legal, and media records. The absence of a coherent framework between 2014 and 2024 led to the Uber fiasco, the lack of social protection for platform workers, and competitive advantages for platforms that were willing to operate in the gray zone (Makó et al., 2023). The InCoding evidence from Hungary (Makó et al., 2024) confirms the underdevelopment of worker information and consultation rights on algorithmic systems even where formal mechanisms exist. A key test of whether the regulatory learning from the Uber episode has become institutionalized will be Hungary’s transposition of the Platform Work Directive (deadline: December 2026). Given the country’s record on transposing EU labor law (documented by ETUI (2024) as being among the slower Member States), this is a question for continued monitoring rather than confident prediction.

8. Conclusions and Future Research

Drawing on 43 interviews, participant observation and documentary analysis across six Hungarian cases from the CrowdWork21 and InCoding projects, this paper examined how the five principles of algorithmic management identified by Stark and Vanden Broeck (2024), namely organizational form, object of management, ideology, modality and accountability, operate across platform and traditional organizations. The short answer to the research question is that AM appears to travel across organizational forms, but not uniformly: only the object of management travels intact, ideology and modality travel in form but are attenuated by institutions, and organizational form and accountability require re-specification. We capture the re-specification of organizational form as the partial Möbius effect, and that of accountability as hybrid, partially re-anchored accountability, the former positioned in Section 6.3 in relation to the recombinant-control, fissuring and platformization studies, and the national institutional context mediates the realization of all five (see Section 5, Section 6 and Section 7). We stress that the partial Möbius effect is stated here as a testable hypothesis and not as a final theoretical achievement of this paper. It was inductively derived from two traditional cases, Data Analytics and ConLog, and has not been tested on a third, independent case. Whether it holds, and under what organizational and institutional conditions, is a task for subsequent articles and research, not a matter settled here. Whether the resulting hybrid configuration is a temporary stage on the way to full platformization or a lasting one cannot be determined from a cross-sectional design covering six cases in a single country. We therefore leave the question open rather than advancing a proposition about it, and set out in Section 6.4 the longitudinal and cross-national evidence that would be needed to answer it.

Limitations and Future Research

Four limitations of the present study define the agenda for future research.

First, the design is cross-sectional and confined to a single national context. Longitudinal, multi-country research is needed to test whether the partial Möbius effect is institutionally stable or instead marks a trajectory toward fissuring and re-platformization. For the same reason, the partial Möbius effect is presented in this paper as an analytically useful interpretation of the six cases rather than as an independently validated construct: it was developed from the DA and ConLog material and has not yet been tested against a case that did not contribute to its formulation (see Section 3.3).

Second, the traditional-firm evidence rests on two organizations in knowledge services and logistics; sector-level variation across healthcare, public administration, and manufacturing remains underexplored in the CEE region. Relatedly, the evidence base is uneven across cases: the Uber case is based on documentary rather than interview data (Section 3.3), and the platform and traditional cases vary in size and technology. Claims involving Uber are accordingly confined to the organizational form, discursive-ideology, and accountability dimensions that public records can support, and should be read as an interpretation of one documented regulatory episode rather than as generalizable evidence about Uber as a workplace. We deal with this by mapping all cases onto the common scale of the five principles and by limiting each case’s claims to the evidence it actually supports, but a fully matched, interview-based comparison across all six sites remains desirable and is a task for future multi-site research.

Third, the findings are predominantly structural in character. The data capture patterns of AM deployment, accountability displacement and institutional mediation, but register worker agency (resistance, workarounds, informal accommodation) only in scattered and indirect form, because CrowdWork21 and InCoding were designed primarily to document AM mechanisms and governance structures rather than to reconstruct workers’ tactical repertoires. Couriers in our sample described awareness of the “game” they were enrolled in and expressed skepticism about the neutrality of algorithmic decisions, but the interview protocols did not systematically probe strategies of resistance or mutual aid. Future research should build on emerging accounts of worker-led platform governance experiments (De Stefano, 2016), algorithmic resistance in warehousing, and collective organizing in gig work to develop a more dynamic picture of the partial Möbius effect as a contested terrain shaped from below.

Fourth, the two-axis space proposed for the partial Möbius effect, defined by the level of management automation and the extent of joint regulation, is conceptual work only. Neither axis has been operationalized, and the paper offers no measurement of either. Making the framework empirically usable would require, in sequence, indicator development for each axis, validation of those indicators across organizations, and only then an assessment of whether transition points between the configurations exist and can be located. Survey-based and multi-country comparative designs are the obvious route, and until that work is done, the space should be read as a way of ordering cases rather than as a model with thresholds.