The algorithm becomes the boss: the freedom of work must be renegotiated
The real question of algorithmic management is how decision rights are divided between software and people. According to research, the same system can both narrow and widen room for manoeuvre; the difference lies in the rules on measurement, exceptions and review.
I am looking for the point at which software that helps with work becomes software that directs work. Imagine an employee who opens the task manager in the morning. The system has ranked their cases, estimated the time to resolve them, and warned about a delay. The recommendation is useful because it takes part of the burden of organising off them. Over time, however, the same interface may also determine what work they may do, how they are evaluated, and what deviation they have to explain. I therefore see the real question of algorithmic management as the distribution of decision rights. People still decide behind the software what is measured, whom is heard, and what work is considered valuable, but these decisions move behind the interface, where they are harder to question.
The leadership decision becomes part of the workflow
Algorithmic management is a much broader category than generative AI. Even a simple rule system can assign tasks or score performance. Language models offer a new way to summarize situations and phrase instructions, but the organizational question stays the same: how much room for manoeuvre does the system leave to the people who do the work?
Katherine Kellogg, Melissa Valentine and Angèle Christin's 2020 review, Algorithms at Work: The New Contested Terrain of Control, examines algorithms as new instruments of workplace control. In their approach, systems can, among other things, direct action, record and evaluate performance, and mediate rewards and sanctions. Workers' responses and forms of resistance also shape how the systems operate, so the technical solution can be understood as part of organisational power relations 1.
A task ranking, for example, can carry the organization's time pressure, customer policy and risk preferences. Whoever accepts the ranking accepts these choices too. Managers should therefore name which goals have been built into the system and in what situations deviation from them is allowed. This can help workers understand the purpose of the rule as well, and make the exception professionally justifiable.
The 2023 systematic review by Niilo Noponen and colleagues, Taylorism on steroids or enabling autonomy? A systematic review of algorithmic management, which analysed 172 studies, shows how open the question of room for maneuver remains. According to the authors, algorithmic management both restricts and enables worker autonomy, while income dependence and other factors force inflexible work practices. The literature reviewed describes a complex, digital version of Taylorism, the form of factory organisation that breaks work into small, measured steps, and finds that these systems create new organisational power structures while reshaping existing ones. According to the review, companies have so far mostly used algorithmic systems in a controlling way and have neglected the enabling side of the technology 2. From this I conclude that deployment recasts questions of power in a new form and leaves their resolution to the organisation.
Autonomy is the experience of owned action
The workplace application of Edward Deci and Richard Ryan's self-determination theory was worked out in detail by Marylène Gagné and Deci in their 2005 paper Self-Determination Theory and Work Motivation. In this framework, autonomy means that people can take ownership of their actions as their own, and the needs for competence and relatedness also play an important role. Work that conforms to external expectations can also be relatively autonomous if the person understands and accepts its purpose. Autonomy thus means something different from operating free of all constraints 3.
A colleague may accept a strict security procedure because they understand its importance, yet experience a small, unjustified click requirement as highly controlling. Beyond the element of choice, the justification, the purpose and the quality of the relationship also shape the difference between the two situations. A corporate system can therefore support work by making the role of its constraints understandable.
AI-mediated feedback shows a similar difference. The statement "you were slower than average" ranks people, whereas an analysis that also shows the complexity of the cases resolved and the time taken to fix them can create a learning situation. Managerial intent is made credible by the specific choice of data and by the consequences, so a polite tone alone is not enough for the feedback to be supportive.
The need for autonomy also differs from person to person. The 2023 qualitative study by Bruno Félix and colleagues is based on interviews with 57 Brazilian platform workers, that is, workers who make a living from occasional jobs assigned through an app. Based on the interviews, both workers and platform organizations tend toward a working relationship that offers either more autonomy or more security. Where the two preferences coincide, the worker reports greater well-being; where they differ, repeated grievances erode it. The authors note, however, that a match does not protect against everything either: if both parties choose more autonomy and less security, the worker may become vulnerable, for example to an unsustainable workload 4. This is interview material from a single country, so I read it as a pattern, and to me it suggests that the balance between freedom and protection is worth renegotiating from time to time.
Flexibility and dependence can grow together
Alex Wood and colleagues examined the quality of remote platform work among workers in Southeast Asia and sub-Saharan Africa. The study, published in a journal in 2019, Good Gig, Bad Gig: Autonomy and Algorithmic Control in the Global Gig Economy, combined interviews and survey data. Participants' work may have offered considerable flexibility and task variety, while algorithmic control, client demands and labour-market power relations were also associated with overwork and unpredictable working hours 5.
The observation makes an important conceptual distinction visible. Choosing when to start work and shaping the conditions that determine one's livelihood are two different freedoms. Someone can in principle log off at any time, yet still feel strong pressure to be available for a rating or a new task. Local autonomy should therefore be examined together with the broader relationships of dependence.
A 2023 qualitative study by James Duggan and colleagues describes the same duality from the perspective of workers who work through an app. The authors call "algorithmic HR control" a ubiquitous system that allocates and coordinates tasks, manages performance and rewards, and aligns workers' actions with organisational goals. According to their description, it differs from traditional control in two respects: it relies on algorithmic technologies, and the direct and indirect influence of parties outside the organisation is also significant in it. In the workers' accounts, the system appears rigid and complex 6.
A traditional company can take from this the question of how the new system changes bargaining positions: who can interpret the performance data, who can modify the target value, and what follows if someone progresses more slowly for professional reasons. The answers reveal more about the actual room for manoeuvre than the interface's customizability does.
The job as a whole may matter more than the task count
J. Richard Hackman and Greg Oldham's 1976 paper, Motivation through the Design of Work: Test of a Theory, describes a job design model that links skill variety, task identity, task significance, autonomy and feedback to motivation. The model highlights the mediating role of meaningfulness, felt responsibility for outcomes and knowledge of results. The authors also take individual differences into account, so the same job design may affect different people differently 7.
In a customer service operation reorganized around AI, for example, the employee may take over the handling of more complex situations and gain more latitude in resolving them. In another design, they perform a long series of small checks while seeing almost nothing of the customer's history. Both systems may use the same model, yet the work experience will differ.
In work planning, it is therefore also worth asking what adds up to a meaningful whole for the person. Closing a case that can be seen in full carries a different responsibility from accepting isolated sentences. If a process is split into several parts, a separate solution is needed so that feedback reaches whoever carried out the sub-task, who can then learn how their work contributed to the final result.
Accountability can work in two directions
Managers rightly want to know how the work is progressing. Employees may have an equally understandable need to know how the assessment of them came about. One practical form of mutual accountability is presenting the evaluation criteria in advance and offering a way to correct contestable cases.
Imagine a system that flags underperformance based on the number of closed cases, while the employee has been working on a rare problem that requires several departments to cooperate. In a good review, the organization examines what the measurement was missing. Besides correcting the individual assessment, the metric itself may change, and employee feedback becomes a source of knowledge about the workflow.
A review process works if it has an owner, time and consequences. An empty feedback box offers little guarantee that a substantive observation reaches the decision-maker. It is therefore worth pairing the technical capability with an organizational commitment, because a system's improvability becomes visible in the cases that were actually corrected.
Transparency is necessary for this, but not sufficient. A 2022 survey study by W. Alec Cram and colleagues, based on the responses of 621 Uber drivers in the United States, distinguishes two kinds of algorithmic control, the gatekeeper and the nudger. Both were associated with technostress experienced as a challenge and as a threat, that is, the strain arising from the use of technology. The authors also report a result that challenges the assumption that transparency of control mitigates the adverse effects of stress experienced as a threat 8. These are self-reported responses from the drivers of a single platform, so the relationship is an association. For me, it is nonetheless a warning: showing the rule is not enough if the worker is unable to change the rule once they know it.
Regulatory thinking arrives at the same point. The 2023 proposal by Jeremias Adams-Prassl and colleagues, Regulating algorithmic management: A blueprint, identifies two gaps: harms to privacy together with growing information inequality, and the loss of human agency. The authors tie human involvement to four points of the decision. Before the decision they propose information and consultation rights; in the decision, a ban on fully automated dismissal; after the decision, a right to meaningful review; and at the system level, an impact assessment 9. This is a policy blueprint, so it cannot be regarded as law in force, yet its structure is well suited as the skeleton of an internal policy.
The balance of support and control can be adjusted
Algorithmic management can have real benefits. It can make workload more transparent, reveal persistent bottlenecks, and reduce the role of occasional managerial impressions. These benefits, however, depend on the selected data and decision rules, since a bad yardstick applied consistently steers the work in the wrong direction just as well.
In a usable rollout, therefore, employees can take part in the design by describing common cases and exceptions. During the pilot, it is worth observing when the system helped, when it caused extra work, and in what situations its suggestion had to be overridden. From this feedback, management can learn about the real conditions of operation.
An organization can also interpret overrides in several ways. They may be rule violations, but they may also signal missing information or a well-justified professional exception. Proper interpretation requires examining the decision case by case. An automated ranking thus assists management, while managerial responsibility remains for the interpretation and for shaping the consequences.
Where feedback has no internal route, workers look for a route outside the organisation. Lorenzo Cini's 2022 study analyses the collective action of two groups of platform workers, Amazon Mechanical Turk crowdworkers and food couriers, based on the literature and policy reports. According to the author, these workers were able to act against algorithmic control even without involving traditional trade unions, and supportive communities and traditions of political activism played a decisive role in this 10. Together with the finding by Kellogg and co-authors, I conclude from this that resistance is a normal part of how the system operates, and one that the organisation can read as an early signal.
Four decision points in the organisation's hands
From all this, four decision points emerge for me that the organisation would do well to take in hand before deployment. The first is what the system measures and what it leaves unmeasured, because work that falls outside the measurement becomes invisible in the ranking. The second is who may deviate from the recommendation, and on what grounds. The third is who reviews a contested assessment, within what time and with what consequence. The fourth is who may modify the metric itself if recurring exceptions point to a flaw in the yardstick. Together, the four points form a feedback loop: an exception leads to a review, and a review leads to a rule change. Where the loop is broken, the system's errors appear as the individual performance problem of the worker, and the organisation loses the information it could learn from.
The decision-maker must also designate which questions remain in human hands. Michael Kearns and Aaron Roth's 2019 book The Ethical Algorithm, following the pattern of thinking about the moral limits of markets, raises the possibility that there are decisions an algorithm should not make, and that this calls for public debate 11. In the workplace, management and employees can answer this question together. As I see it, the answer must be put in writing before deployment, and dismissal, disciplinary consequences and pay-setting are decisions where removing human judgment also makes it unclear who is responsible. For the HR profession, this means that besides the rules of performance evaluation, the system's settings also become a matter of agreement between employer and employees.
Freedom becomes concrete in the conversation about work
Workplace autonomy shows up in practical questions: whether workers can deviate with justification, whether they get the information they need, and whether they can ask for a flawed rule to be changed, and whether the organization can tell convenient compliance from a sound professional decision. A system's interface and a conversation with a manager together answer these questions.
AI-based work organization offers an opportunity to reduce unnecessary administration and make work more meaningful. To realize the gain, however, how knowledge, decisions and responsibility are shared must be designed. An algorithm becomes a good colleague when, in using it, people also come to understand better, and can shape, the work for which they take responsibility. In my view, therefore, the agreement on the freedom of work must be renegotiated whenever any algorithmic system is introduced, and that agreement must set out the rules for measurement, exceptions and review.
Recommended reading
- Algorithms at Work: The New Contested Terrain of Control (Kellogg, Katherine C.; Valentine, Melissa A.; Christin, Angèle). A foundational review of algorithmic workplace control, which discusses the tools of direction, evaluation and discipline together with workers' responses.
- Self-Determination Theory and Work Motivation (Gagné, Marylène; Deci, Edward L). The foundational text on applying self-determination theory in the workplace, on what turns an external expectation into an action one takes on as one's own.
- Motivation through the Design of Work: Test of a Theory (Hackman, J. Richard; Oldham, Greg R). The classic model of job design, showing which job characteristics are linked to the experience of meaningfulness and responsibility.
- Taylorism on steroids or enabling autonomy? A systematic review of algorithmic management (Niilo Noponen, Polina Feshchenko, Tommi Auvinen, Vilma Luoma‐aho és munkatársai). A systematic review of 172 studies on how algorithmic management both constrains and enables worker autonomy.
- Regulating algorithmic management: A blueprint (Jeremias Adams‐Prassl, Halefom H. Abraha, Aislinn Kelly‐Lyth, Michael ‘Six’ Silberman és munkatársai). A regulatory draft that assigns human involvement to points before, during, after and above the decision, in a structure that can also be used for organisational policies.
Sources
- [1] Kellogg, Katherine C.; Valentine, Melissa A.; Christin, Angèle (2020): Algorithms at Work: The New Contested Terrain of Control. Academy of Management Annals, 14(1), 366–410.
- [2] Niilo Noponen, Polina Feshchenko, Tommi Auvinen, Vilma Luoma‐aho és munkatársai (2023): Taylorism on steroids or enabling autonomy? A systematic review of algorithmic management. Management Review Quarterly.
- [3] Gagné, Marylène; Deci, Edward L (2005): Self-Determination Theory and Work Motivation. Journal of Organizational Behavior, 26, 331–362.
- [4] Bruno Félix, Diego Dourado, Valcemiro Nossa (2023): Algorithmic management, preferences for autonomy/security and gig-workers' wellbeing: A matter of fit?. Frontiers in Psychology.
- [5] Wood, Alex J.; Graham, Mark; Lehdonvirta, Vili; Hjorth, Isis (2019): Good Gig, Bad Gig: Autonomy and Algorithmic Control in the Global Gig Economy. Work, Employment and Society, 33(1), 56–75.
- [6] James Duggan, Ronan Carbery, Anthony McDonnell, Ultan Sherman (2023): Algorithmic HRM control in the gig economy: The app‐worker perspective. Human Resource Management.
- [7] Hackman, J. Richard; Oldham, Greg R (1976): Motivation through the Design of Work: Test of a Theory. Organizational Behavior and Human Performance, 16(2), 250–279.
- [8] W. Alec Cram, Martin J. Wiener, Monideepa Tarafdar, Alexander Benlian (2022): Examining the Impact of Algorithmic Control on Uber Drivers’ Technostress. Journal of Management Information Systems.
- [9] Jeremias Adams‐Prassl, Halefom H. Abraha, Aislinn Kelly‐Lyth, Michael ‘Six’ Silberman és munkatársai (2023): Regulating algorithmic management: A blueprint. European Labour Law Journal.
- [10] Lorenzo Cini (2022): Resisting algorithmic control: Understanding the rise and variety of platform worker mobilisations. New Technology Work and Employment.
- [11] Michael Kearns; Aaron Roth (2019): The Ethical Algorithm: The Science of Socially Aware Algorithm Design.