Beyond Platform Work: Algorithmic Management in Platform Labor and Traditional Organizations—The Hungarian Experience

Abstract

Algorithmic management (AM) is spreading rapidly in platform work and traditional employment, yet empirical research remains platform-centric and skewed towards Western Europe, North America and China, leaving Central and Eastern Europe, especially Hungary, largely absent. This paper reports qualitative fieldwork on AM in two contrasting Hungarian firms, Data Analytics, a knowledge-intensive business services company, and ConLog, a contract logistics provider. Fieldwork in the EU-funded INCODING project (January 2022–July 2023) combined 15 semi-structured interviews, twelve internal and three external, with participant observation and documentary analysis. A sample limitation applies to one of the two firms: the Data Analytics evidence rests on only four internal interviews (two employees and two managers), so the findings for that case are indicative and exploratory rather than firm-level conclusions. A deliberately asymmetric comparison sets these findings against Hungarian platform labor, using secondary evidence on Uber, Wolt and Bolt from the CrowdWork21 project (2019–2021). Results show heterogeneity. In traditional firms, algorithmic systems sit within managerial hierarchies and produce hybrid human-algorithmic control. The older platform evidence documents real-time monitoring, opaque decision rules and market-mediated feedback that tighten control and erode autonomy. The contrast is a proposition for testing, not a measured difference. Workers in both settings adapt opportunistically, accommodating or partially resisting control. Hungary’s weak industrial relations institutions remain theoretically instructive despite limited generalizability. The contribution lies in the platform versus non-platform comparison, the institutional-bypassing argument and the design authority continuum.

1. Introduction

The digitalization of work is no longer just the preserve of Silicon Valley start-ups or gig-economy couriers. Algorithmic management (AM), that is, the use of software to automate or semi-automate organizational functions traditionally performed by human managers (Potocka-Sionek & Aloisi, 2024), is already working its way through supply chains, logistics hubs and knowledge-intensive business services throughout Europe. This change has significant consequences for the organization, monitoring and governance of work, and for the distribution of power between employers, algorithms and workers.

However, the academic literature on AM is still very platform-centric. The bulk of empirical work focuses on food-delivery couriers and ride-hailing drivers under near-complete real-time algorithmic control (Alasoini et al., 2023; Kellogg et al., 2020; Stark & Pais, 2020). Researchers increasingly argue that the traditional office should also be seen through this lens, as AM tools, such as dashboards, project-planning systems and warehouse-management software, are now influencing the work of millions of employees working under common employment contracts (González Vázquez et al., 2025; Krzywdzinski et al., 2024; Dupuis, 2024). Studies on AM in Central and Eastern Europe (CEE) remain comparatively scarce and unevenly distributed, despite the region’s post-socialist institutional legacies, weak industrial relations and rapidly growing logistics and platform sectors.

But the region is not unstudied, and the claim of novelty proffered here must be made against the regional record rather than in the absence of it. The ETUI Internet and Platform Work Survey (Piasna & Drahokoupil, 2019) provides evidence on the extent and composition of digital and platform labor in Poland, Hungary, Bulgaria, Latvia and Slovakia, and Kahancová et al. (2020) offer a comparative analysis of precarity in on-demand platform work in Hungary and Slovakia predating the CrowdWork21 project, and draw conclusions on autonomy and interest representation that anticipate several of ours. In terms of logistics, Polish warehouse work has emerged as a significant empirical field in its own right. Miszczyński and Zanoni (2025) examine the interplay of coercion and consent under algorithmically mediated control in an Amazon warehouse in Poland. The present paper attempts what this literature has not yet provided. The survey work establishes prevalence and precarity but does not open up the workplace. The Amazon studies open up the workplace, but at its most extreme and least representative point, a fulfillment center of a global e-commerce firm whose control regime is exceptional rather than typical. None of this work places platform and non-platform AM in the same national institutional setting within a single comparative frame. We therefore do not seek to fill a gap in CEE research, but rather to fill a gap in worker-level evidence on AM in ordinary CEE workplaces, in a domestically owned knowledge-intensive firm and a mid-sized logistics subsidiary, read against platform labor in the same country.

The paper makes three contributions: this offers new case-study evidence on algorithmic management in traditional workplaces in a CEE country at an empirical level and organizes this evidence through a deliberately asymmetric comparison with the existing record on Hungarian platform labor. In theory it brings together labor process theory, socio-technical systems thinking and institutional analysis to explain divergent outcomes of the same underlying logic of datafication in organizational and institutional settings. From a policy perspective, it derives lessons for the governance of AM in supply chains and logistics, which are timely, given the EU AI Act (2024) and the EU Platform Work Directive (2024). One constraint defines what this paper can and cannot claim and we state it upfront rather than only in the methodology: the platform-traditional comparison rests on asymmetric evidence (Section 3.5). The empirical core comprises two traditional firms studied directly in 2022–2023 through interviews and observation, material from the INCODING fieldwork also reported in the companion study by Makó et al. (2025) (Section 7.5); the three platform cases draw on platform artifacts and secondary evidence from the earlier CrowdWork21 project (2019–2021), with Uber as a historical and regulatory reference point after its exit from the Hungarian market in 2016. All comparative claims in Section 6 and Section 7 are thus propositions based on primary evidence on one side and synthesized secondary evidence on the other. The paper’s unique contribution is not in previously unpublished empirical findings, but in the platform/non-platform comparison and theorization, the institutional bypassing argument, and the design authority continuum. The title “Beyond Platform Work” should therefore be read as a shift away from the platform-centric focus of the AM literature to traditional workplaces, rather than a claim to supersede platform-work research or to treat both settings with equal empirical depth. Hungary provides a strategically relevant research setting: it has a large and growing logistics sector that is deeply embedded in European supply chains, its platform economy was shaped by a high-profile regulatory clash (the Uber debacle of 2016), and its industrial relations system is marked by extremely low union density and low coverage of collective bargaining (Berki, 2023; Meszmann & Szabó, 2023). It is thus an instructive, if not necessarily representative, case of AM implementation under weak collective worker representation—conditions present to varying degrees across a CEE region that nevertheless varies substantially in union density, bargaining structures and platform regulation.

We ask four questions: (RQ1) In what ways do the forms and functions of AM diverge across digital labor platforms and traditional firms? (RQ2) How do workers negotiate algorithmic control and resist in these contexts? (RQ3) What is the effect of institutional and regulatory settings on the implementation and outcomes of AM? (RQ4) What governance implications emerge from a comparison of platform and non-platform AM in the Hungarian institutional setting?

The paper is organized as follows. Section 2 is the theoretical framework. Section 3 deals with data and methodology. Section 4 and Section 5 review results of AM in platforms and traditional workplaces, respectively. Section 6 provides a comparative analysis across six dimensions. Implications are discussed in Section 7. Section 8 concludes with future research directions.

2. Theoretical Framework

2.1. Algorithmic Management: Core Concepts

Algorithmic management (AM) is the use of digital instruments underpinned by data collection, machine-learning algorithms and automated decision-support systems to perform or support tasks that were previously performed by human managers. Mateescu and Nguyen identify five core features of AM: the collection and surveillance of large datasets, real-time data responsiveness, automated or semi-automated decision making, the transfer of performance evaluations to rating systems and the use of “nudges” and penalties to incentivize worker behavior (Mateescu & Nguyen, 2019).

More recently, Business Europe (2023) offers a definition better suited to traditional workplaces, namely that AM concerns “the use of artificial intelligence to automate managerial tasks that are directly related to the coordination of labor input in the workplace”. Stark and Vanden Broeck detail four principles of AM: data-driven, continuous, scalar, and performative, meaning the act of measuring performance shapes the behavior it ostensibly records. These features make AM a management tool and disciplinary mechanism and a site of contested meaning between employers and workers (Stark & Vanden Broeck, 2024).

Baiocco et al. propose a four-level typology of automation relevant to AM: (1) no automation, (2) management assistance, (3) partial and conditional automation and (4) full automation. In our empirical cases, we find elements across all four levels, often within the same organization, suggesting that AM operates as a continuum rather than a dichotomy (Baiocco et al., 2022).

2.2. Platform Labor and Algorithmic Control

The literature on platform work and AM has grown substantially. Digital labor platforms (food delivery (Wolt), ride-hailing (Uber, Bolt), online project work (Upwork), etc.) are based on “algorithm-based management” as the predominant mode of worker coordination, instead of human supervisors and formal HR systems that are typical of employment (Jarrahi et al., 2021). The key instruments of platform AM are rating systems, surge pricing, geo-location tracking and automated deactivation (Makó et al., 2022; Pap et al., 2021).

A basic structural characteristic of platform AM is its opacity: workers are subject to algorithmic rules that they cannot inspect or challenge. Platform operators are therefore not depicted as neutral intermediaries but as real market-makers who “frame the entire institutional and regulatory framework of the platform economy” (Grabher & van Tuijl, 2020, p. 1012). This opacity creates a fundamental power imbalance that differentiates platform AM from AM in traditional workplaces (Alasoini et al., 2023; Vallas & Schor, 2020).

2.3. Algorithmic Management in Traditional Firms

AM is no longer only for platform companies. Recent EU-wide evidence shows that AM has moved well beyond platform companies: 24% of workers report that their working time, rosters or shifts are allocated automatically, 21% report automated task allocation, and 12% report automated benchmarking through leaderboards or dashboards (González Vázquez et al., 2025). Tools that automate the planning, staffing, coordinating and controlling functions across sectors now include enterprise resource planning (ERP) systems, project planning dashboards, warehouse management systems and KPI-tracking tools.

Most importantly, AM in traditional workplaces is framed within a very different institutional context: there is labor law, collective bargaining may be available, and workers are entitled to conventional employment rights. Krzywdzinski et al. argue that the degree of human managers’ mediation of algorithmic outputs, which they term “hybrid human–algorithm control”, is the key variable that differentiates progressive from regressive AM implementations (Krzywdzinski et al., 2024). As Dupuis demonstrated, AM can serve as a top-down control mechanism or as a participatory tool, depending on the existence of unions (Dupuis, 2024).

2.4. Worker Autonomy, Agency, and Resistance

A common finding across platform and traditional AM research is that workers are not passive recipients of algorithmic control. Drawing on the tradition of the labor process (Burawoy, 1979), scholars have identified a number of worker strategies including “multi-homing” (i.e., working for multiple platforms at the same time), selectively ignoring low ratings, gaming performance indicators, and using peer networks to share information about algorithmic rules (Alasoini et al., 2023). These practices are theorized by Alasoini et al. as “digital agency”, the capacity of workers to act intentionally within and against algorithmically structured work environments.

In traditional workplaces, Krzywdzinski et al. find similar patterns of “negotiated compliance” and “informal workarounds” through which workers accommodate AM while maintaining pockets of discretion (Krzywdzinski et al., 2024). Burawoy introduced the idea of “making out”, workers finding ways to meet targets while reducing effort intensity, which subsequent scholars have applied to digital contexts. This is a useful analytical lens for our cases (Burawoy, 1979).

2.5. Institutional and Regulatory Perspectives

Institutions do not just provide the stage upon which AM is implemented; they actively construct its forms and effects. The amount of space workers have to contest algorithmic governance is determined by national industrial relations systems, relative density of unions, scope of collective bargaining and strength of labor law (Potocka-Sionek & Aloisi, 2024; De Stefano & Taes, 2023). Hungary is an extreme example: union density has fallen to below 10%, collective bargaining coverage has declined from 47% in 2000 to 22% in 2020, and the bargaining system is decentralized and mainly restricted to single-employer agreements (Berki, 2023; Gyulavári & Kártyás, 2023).

At the European level, the EU AI Act (2024) and the EU Platform Work Directive (2024) are important regulatory responses to AM governance challenges, but their implementation in CEE settings is still lacking. These frameworks emphasize transparency, explainability and worker participation as core principles—principles against which our Hungarian cases can be meaningfully assessed.

2.6. Conceptual Model

Our analytical framework positions AM as a socio-technical system (STS) shaped by three interacting domains:

• Technology: algorithmic decision-making, data-driven performance monitoring, automated feedback loops.
• Organization: management structures, rule transparency, human–algorithm mediation, power relations.
• Institutions: industrial relations systems, worker representation, labor law, platform governance gaps.

The model predicts platform AM to be highly opaque, with real-time, market-based feedback and few institutional constraints. In traditional workplaces, AM will be integrated into existing hierarchical structures, resulting in hybrid control systems with varying degrees of managerial involvement and strength of industrial relations. Both settings generate worker agency and adaptation but the repertoires of agency vary across institutional contexts. This framework guides our comparative analysis in Section 4, Section 5 and Section 6. The three domains are not used illustratively but analytically: Section 5.5 reads each traditional-firm case through them in turn, and Section 6.4 draws on the resulting pattern to argue that the domains are not independent, but covary with the locus of AM design authority.

3. Methodology and Data

3.1. Research Design

This paper applies a comparative multi-case study design (Yin, 2014). Tomory indicates that the case study method is appropriate for “how or why” questions of contemporary phenomena in their real-world context, when the researcher has limited control over behavioral events (Tomory, 2014). Our approach is exploratory–explanatory, following Huws and Dahlmann (Huws & Dahlmann, 2007). We use the cases to explore new phenomena (AM in Hungarian workplaces) and to explain the mechanisms through which institutional context affects AM outcomes.

The selection of cases was purposive. For platform labor, we selected three geographically mobile platforms operating in Budapest (Uber, Wolt, Bolt), all representing the food delivery and ride-hailing segments of the Mobile Labor Market (MLM) (Codagnone et al., 2016). For traditional firms, we selected two contrasting cases: Data Analytics (DA), a high-involvement, knowledge-intensive business services company; and ConLog, a low-involvement contract logistics provider. This contrast allows us to examine AM at two contrasting points on the spectrum of skill levels, employment types and organizational logics present in the Hungarian economy.

3.2. Data Sources

Data were collected between January 2022 and July 2023 through four methods:

• Semi-structured interviews: fifteen interviews were conducted in total, seven in the DA strand (four with DA personnel and three with external specialists) and eight at ConLog (see

  Table 1. Summary of the interview details.

  Interviewee	Highest Educational Attainment	Age	Status (Self-Definition)	Work Arrangement
  ConLog 1	University	40–49	Standardization Lead	Full-time
  ConLog 2	Secondary	40–49	Site Manager	Full-time
  ConLog 3	University	40–49	Site Manager	Full-time
  ConLog 4	College	50–59	Facility Specialist	Full-time
  ConLog 5	Secondary	40–49	Supervisor	Full-time
  ConLog 6	Vocational	60–69	Operator	Full-time
  ConLog 7	Vocational	40–49	Operator	Full-time
  ConLog 8	Secondary	20–29	Operator	Full-time
  Data Analytics 1	University/Legal Studies	40–49	AM Legal Expert	Full-time
  Data Analytics 2	University/Sociology	40–49	Founder and Owner	Full-time
  Data Analytics 3	University/Social Sciences	50–59	Tool Developer	Full-time
  Data Analytics 4	University/IT Engineering	30–39	Operations Director	Full-time
  Data Analytics 5	University/Engineering Management	30–39	Account Representative/Senior Data Processing Specialist	Full-time
  Data Analytics 6	University/IT/Business Administration	40–49	AI Ethics Lead, Global Cloud Services Provider	Full-time
  Data Analytics 7	Technical High School	40–49	Works Council President, Large ICT-Services Employer	Full-time

  Note: identifiers denote the fieldwork strand, not the employer. Data Analytics 1, 6 and 7 are external specialists employed by neither case firm (see Section 3.2). ConLog 1 and ConLog 4 hold elected trade-union representative roles in addition to the operational position shown.

  ). The interviewee identifiers in Table 1 reflect these two strands of fieldwork, not employment at the firm indicated in the identifier. Interview guides were developed in the INCODING consortium and iterated through seven versions in four countries (Hungary, Germany, Denmark, Spain). The eight ConLog interviews comprise three warehouse operators, three managers (two site managers and one supervisor) and two employees who also hold elected trade-union representative positions at the site (ConLog 1 and ConLog 4). Tenure and gender are not reported at the individual level because the inclusion of either would render several of the interviewees identifiable in organizations of this size, and with role designations as specific as those appearing in Table 1. The DA strand comprises seven interviews, four with DA personnel—two rank-and-file employees (a tool developer and an account representative/senior data processing specialist) and two managers (the founder-owner, who also is CEO, and an operations director)—and three external specialists who do not work for either case firm: a legal expert in algorithmic management (Data Analytics 1), an AI ethics lead at a large cloud services provider (Data Analytics 6) and a works council president at a large ICT-services employer (Data Analytics 7). These three were added to this strand because DA has no union, works council or collective agreement (Section 5.1), and so legal, ethical and sectoral perspectives on interest representation could not be elicited within the firm. They do not underlie any claim about DA as a case, which is based on four internal accounts.
• Participant observation: a key researcher had several years of prior employment at the DA firm (2014–2017), providing rich insider knowledge of the evolution of AM practices at that site. This employment predated the 2022–2023 data-collection window by five to eight years. The observational material consists of contemporaneous field notes and was consequently considered as longitudinal background on the evolution of the PPS rather than evidence of practice during the study period (see the reflexivity discussion in Section 3.5 and Section 3.6). The results in Section 5 are not based solely on this material. Where it is used, it is to establish the developmental history of the PPS between 2014 and 2017, and every assertion about the operation of the system during the study period is backed up by interview or documentary evidence gathered in 2022–2023.
• Documentary analysis: company reports, platform terms and conditions, app screenshots, regulatory texts and media coverage of the Uber ban were systematically reviewed.
• Platform artifacts: screenshots of Bolt, Wolt and Upwork rating interfaces, surge pricing notifications and app-mediated worker communication were collected and analyzed.

For the platform cases, data were drawn from the parallel CrowdWork21 project (2019–2021) and previously published case studies (Pap et al., 2021; Makó et al., 2022), supplemented by document review and secondary analysis. Interviews were recorded with consent, transcribed and analyzed. The evidentiary status of the platform material bears directly on what can be supported by the comparison, and we state precisely what was and was not done with the material. We did not re-analyze the raw CrowdWork21 interview corpus: this material was produced in a separate project with different instruments and was not available to this study in transcript form. What we directly analyzed were platform artifacts and documents, specifically the terms and conditions in force for Bolt and Wolt couriers and drivers, in-app screenshots of rating displays, surge notifications and deactivation messages, platform communications to workers and the regulatory and media record surrounding Uber’s exit. These were coded against the same four analytical dimensions as the traditional-firm cases (Section 3.3), and this coding is the new analytical work here. The second source was the previously published case studies (Pap et al., 2021; Makó et al., 2022), where their findings were considered as evidence to cross-check against the artifacts, instead of conclusions to be re-stated. This leads to a secondary analysis where we apply our own coding frame.

3.3. Analytical Strategy

The analysis was carried out by means of an abductive approach (Alvesson & Kärreman, 2011) in an iterative process between empirical observations and theoretical concepts. Thematic coding was applied to each case along four analytical dimensions: (1) forms and functions of AM, (2) worker autonomy and constraints, (3) worker agency and resistance and (4) institutional embedding. Cross-case comparison was performed along six analytical dimensions (see Section 6), using the comparative logic of “most different cases” to identify the institutional factors explaining variation in AM outcomes. Findings were validated by triangulation across data sources. The six dimensions of comparison were derived from the four coding dimensions as follows. During analysis, the coding dimension (1), forms and functions of AM, was disaggregated into two comparison dimensions, as opacity emerged from the coded material as an analytically distinct property of AM architectures. Coding dimensions (2) and (3) directly map onto the dimensions of worker autonomy and worker agency comparison. Coding dimension (4) was divided into institutional embedding and governance challenges, thus distinguishing the descriptive findings on institutional context from their regulatory implications. In the two traditional-firm cases, the findings were validated by triangulation across independent types of data sources (interviews, participant observation, and documents). This form of triangulation was not possible for the platform cases where the evidence base is secondary throughout, and validation was limited to the cross-checking of the published case studies against documentary sources and platform artifacts.

The coding procedure was performed in three steps. The first stage, description, involved reading the interview transcripts, field notes and documentary material in full, and segmenting them into units of meaning, to which temporary descriptive codes were assigned (e.g., “real-time metric visibility”, “renegotiation with the client”, “bonus calculation”, “informal grievance route”). In the second, structuring stage, the provisional codes were condensed into a written codebook in which each code was defined, given an inclusion rule, an exclusion rule and an anchor quotation from the corpus; the codes were then grouped under the four analytical dimensions listed above. In the third, interpretive stage, the coded segments were re-read in the light of the constructs of labor process theory, socio-technical systems theory and institutional analysis, and both the codebook and the constructs revised where the material did not fit them; following the abductive logic of the design, a further round of coding was undertaken following each revision. The coding was performed using ATLAS.ti. The entire corpus, consisting of the fifteen interview transcripts, the field notes and the documentary material, was imported into one project file. The codebook, the code definitions with their inclusion and exclusion rules, the anchor quotations and the coded segments were kept there throughout. The software was used to manage the code system through the successive rounds of coding described above, to group codes under the four analytical dimensions, and to retrieve coded segments by code, by dimension and by case for the cross-case comparison reported in Section 6. It also supported the double-coding exercise described below, in that the two coders worked in separate copies of the project and the outputs were compared segment by segment on the basis of exports from them, and it supports the chain of evidence referred to in the validity checks, since every claim advanced in Section 5, Section 6 and Section 7 can be traced back to the coded segments and source documents from which it derives.

The codebook was developed by a team rather than a single analyst. Two researchers independently coded a sub-sample of the transcripts, designed to include both firms and each type of interviewee, and then compared the two outputs segment by segment. Agreement was calculated as the proportion of segments that were assigned to the same analytical dimension and to the same code within that dimension. Disagreements took two forms. Boundary disagreements, over the start and end points of a coded segment, were treated as non-substantive; substantive disagreements, over which code applied, were not. All substantive disagreements were resolved in joint discussion and the resulting decisions were written back into the codebook as sharpened inclusion and exclusion rules prior to coding the remaining transcripts. Because of the abductive nature of the design and the fact that the code set was not fixed but was allowed to evolve during the analysis, we did not treat coding as a fixed instrument measurement task. We therefore report agreement in a descriptive manner rather than as a chance-corrected coefficient; the operative criteria were the stability of the codebook across rounds and the resolution of every substantive disagreement by consensus.

Given the systematic difference in size, sector and involvement regime between the two firms, saturation was assessed within each case rather than across the pooled sample. Codes were developed interview by interview and a case was considered thematically saturated when two consecutive interviews did not generate new codes for any of the four analytical dimensions and only variations of existing codes were generated. Saturation was reached on this measure at both sites within the interviews conducted. We make the modest claim that the four interviews with DA staff and the eight at ConLog, read alongside the three external specialist interviews, are sufficient thematically for the constructs that structured the interview guide, but not for the full range of experience of AM in either organization. Themes which the material does not include may be contained in perspectives not represented in the corpus, especially those of workers who had already left the firms.

Four validity checks were performed. First, data triangulation: wherever possible, each analytic claim about the traditional-firm cases was cross-checked with at least two independent source types (interview, participant observation, document), and wherever this was not possible, claims corroborated in only one source type are identified as such. Second, investigator triangulation (through independent coding and joint interpretation sessions as outlined above). Third, negative-case analysis: material that ran counter to the emerging interpretation was not discarded but was retained and analyzed. The clearest example is the ConLog finding that an imposed, globally standardized AM system nonetheless increased perceived fairness in bonus calculation among the operators interviewed (Section 5.3), which contradicts a straightforward reading of the low-involvement thesis and is reported as a deviant case. Fourth, the chain of evidence was maintained, linking each claim in Section 5, Section 6 and Section 7 to the coded segments and source documents on which it was based. External validity is analytic rather than statistical: the cases support generalization to theoretical propositions rather than to a population of firms (Yin, 2014).

3.4. Case Profiles

Table 2. Summary of the five cases investigated in this study.

Case	Sector	Type	AM System	Employees/Interviewees
Uber Hungary	Ride-hailing	Platform (MLM)	Dynamic pricing, geo-tracking, rating deactivation	N/A (exited 2016); document analysis
Wolt Hungary	Food delivery	Platform (MLM)	Surge incentives, GPS routing, five-star ratings	Secondary data; published case study
Bolt Hungary	Ride-hailing/food delivery	Platform (MLM)	Rating system (1–5), GPS, automated incentives	Secondary data; published case study
Data Analytics (DA)	Business services/IT	Traditional (high involvement)	Project Planning System (PPS); KPI dashboards	65 employees; four internal interviews (two employees, two managers) + three external specialist interviews
ConLog	Logistics/warehousing	Traditional (low involvement)	Global warehouse management system; standardized KPIs	~200 employees; eight internal interviews (three operators, three managers, two elected trade-union representatives)

Source: Own design.

presents the five cases examined in this paper, chosen to highlight variations in employment arrangements and the extent of algorithmic control. Uber, Wolt and Bolt Hungary are platform-mediated labor markets where the work is coordinated through dynamic pricing, GPS tracking and customer ratings. Data Analytics (DA) and ConLog are conventional employers but they have a different degree of employee involvement in their systems. The table further gives the size of each organization and the empirical basis on which it was reconstructed.

Table 2 indicates that the cases do not represent a singular organizational form, but a wide range. The platforms delegate almost all of the coordination function to the system, while ConLog employs similar technologies in a conventional employment relationship where managers and union representatives are part of the decision chain, with DA in an intermediate position. Such variation enables one to separate the effects of the technology from those of the employment arrangement. The following sections address each case in turn.

3.5. Limitations and Evidentiary Basis of the Comparison

The five-case comparison presented in this paper combines two distinct types of evidence, and this asymmetry should be made explicit. The two traditional-firm cases (DA and ConLog) are based on newly conducted fieldwork: 15 semi-structured interviews, of which 12 were with personnel of the two firms and three with external specialists, together with participant observation and documentary analysis, carried out between January 2022 and July 2023. The three platform cases (Uber, Wolt, Bolt), by contrast, are analyzed primarily through platform artifacts, documentary review and secondary data drawn from the earlier CrowdWork21 project (2019–2021) and previously published case studies (Pap et al., 2021; Makó et al., 2022). The Uber case in particular should be read as a historical and regulatory benchmark rather than as evidence of AM as currently experienced by Hungarian platform workers, since Uber withdrew from the Hungarian market in 2016, it illustrates how platform AM interacts with regulatory institutions rather than documenting ongoing worker experience. We keep the Uber case, even if it is dated, for one specific and limited analytic reason. This is the only case of platform AM in Hungary that has faced a decisive institutional response, and thus is the only observation in our material where the institutional variable is not held constant. Bolt’s next move and consolidation can only be understood in relation to that benchmark. In the absence of Uber, it seems to be a property of Bolt rather than an institutional effect that the company is able to operate within the constraints of Hungarian regulation. Uber is thus used only as a regulatory point of reference, in Section 4.1 and Section 4.4, and no claim about contemporary AM worker experience anywhere in this paper depends on it. A symmetric design could include primary interviews with Wolt couriers and Bolt drivers and riders, using the tools used in the traditional firms. Such fieldwork is outside the scope and funding period of the INCODING project and we identify it as the immediate next step in this research agenda (Section 8). In the absence of such data, each platform-traditional contrast in this paper is offered as an exploratory hypothesis rather than as a comparison of like with like. A consequence of this asymmetry is also a question of time: the platform evidence comes from before the fieldwork in the traditional firm, by some three to six years, and was produced in a different project with different instruments. We cannot exclude that Hungarian platform AM has changed itself since 2021, and the comparison should be interpreted with this time gap in mind.

A second limitation concerns sample size and generalizability. The traditional workplace findings rest on 12 interviews inside two purposively selected firms, read alongside three external specialist interviews, the firms having been chosen to represent two contrasting positions on a high-involvement/low-involvement spectrum. However, these twelve interviews are not uniformly distributed, with eight at ConLog and just four at DA. A four-person internal sample, including two managers and the founder–owner who commissioned the system, cannot tell us how AM is experienced across DA’s 65 employees, and is weighted toward those with the greatest stake in the system’s participatory self-description; no former employees were interviewed. The DA findings are therefore offered as a qualified single-firm account. The term “participatory” in relation to the PPS (Section 3.6 and Section 5.1) is used in a strictly comparative rather than descriptive sense of employee involvement in design. The resulting typology (Section 7.1) should therefore be read as identifying plausible mechanisms and patterns that warrant testing in other organizations and sectors, rather than as an exhaustive map of AM practice in Hungarian traditional workplaces, nor is Hungary a representative CEE case. Union density and collective bargaining coverage are at the lower end of the regional distribution, sectoral bargaining is largely non-existent, and the platform economy was shaped by an unusually abrupt regulatory intervention in 2016. Hungary is thus better understood as a limiting case, useful for observing what AM does where institutional counterweights are weakest, than as a proxy for a region that varies substantially in union density, bargaining structure and platform regulation. Claims in this paper, which extend beyond Hungary, are therefore formulated as hypotheses for comparative testing rather than as regional findings.

Third, one researcher was working at DA during the period 2014–2017, prior to the data collection period. This relationship facilitated rapport and provided longitudinal insight into the development of the firm’s AM practices, but also raises a reflexivity consideration in that the researcher’s prior involvement may have biased the analysis toward the framing of the Project Planning System as a participatory innovation by the firm itself. The mitigation is based on the four internal DA interviews at all levels of the organization, on the documents, and on the procedures of coding, counter-reading and memo-writing described in Section 3.6, to which the reader is referred; the three external specialist interviews play no role in it, since they do not support any claim about DA as a case (Section 3.2). This previous relationship should be taken into account when evaluating the DA findings.

3.6. Reflexivity and Researcher Positionality

More than recognition is due for the earlier association of one team member with DA, as detailed above, since it bears on the case that holds most of the paper’s positive findings. The researcher was employed by DA during the development of the Project Planning System (PPS) from 2014 to 2017 and had no ownership interest in the company and no contractual or advisory relationship with it during the study period of 2022–2023. The relationship shaped the research in three identifiable ways: it informed the selection of DA as a case, it facilitated access and rapport, and it provided a longitudinal understanding of the PPS that no interview conducted in 2022 could have produced. The first two are benefits of the design. The third is also the main risk, because the developmental history of a system, told from within, tends to be told as a history of intentions, and the intentions at DA were participatory.

We adopted four procedures to keep risk visible and not just declared. First, the DA interviews were coded jointly with a researcher who had no prior relationship with the firm and no analytic claim about DA rests solely on the insider researcher’s coding. Second, the insider’s memories and contemporaneous notes from 2014 to 2017 were treated as background material on the development of the PPS and not admitted as evidence of practice during the study period (Section 3.2). Thirdly, we deliberately counter-read the DA material: one team member developed the strongest possible argument that the PPS was a tool of work intensification rather than participation and tested this against the coded segments. The more intense findings reported in Section 5.2 are a product of and survived that exercise and are presented as a substantive result rather than a concluding caveat. Fourth, interpretive decisions about DA were recorded in reflexive memos that pointed out where the insider’s prior knowledge had contributed to a reading, so that the other authors could revisit those readings.

These procedures do not neutralize the position, and so we explicitly qualify our use of the word “participatory”. The term is comparative, not absolute: DA employees exercise more discretion over the use of algorithmic outputs than ConLog’s warehouse operators or the platform workers discussed in Section 4, and it is that difference the term marks. That does not mean that DA employees were involved in the design of the system. They did not choose the metrics, and the decision to count client-billable time while undercounting internal training, innovation and collegial support (Section 5.2) is a managerial specification of what counts as performance, made before any employee interacts with the dashboard. Participation is also not institutionally secured: DA has no union, no works council and no collective agreement (Section 5.1), so employee influence on the PPS is a revocable feature of the founders’ management philosophy rather than a right that would survive a change of ownership or strategy. Finally, the same design feature that interviewees experienced as transparency, the continuous visibility of individual performance data to managers and colleagues, is equally a mechanism of peer-visible surveillance, and interviewees recruited with the firm’s cooperation are more likely to articulate the first reading than the second. We have thus treated the participatory nature of DA’s AM as a claim under test throughout the analysis rather than as a case property. Section 5.1 and Section 5.2 are to be read as a single argument, the latter half qualifying the former.

4. Findings: Algorithmic Management in Platform Labor (Secondary Evidence Base, 2019–2021)

4.1. Forms of Control: Real-Time, Opaque, Market-Mediated

All three Hungarian platform cases employ AM architectures that share a common structural logic: real-time data collection from mobile devices, machine-learning-driven task allocation and pricing, customer-generated ratings as the primary performance metric, and automated deactivation as the ultimate disciplinary tool. Unless otherwise indicated, the findings in this section rest on secondary evidence and platform artifacts from 2019 to 2021 (see Section 3.5), and present-tense formulations refer to platform AM as documented for that period, not to current ethnographic observation.

Bolt and Wolt use a five-star rating system where customers review every finished service. If the average rating of a driver drops below a certain threshold after 40–50 completed trips (Pap et al., 2021), the platform automatically terminates the relationship without any possibility of appeal. Importantly, the rating criteria are not revealed to the driver and the platform’s judgment confounds platform-level failures (e.g., the app locating the passenger incorrectly) with driver performance, generating systematic measurement error that workers cannot challenge (Pap et al., 2021). Wolt has a similar mechanism: ratings below the platform’s internal threshold trigger contact from customer service and, if they persist, termination of the contract, for couriers and restaurant partners alike. The point at issue is that the reviewed platform documentation does not publish that threshold itself.

A second form of algorithmic control is surge pricing. When demand exceeds supply, Bolt and Wolt activate surge multipliers and communicate them via SMS, in-app notifications and social media. Workers are “nudged” rather than told, but the financial incentive is structured to be hard for workers reliant on income from the platform to refuse. This is an example of what Mateescu and Nguyen call the use of “nudges and penalties to indirectly incentivize worker behavior”, control through market signals rather than managerial authority (Mateescu & Nguyen, 2019).

The Uber case, by contrast, concerns the consequences of AM opacity when it is combined with regulatory non-compliance, and is presented here as a historical and regulatory benchmark (see Section 3.5). Uber entered Budapest’s market in November 2014 as a neutral technology platform and not as a transportation company, and therefore claimed an exemption from the whole range of taxi regulations (Makó et al., 2022). Its algorithmic management system, dynamic pricing, driver geo-tracking and automated rating-based deactivation, was designed for a regulatory environment in which the platform itself set the rules. After the Hungarian taxi industry staged a major protest in January 2016 and parliament passed legislation that effectively banned the Uber model, Uber chose to leave Hungary in July 2016 rather than comply. Bolt (then Taxify) filled the vacuum but did so by accepting regulatory constraints—a key lesson in how institutional context shapes platform AM governance.

4.2. Worker Autonomy and Constraints

While platform workers in Hungary have high temporal autonomy (e.g., they can choose when to work), their procedural autonomy is highly limited: the algorithm determines which tasks to take, how to perform them, how to price them, and what the consequences of underperformance will be. Employment status adds to this asymmetry: in Hungary all platform workers are legally self-employed and therefore do not benefit from standard employment protection such as minimum wage, sick leave and collective bargaining rights (Makó et al., 2022).

Gamification elements, streak bonuses, achievement badges, leaderboard comparisons, extend algorithmic control into workers’ motivational and emotional lives, intensifying performance pressure without triggering the formal performance management processes that apply in standard employment. In the 2019–2021 CrowdWork21 material, workers reported that the combination of rating anxiety, surge pressure and gamified incentives produced a form of “self-intensification”, in which the algorithm effectively outsourced managerial discipline to the workers themselves (Pap et al., 2021; Makó et al., 2022).

4.3. Worker Agency: Multi-Homing, Workarounds, and Community

Platform AM has a large structural power asymmetry, yet Hungarian platform workers develop different strategies of agency. The most common is “multi-homing” or signing up to multiple platforms (Bolt, Wolt, Uber Eats) at the same time to become less dependent on any one algorithmic system and to benefit from the surge-pricing differences across platforms.

Workers also communicate outside the app through informal Facebook groups and WhatsApp channels to share information about surge locations, informally rate customers, and coordinate informal collective responses to changes in platform rules (Alasoini et al., 2023). The specific Hungarian instantiation of these practices is inferred from platform artifacts and the secondary case material rather than directly documented, and is advanced accordingly.

Some workers engage in what Burawoy (1979) calls “making out”, finding the least effortful way to meet platform metrics without maximizing actual performance (Burawoy, 1979). For Bolt drivers, this could be in the form of accepting only high-rated passengers; for Wolt couriers, it is about spatial positioning in close proximity to zones with high demand, learned through collective intelligence and not algorithmic guidance. While these strategies do not constitute formal collective resistance and Hungarian trade unions have little interest in organizing platform workers, they are a form of distributed, digitally-mediated agency that partially counterbalances algorithmic control.

4.4. Institutional Misalignment

The Hungarian institutional context provides particular conditions for platform AM governance. In the absence of sectoral collective bargaining and with weak trade unions, platforms face little organized worker resistance. The main regulatory framework, a binary classification into “employee” (covered by the Labor Code) and “self-employed” (covered only by civil law), leaves platform workers in a legal limbo that platforms actively exploit to avoid employment obligations (Makó et al., 2022).

The Uber case illustrates that this institutional void can be exploited to the point of regulatory backlash: when the platform’s AM architecture conflicted sufficiently with incumbent economic interests (the taxi industry) and tax revenue expectations (the government), the institutional response was swift and decisive. Bolt’s subsequent success demonstrates that platform AM can operate sustainably in Hungary, but only when the platform accepts the basic institutional constraints of the national context, a lesson with broader implications for the governance of AM in supply chains and logistics.

5. Findings: Algorithmic Management in Traditional Workplaces (Primary Evidence Base, 2022–2023)

5.1. Data Analytics (DA): Qualified Participatory AM in a High-Involvement Setting

Data Analytics (DA) is a Hungarian private company established in 2007, growing to 65 highly skilled employees by 2023. The firm provides market research, data processing, analytics and IT development services. There are no formal institutions of collective interest representation at DA. The DA case relies on the smaller of the two internal samples, four interviews with firm personnel, two rank-and-file employees and two managers, one of whom is the founder–owner who commissioned the system. Therefore, two of the four internal accounts are managerial, and the three other interviews linked to this strand are with external specialists not employed by either case firm (Section 3.2). The results reported below are therefore provisional and particular to a firm, not an account of how AM is experienced across DA’s workforce as a whole.

The primary AM tool at DA is the in-house developed Project Planning System (PPS), which encompasses project delivery, financial performance and HR functions. The PPS tracks individual and team hours, project budgets, client interactions and productivity measures in real-time, so performance data are constantly visible to employees and managers alike. Color-coded dashboards alert staff when projects are over budget or behind schedule, prompting direct re-negotiation with clients—a task once restricted to management.

The PPS was introduced due to rapid organizational growth. As the company expanded from 8 to 10 to 65 employees, centralized managerial control was no longer feasible. The DA’s founders’ philosophy was a deliberate delegation of decision-making power to teams and individual employees by building project management authority into the AM system. This logic was well explained by the CEO: “This system, to have a clear view of the business, only works if the employees are part of its operation”.

The central observation at DA, advanced as a provisional case-specific finding rather than a general claim, is that AM can operate in a comparatively participatory manner where its design reflects a high-involvement management philosophy. The term “participatory” is used here in a qualified, comparative sense as set out in Section 3.6: it refers to greater employee discretion over the use of algorithmic outputs than in the other cases examined, not employee involvement in designing the system or in selecting its metrics, and it describes a feature of the firm’s management philosophy rather than an institutionally secured right. Employees gained meaningful new roles: they now negotiate directly with clients on project terms, timelines, and pricing, tasks that require commercial judgment and interpersonal skill. All four of the internal DA interviewees described AM transparency, the visible alignment between effort, output, and reward, as making performance evaluations fairer and less ambiguous than the pre-PPS arrangements. This convergence across four accounts, two of which are managerial and one of those the founder-owner who commissioned the system, is reported as a pattern within a very small sample rather than as a firm-level finding, and it is not evidence of how the PPS is experienced across DA’s 65 employees.

“The truth is, if only the team leader is business-oriented, or if it’s just me (the founder), it certainly won’t work with 60 people, as they have been delegated the role to negotiate the price of our services on a project level. So basically, everyone has to become one. There is already an aspect of PPS, how to transform the entire company into a business-minded community.”

(DA 2)

5.2. AM and Job Quality at DA: Autonomy and Intensification

The PPS’s positive effects on autonomy and transparency are accompanied by a significant negative consequence: work intensification. Because the system tracks only time booked to client projects, complementary activities, internal training, innovation and supporting colleagues are systematically under-counted. Two of the four internal interviewees described a resulting pressure to exceed contracted hours in order to demonstrate adequate productivity.

This is consistent with Krzywdzinski et al., who argue that algorithmic management (AM) may improve workers’ autonomy in some dimensions (task organization, client relations) but undermine it in others (control over effort intensity, work–life balance) (Krzywdzinski et al., 2024). The DA case shows that participatory AM design does not solve the inherent tension between datafication and human complexity but rather relocates and partially conceals this tension. The qualification should be drawn out as an implication for the framing of the case. Intensification at DA is not an incidental cost attached to an otherwise participatory system, it is the same design decision that yields the participation. The metric that gives employees the standing to renegotiate a project with a client is the metric of client-billable time, and it is precisely because that metric was made the operative measure of contribution that internal training, innovation and collegial support fall outside it. On the accounts given in the four internal interviews, DA employees were free to respond to the numbers while losing the ability to have work that the numbers do not capture recognized as work. Read with Section 3.6, this means that the participatory character of DA’s AM is a finding about the distribution of discretion within a given definition of performance, not about the definition of performance itself, which remained a managerial prerogative and which no interviewee described as negotiable.

“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%, then 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.”

(DA 4)

“After a while, it turned out that in the reports, the company was only occupied 60% of the time with chargeable hours, but in fact, they were working overtime to deliver the assigned projects.”

(DA 4)

“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.”

(DA 3)

5.3. ConLog: Standardized AM in a Low-Involvement Setting

ConLog is a Hungarian subsidiary of a multinational logistics company with about 200 warehouse workers. The AM system was implemented by the parent company as part of a global standardization effort, with only minimal local adaptation and without the participation of Hungarian workers or their representatives in the design process. The system tracks individual and team productivity metrics in real time and feeds these data into an automated performance management system.

Unlike DA, ConLog’s AM was not designed to empower workers; it was designed to enforce global corporate standards. Site managers interpret algorithmic outputs and translate them into individual performance assessments, creating a hybrid human–algorithm control structure in which algorithms set the standards and humans apply them. The five non-managerial interviewees at the site, three warehouse operators and two employees who also hold elected trade-union representative positions, reported that the system increased the transparency and predictability of performance evaluation, in that they know what is expected of them and how their performance is measured, while also reducing the discretion they had previously exercised in organizing their own work routines. This is a convergence across five of the eight ConLog interviews and not offered as an account of how AM is experienced across the site’s roughly 200 warehouse employees.

Notably, even in this low-involvement setting, the AM system produced some unexpected positive outcomes. The three warehouse operators interviewed reported that real-time performance feedback made it easier to understand the link between effort and earnings, reducing the opacity that had previously characterized performance-related pay. As one of them put it, the system made the bonus calculation transparent in a way that informal managerial discretion never had been.

“The special site of the company is motivated to use corporate standard solution … The system is an Oracle Software that is customized for the Company. The actual site is not developing, only using it.”

(ConLog 2)

“The standardization drive originated in acquisitions and in the tender risk attached to a major global client, and subsequently became part of corporate strategy—that is, it was imposed globally rather than co-designed locally.”

(ConLog 1)

“I can see my real-time performance, the percentages. I feel that its 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)

“Payroll is using the system … directly affecting my premium that is linked to performance. I think that the system shows an objective picture of my work.”

(ConLog 8)

“It helps me to work without mistakes, the system does not allow incorrect picking, it stops me immediately.”

(ConLog 6)

5.4. Institutional Moderators at ConLog

ConLog has a trade union, a major exception in the otherwise poorly organized Hungarian logistics sector, with around 30 years of sectoral experience. Among the eight interviewees from ConLog, two (ConLog 1, ConLog 4) have chosen to hold employee-representative roles in addition to their operational functions (Section 3.2). The quoted passages in this section are from operators. The union’s presence appears to have had an indirect effect on the implementation of AM, acting as a channel of communication for individual complaints and keeping informal contact with management over working conditions, even if the union has not formally negotiated the terms of AM implementation. This finding aligns with Dupuis’ assertion that union presence moderates AM outcomes in the absence of formal bargaining rights (Dupuis, 2024).

“We have a workers union. We communicate with them frequently face to face. They are involved not only in the employment related things, but also in the work related things, e.g., the actual processes in the warehouse … However, the system related issues/things are communicated directly to the WH Management and the System Deployment teams.”

(ConLog 6)

“They shall keep us informed, and shall represent our interest towards the management.”

(ConLog 6)

“I’m not aware of any employee representatives here. I’m just a simple worker … If I have any trouble I talk to the shift leader or the HR Assistant.”

(ConLog 8)

The strength of the moderation claim should be as strong as the evidence really supports, which is less than the preceding paragraph implies. The material establishes that there is a union at the site, that it has regular face-to-face contact with management and that it deals with individual grievances. What it does not establish is that this contact changed the AM system or its application. None of the interviewees reported a metric, target, monitoring practice or sanction that was changed as a consequence of union intervention. The union was not consulted in the design phase. There is no collective agreement clause and no written protocol on the system in the documentary material. System-related issues were, according to the interviewees themselves, routed directly to warehouse management and the system deployment teams, rather than via representation. The third quote above is another challenge to the claim. An operator at the same site said he had no idea that employee representatives even existed. This implies that despite whatever moderating power the union has, it is not evenly distributed across the workforce, and does not extend to at least some of the workers most directly exposed to the system. We therefore report the role of the union as a possible but not confirmed moderator. We consider the argument of Dupuis (2024) as compatible but not corroborated by the ConLog case. The corpus provides no evidence that any AM parameter changed as a result of representation, and such evidence would be required to demonstrate moderation. The contrast between DA and ConLog suggests, on the evidence of two firms, that the impact of AM in traditional workplaces is mediated both by the design philosophy of the AM system (participatory in the qualified, comparative sense set out in Section 3.6, as against imposed) and by the industrial relations context (high-involvement with informal participation, as against low-involvement with formal but limited union presence). Neither half of this proposition is established by the present material, and the first rests on four internal interviews at DA.

5.5. Reading the Two Cases Through the Conceptual Model

The three domains discussed in Section 2.6 can now be applied to the two traditional-firm cases. The value of the model is that the domains do not line up the same way in each. The two systems are more similar to each other in the technology domain than the cases as a whole would suggest. Both the PPS and ConLog’s warehouse management system continuously measure at the individual level, make performance visible in real time and feed the resulting data into pay-relevant calculations. Neither automates a dismissal decision and in the terms of Baiocco et al. (2022) both are at the level of management assistance rather than partial or full automation. If the AM outcomes were only driven by the technology domain, the two cases should be similar. They are not.

The organization domain produces the divergence, and the institutions domain accounts for why that divergence was possible. The crucial difference in the organizational form is the point at which the specification of the metric was made and by whom. At DA the founders designed the PPS in-house and coupled it to a delegation of client-facing decisions to project teams, so that the algorithmic output enters the labor process as an input to an employee’s own commercial judgment. At ConLog the metric is provided by the parent company, and the output of the algorithm is used as an input to the site manager’s assessment of the operator. Institutionally, neither firm’s employees had an enforceable claim on that design decision. DA has no union, works council or collective agreement (Section 5.1), and ConLog’s union, while present, was not party to the design of the system, and only engaged with it after implementation (Section 5.4). The comparison thus reveals the working of the institutional domain in a certain manner. It did not affect the design of either system. It determines whether the organizational choice that did affect the design is revocable at the employer’s discretion. At DA the participatory features of the PPS are only as lasting as the management philosophy of the founders; at ConLog the imposed features are equally beyond challenge. This provides the empirical basis for the argument in Section 6.4 and Section 6.5.

6. Comparative Analysis: Platforms vs. Traditional Firms

The evidentiary posture set forth in Section 3.5 applies to this entire section and to Section 7. The two columns of

Table 3. Summary of the comparative findings across six analytical dimensions. Given the asymmetric evidence base set out immediately above, entries should be read as propositions rather than as measured contrasts.

Dimension	Platform AM (Uber/Wolt/Bolt)	Traditional AM (DA/ConLog)
Mode of Control	Real-time, rigid, market driven; the algorithm is the primary manager	Hybrid human–algorithm; managers interpret algorithmic outputs
Transparency and Opacity	Highly opaque; criteria undisclosed; no grievance channel	Outcome transparency rather than design transparency. DA: high on both, system designed in-house. ConLog: dashboard visibility of scores coexists with no access to how the global parent specified them (see Section 6.2)
Worker Autonomy	High temporal, low procedural; no employment rights	Variable: DA increases both; ConLog reduces task discretion but clarifies expectations
Worker Agency	Multi-homing, out-of-app coordination, metric gaming	Negotiated compliance, informal workarounds, occasional formal grievance (ConLog union)
Institutional Embedding	Operates above/around institutions; exploits regulatory gaps	Constrained by the Labor Code and employment contracts; union present at ConLog but not demonstrably moderating AM design (see Section 5.4)
Governance Challenges	Misclassification; no collective rights; EU Platform Work Directive not yet effective	Limited worker participation in AM design; transparency needs formal channels; EU AI Act relevant

Source: Own design.

do not have equal evidentiary weight: the traditional-firm column summarizes primary fieldwork conducted in 2022–2023, while the platform column synthesizes secondary evidence and platform artifacts largely dating from 2019–2021, with Uber as a pre-2016 historical benchmark.

6.1. Mode of Control

Platform AM is characterized by what Stark and Pais call “cybernetic control loops”. The algorithm gathers data, measures performance and enforces consequences (deactivation, withdrawal of surge access) without human involvement (Stark & Pais, 2020). In a traditional office, the output of an algorithm becomes the input for a human manager who interprets the output and acts on it. Thus the main structural difference lies in human mediation, which generates accountability, institutional constraints and space for negotiation.

6.2. Transparency and Opacity

The key differentiator between platform and traditional AM is opacity. Hungarian platform workers do not have access to the criteria that determine whether they will be kept on or not, they cannot challenge deactivation and they have no way of raising concerns about algorithmic errors. By contrast, DA’s PPS is intended to be visible to all employees. ConLog’s system is not participatory by design, but it is at least predictable. That predictability should not be called partial transparency without qualification, because one word is measuring two different things. The first is outcome transparency: can a worker see the score, know which behaviors move it and know what the consequence is. This is where ConLog performs well, as the operators quoted in Section 5.3 confirm. Second is transparency of design: can a worker figure out how the metric was defined, why that threshold and not another, on what evidence the goal was set, and to whom the specification can be communicated. ConLog offers nothing on this second dimension, since the specification was made by the global parent and neither the workers nor local management took part in it. A dashboard gives the first and can leave the second as is, and it is the second that determines if a metric can be contested at all. In our material, the two only coincide in the case of DA, and they do so because the people that the employees can talk to designed the system in the same building. We have therefore replaced the single label in Table 3 with this distinction. It also qualifies the policy inference that follows: an obligation to explain how a system works to those subject to it concerns outcome transparency and does not, by itself, give them standing at the point where the system is designed. This distinction has direct policy implications. If properly implemented, the transparency provisions of the EU AI Act will have a major impact on platform AM but more limited effects on the relatively more transparent traditional workplace applications.

6.3. Worker Autonomy and Agency

These findings question the assumption that AM systematically undermines worker autonomy. At DA, AM delegated commercial decision-making authority to the project teams to give employees more autonomy. ConLog decreased task discretion but increased pay transparency. Platform AM preserved workers’ temporal flexibility, but at the cost of their procedural autonomy and their employment rights more generally. The relevant dimension of autonomy is therefore context-dependent: the same AM tool can increase autonomy in some domains and decrease it in others, depending on organizational design choices and institutional constraints.

6.4. Institutional Embedding

The weak industrial relations system in Hungary is not a passive background but an active shaping force. The lack of sectoral bargaining means there is no organized resistance to the implementation of AM by platforms, and traditional firms like ConLog can implement AM without meaningful worker involvement. By contrast, Thelen’s account of sustained union opposition to Uber in Germany underlines the importance of institutional capacity for AM governance (Thelen, 2018). Such institutional variation may present governance challenges and regulatory arbitrage opportunities for supply chain and logistics operators across European jurisdictions, although this is an inference from a single national case: our material offers no evidence from any jurisdiction other than Hungary, and we did not examine operators’ cross-border decisions. Hungary is situated at the weak end of the spectrum of collective representation in CEE, so this dynamic should be read as a mechanism identified in a limiting case, not as a characterization of the whole region. The six dimensions in Table 3 are not independent: opacity, mode of control and institutional embedding cluster together. Where design authority is readily available and institutionally visible, as at DA, transparency is greater and control more participatory. Where design authority is remote, whether structurally (platforms designed abroad) or organizationally (mandate issued by ConLog’s parent company), opacity is increased and worker influence reduced. The platform/traditional split is therefore a useful analytical starting point, but ultimately a proxy for a deeper variable: the institutional accessibility of the point where AM systems are designed.

6.5. Institutional Bypassing Through Global Value Chains

The comparative analysis reveals a mechanism that neither the platform-centric nor the traditional-workplace AM literature has adequately theorized, and which we term institutional bypassing through global value chains. In the platform cases, the bypassing of national institutions is direct and well documented: platforms exploit the binary employee/self-employed classification to place workers outside the protective scope of the Labor Code (Makó et al., 2022). The ConLog case, however, shows a second, subtler bypass route operating inside standard employment. ConLog’s AM system was specified by its multinational parent as part of a global standardization initiative, designed without reference to Hungarian industrial relations institutions and implemented with no involvement of local workers or their representatives. The material itself makes it clear that there was no local involvement: no interviewee at the site, managerial or otherwise, stated they were involved in, or consulted on, the specification of the system, and no local document records any involvement. We have not interviewed anyone at the parent firm and we have no documentation of the design process itself, so the location of that process, and the constraints that operated on it, are inferred from local evidence rather than observed. Hence, the system is formally regulated by Hungarian labor law, while the decision-making process that formed its architecture was not exposed to Hungarian institutional influence. This inference is the weakest link in the argument we develop below, and would require access at the parent company to test it, which we identify in Section 8 as a necessary next step.

Theoretically, to locate this mechanism is to engage the literature on global value chains and global production networks, which has long analyzed governance exercised across firm and national boundaries, but has not, to our knowledge, been applied to algorithmic management. Gereffi et al. (2005) develop a typology of governance of value chains based on transaction complexity, the codifiability of information and the capabilities of the supply base, the latter of which is directly involved here: AM standardizes the labor process, turning task execution into standardized, machine-readable parameters. Codification is what renders a labor process transferable across sites and, in the opposite direction, what renders the specification of that process portable. The parameters travel with the software, so the parent company can set work parameters at a subsidiary as a system configuration rather than a management instruction and without local negotiation. In this respect, ConLog is a captive node. The parent prescribes the process and keeps control of it, while the local unit carries it out. What the value chain typology leaves implicit, and that Coe and Yeung (2015) provide, is that such arrangements are embedded in territorially specific institutional environments whose boundaries do not coincide with those of the network.

This permits the mechanism to be stated explicitly. The institutional constraints that would bind the design of the labor process and the ones that bind the employment relationship become decoupled in cases where the AM parameters are set at a network node that is outside the territory in which the work is performed. The employment relationship at ConLog is governed by Hungarian labor law, Hungarian collective bargaining and Hungarian enforcement capacity, but none of these apply to the specification decision, because it is not an employment act performed in Hungary but a configuration decision made elsewhere and delivered as software. The geography of the value chain thus changes institutional constraint in a particular way: it does not relax the constraints, it relocates the object to which they would have to attach. This is what sets institutional bypassing through global value chains apart from the more familiar regulatory arbitrage where production is relocated to a weaker regime. Here, production stays, workers stay and employment rights stay, only the design authority moves, which is why the mechanism is invisible to instruments that regulate employment status and why it would stay even if Hungarian bargaining coverage were to increase substantially.

The two routes differ in mechanism but converge in outcome. Platform bypassing works through employment classification: the worker is placed outside the institutions. Value-chain bypassing works through corporate geography: the decision is placed outside the institutions, while the worker remains formally inside them. This distinction matters for governance. Regulatory instruments targeting the first route do not reach the second, because ConLog’s workers are already employees. Conversely, instruments targeting workplace AM transparency and consultation, such as the high-risk provisions of the EU AI Act, apply in principle to the second route but presuppose national enforcement capacity and workplace representation structures that, in Hungary’s case, are weak precisely where multinational logistics chains are concentrated.

This mechanism also qualifies the common assumption that traditional-workplace AM is institutionally embedded while platform AM is institutionally disembedded (cf. Table 3). Our evidence suggests a continuum: the degree of institutional embedding of AM depends not only on the employment relationship but on where in the value chain the system’s design authority sits. Where design authority is local (DA, whose PPS was developed in-house by the firm’s founders), weak national institutions may be partly offset by organizational-level discretion, although at DA that discretion was a revocable managerial choice rather than a secured form of participation (Section 3.6). Where design authority sits abroad (ConLog), formal institutional coverage coexists with substantive institutional bypass. We advance this as a proposition for testing in other CEE logistics and manufacturing subsidiaries: the more distant the locus of AM design authority from the workplace, the weaker the effective (as opposed to formal) institutional constraint on AM implementation.

7. Discussion

7.1. AM Is Not Monolithic

The basic conclusion of this comparative study is that AM is not a uniform system of control. It takes different forms, has different functions and different consequences in different organizational contexts, institutional settings and levels of worker skill. This finding resonates with recent theoretical arguments advanced by Krzywdzinski et al. and Dupuis, but adds empirical depth from a CEE perspective that the European literature has largely neglected (Krzywdzinski et al., 2024; Dupuis, 2024).

The Hungarian evidence suggests a typology of AM outcomes: (1) opaque, market-mediated control (platform AM, low institutional constraints), (2) hybrid human–algorithm control with limited participation (ConLog, weak IR); (3) participatory, transparency-enhancing AM (DA, high-involvement management despite weak IR). This typology has direct implications for AM governance: effective regulation must address not only the extreme platform case but also the hybrid traditional workplace case where AM is spreading fastest. The three types are built on unequal evidence. Type (1) is a synthesis of secondary material from 2019–2021. Type (2) is based on eight interviews at a single logistics site. Type (3) is based on four internal interviews at a single 65-person company and is the weakest of the three, hence the participatory label attached to it is qualified throughout (Section 3.6 and Section 5.1). The typology is thus posited as a collection of propositions, identifying candidate mechanisms, rather than as findings symmetrically evidenced. To test it we would need to add cases that could disconfirm it: traditional firms with strong representation at the workplace, high-involvement firms whose AM systems were externally specified and low-involvement firms with locally designed systems, the latter being completely absent from our material.

7.2. Convergence and Divergence in AM

Across all five cases, AM shares a common structural feature: datafication. Whether through GPS tracking, star ratings, PPS dashboards, or warehouse KPI systems, the fundamental logic of AM is to convert human labor performance into numerical data that can be monitored, compared, and acted upon in real or near-real time. In this sense, platform and traditional AM are converging on a shared technical infrastructure.

However, the institutional and organizational mediation of this datafication diverges sharply. In platforms, as documented in the 2019–2021 secondary evidence, datafication is fully automated and legally uncontested, in traditional firms, it is mediated by human managers, constrained by employment law, and in favorable cases shaped by worker participation. The convergence of technical infrastructure coexists with a fundamental divergence in power relations and governance. The EU’s regulatory response must grapple with both dimensions simultaneously.

7.3. Implications for Supply Chain and Logistics Governance

The logistics sector represented by ConLog is in a strategic position for the analysis of the governance of AM in this study. What follows should be read as conditional implications, not tested policy recommendations. Our data capture the specification and application of one AM system in one Hungarian subsidiary. We do not measure the effectiveness of regulation, nor enforcement capacity, nor the outcome of any governance intervention. Hungarian logistics companies are nodes in the European and global supply chains, and their AM systems are often specified by multinational parent companies or major customers. This means there is a form of “imported AM governance” where weak local institutions (i.e., limited unionism and low collective-bargaining coverage) are bypassed by global corporate standards. The same corporate standardization might face more resistance where bargaining coverage and enforcement capacity are higher than in Hungary. Thus, the argument is a hypothesis on AM governance through value chains, not a finding on CEE logistics in general.

The cases reveal three governance constraints, which are not derived from the cases. First is transparency. ConLog achieved outcome visibility without having access to the design of the metric. Therefore, a duty just to explain how a system works would leave that situation largely unchanged. What is missing is a channel to the point of specification. The second is involvement. There was no participatory design of ConLog’s system, the union’s role was limited to grievance handling after implementation, and the only participatory case in our material is characterized as such by a founder’s management philosophy, not by an institution. Our only claim is that a channel of communication already in existence is a cheaper starting point for consultation than one that has to be built from the ground up. Third is regulatory harmonization. The EU AI Act classifies many AM systems used in logistics as “high-risk” applications requiring conformity assessment, but the capacity of CEE countries to enforce such requirements nationally is still limited. In our view the binding constraint is enforcement capacity not the rules themselves and no recommendation to firms can substitute it.

7.4. Theoretical Contributions

This paper makes three theoretical contributions. First, it extends labor process theory to the comparative analysis of AM across platform and conventional workplaces, demonstrating that the LPT concepts of control, resistance, and agency remain analytically powerful but require adaptation to the specificities of algorithmic management. Second, it puts pressure on the socio-technical systems framework from an unexpected direction. In our cases the institutional layer mattered through its absence: no institution had standing at the point where either system was specified (Section 7.6, RQ3). What the material shows is therefore not that institutions determine which AM designs are possible, but that where no mandatory consultation point exists, design authority settles wherever the firm’s own structure places it. We propose accordingly that the institutional layer be modeled by the presence or absence of an enforceable claim on system design, rather than as a background condition. Third, it offers institutional theory a hypothesis rather than a documented finding: that weak national institutions may be partially offset by organizational-level practices, as the DA case suggests on four internal interviews, while leaving workers exposed where no such offset exists, as the platform cases suggest on secondary evidence.

7.5. Relationship to Prior and Companion Studies

The DA and ConLog cases draw on the same INCODING fieldwork that underlies Makó et al., which examines algorithmic management in these two firms as a within-traditional-workplace comparison of high- and low-involvement working practices (Makó et al., 2025). The present paper builds on that empirical material but pursues a different analytical purpose: rather than comparing high- and low-involvement traditional workplaces against each other, it places both alongside the Hungarian platform-labor cases (Uber, Wolt, Bolt) within a single comparative framework spanning platform and non-platform AM. Hence, the contribution of this paper is specifically in the platform/non-platform comparison itself: in the typology of AM outcomes developed in Section 7.1, in the analysis of opacity and institutional embedding across the platform–traditional divide in Section 6, and in the resulting implications for AM governance in supply chains and logistics in Section 7.3, rather than in the DA and ConLog findings considered in isolation, which are treated in greater depth in the companion study.

7.6. What the Data Answer: Summary by Research Question

Because the four research questions are addressed at different points in Section 4, Section 5, Section 6 and Section 7, and because they are supported to different degrees, we set out below what the interview, observational and documentary material specifically shows for each, separated from what the theoretical framework would lead one to expect.

• RQ1, forms and functions of AM. The data show, at DA, a system that measures client billable time and makes the resulting figures visible to employees and managers alike, with color coded alerts triggering renegotiation with the client by the employee rather than a manager; and at ConLog, a warehouse management system that measures individual and team output in real time, feeding a performance-related pay calculation, and interpreted by site managers before it reaches the operator. In the platform material, allocation, pricing and deactivation are handled automatically by the system. Accordingly, our data support a functional difference in who acts on the output not what is measured: measurement is broadly similar in all five cases; the location of the decision is not.
• RQ2, negotiating and resisting algorithmic control. Interviewees in DA described working beyond contracted hours to keep the visible percentage acceptable, and managers described recognizing that 100 per cent is a sign of overtime rather than good performance, which is an accommodation to the metric on both sides rather than a contestation of it. ConLog interviewees said they changed their pace on the screen figure through the shift and raised grievances to the shift leader, HR assistant or union, while one operator reported not knowing there were representatives. Interviewees from neither firm described collective action, a refusal, or a change to the system through negotiation. It is not our interviews, but the secondary sources that provide the platform repertoires of multihoming, out-of-app coordination and positioning near high-demand zones.
• RQ3, The effect of institutional and regulatory environment. This is the question our data speaks the most indirectly to, and so we state the mechanism rather than restate that Hungarian institutions are weak. What the interviews reveal is not that weak institutions led to a particular AM design, but that at no point in either firm did an institution enter the process at all. Nobody at DA had a right to be consulted, and the participatory design of the PPS was a founder’s decision that no employee could have compelled and none can prevent from being reversed. During the implementation at ConLog, there was a union with 30 years’ presence in the sector that was not involved in the process, and system issues bypassed representation and went directly to warehouse management and the deployment teams. The mechanism that is observable in the data, then, is not an active institutional weakness, but the lack of a mandatory consultation point: Hungarian law does not place AM design among the issues on which employee representation has enforceable standing. Thus, the design choice is made where the firm’s own structure places it: in the founders’ office at DA and in the parent company at ConLog. The local institutional setting plays no part in either case. From these data, we cannot demonstrate the counterfactual, i.e., how the same systems would have been implemented under stronger institutions. The German contrast in Section 6.4 is taken from the literature (Thelen, 2018) and not from our fieldwork.
• RQ4, governance implications. Our evidence speaks to this question in three particular ways rather than in the sense of general foresight. The first is the enforcement gap, which shows up in the cases as a mismatch of level: the EU AI Act imposes obligations on the deployer of a high-risk system but the design decisions that our cases turn on were made by a parent company outside the jurisdiction of the workplace and, at DA, by a founder, so an obligation discharged at the point of deployment does not get to the point at which the parameters are set. Second, the two firms have different conditions to close. At ConLog, the gap could narrow without any change in the Hungarian institutions, if the home jurisdiction of the parent company or its major clients imposed AM standards through the value chain, which is the same route by which the system arrived. At DA no such external route exists, and the gap would narrow only through general regulation, or widen the moment ownership or management philosophy changes. Third, AM is diffusing in the traditional-workplace case, where employment status already provides formal coverage, and where that formal coverage was shown to coexist with substantive exclusion from design, so our evidence suggests that the traditional-workplace case is the more consequential of the two for governance. These are the implied pathways of the cases, not predictions.

8. Conclusions and Future Research Challenges

This paper has examined algorithmic management across five Hungarian organizational cases, three digital labor platforms and two traditional firms, to answer four research questions about the forms, worker experiences, institutional shaping, and governance implications of AM in a CEE context. The five cases do not rely on comparable evidence, and this makes each finding reported below qualified: the two traditional firms were studied directly in 2022–2023. The three platform cases are a synthesis of secondary material and platform artifacts from 2019–2021, with Uber as a pre-2016 historical benchmark. The platform part of each cross-setting claim in this paper is thus inherited from prior work, not established here.

The key findings are: (1) AM in Hungary takes distinct forms depending on whether it operates through platform market mechanisms or traditional managerial hierarchies; (2) workers in both settings develop adaptive agency, but the repertoires available differ substantially based on employment status and institutional protection; (3) Hungary’s weak industrial-relations system provides limited institutional counterweights to the control-enhancing and autonomy-reducing tendencies of AM in both platform and traditional settings; the DA case suggests that organizational-level practices may partly fill that space, but on four internal interviews at a single firm this remains a hypothesis rather than a finding; and (4) comparing platform and non-platform AM in Hungary points to a probable governance gap. The EU AI Act and the Platform Work Directive are founded on the principles of transparency, accountability and worker participation, but neither firm had a counterpart to these principles at any point in the design of the AM system, with no institution entering the design. We did not measure the gap; we inferred it from the lack of institutional involvement we observed in the cases and from the secondary literature. We did not study enforcement bodies, compliance practice or the operation of either instrument. Whether and how this gap manifests elsewhere in the region is an empirical question that the comparative research agenda outlined below is designed to address.

The study suggests a number of possibilities for future research. The most obvious is to fill the evidentiary asymmetry described in Section 3.5: primary interviews with Wolt couriers and Bolt drivers and riders, conducted with the instruments developed for the traditional-firm cases, would put both sides of the comparison on equal footing and would allow the propositions advanced here to be tested directly rather than inferred from secondary sources. Second, comparative research based on the Hungarian results and applied to other CEE countries (Poland, the Czech Republic, Romania) would help us understand how the institutional variation within the post-socialist region influences AM outcomes. Such an extension would not be merely additive: because union landscapes, bargaining coverage, and platform regulation differ meaningfully across the region, comparative CEE research is the necessary test of propositions that a single-country study can only generate. Third, quantitative survey research could establish the prevalence of different AM types across Hungarian firms, providing a systematic map of the landscape sketched qualitatively here. Finally, interdisciplinary collaboration between data scientists and social scientists is urgently needed to develop AM governance frameworks that are technically credible and institutionally grounded simultaneously.