Does algorithmic management deprive workers of discretion?

Kellogg et al. organized workplace algorithmic control into six functions: restricting and recommending work, recording and evaluating behavior, and replacing and rewarding people. [1] Even within the same app, the authority and psychological load differ between a function that proposes work candidates and one that sanctions those who refuse.

When comparing departments that have implemented "AI" with those that have not, while keeping management functions bundled, it is unclear which functions clarify roles and which deprive discretion. We treat instructions, monitoring, evaluation, rewards, sanctions, and appeals as separate units of change.

This article does not address how generative AI changes total employment or the hiring of young workers. Nor does it limit itself to KPI gaming. It addresses how work discretion, health, voice, and remedies change when management authority is transferred to software in current workplaces.

• DIRECT: Distinguish between showing work candidates and forcing choices.
• OBSERVE: Distinguish between measuring necessary outcomes and monitoring behavior at all times.
• DECIDE: Distinguish between assisting evaluation and automatically determining rewards and sanctions.

Kinowska and Sienkiewicz analyzed 21,869 organizations included in the 2019 European enterprise survey and reported that algorithmic management practices are associated with lower job autonomy and workplace well-being, with the relationship varying by scale. [2] It is important that this includes not only platform workers but also typical organizations.

However, this is cross-sectional organizational-level data and a structural equation model, so it cannot be determined that implementation caused the decline. There remains the possibility that workplaces with poor performance or labor shortages implemented management technology, and confounding by industry, institution, and management quality. [2] Do not transfer coefficients to your company's causal effects or health risk rates.

In practice, take baseline values before implementation, use phased implementation, or place comparable controls. Measure autonomy, role clarity, load, and well-being simultaneously, and track changes after implementation by management function. Combine questionnaires with behavioral records of work changes, breaks, appeals, turnover, and absenteeism.

The paper published by van Zoonen et al. in 2025 analyzes 590 European crowdworkers who completed Clickworker surveys in July and September 2022. The relationship between perceived algorithmic control and work frustration included a supported pathway through reduced autonomy. Control was also associated with difficulty disconnecting during work episodes, but the pathway to frustration through connectivity was not statistically supported. [3] Flexibility in choosing working hours and discretion over how to work should be considered separately.

As a Catalysphere operating hypothesis, test whether short-lived opportunities, response-speed evaluation, and rewards for constant availability make disconnecting harder. This paper did not measure intervention effects of notification or reward design. We propose recording availability and performance separately and checking before deployment whether notification pauses or breaks create disadvantages.

This is a self-reported, two-wave observational study on one platform, with 54.3% attrition from the initial respondents. The authors note that mediation is ideally tested at three or more time points; causal direction remains unresolved. The publisher displayed In Press, Corrected Proof when checked on 1 October 2026. [3] We do not generalize the results to all occupations or connectivity outside work; test fewer notifications or greater discretion as local operating proposals through an organizational intervention comparison.

• AVAILABILITY: Separate response speed from outcomes and do not require constant connectivity.
• PAUSE: Test that breaks, notification pauses, and end of work do not lower evaluations.
• OVERRIDE: Allow changes to assignments, order, and deadlines based on on-site conditions.

For Liang et al.’s 2025 Journal of Business Research paper, this review checked the publisher’s public abstract, introduction, and excerpts from methods and conclusion. The authors report developing scales for instructive guidance, process monitoring, and evaluation feedback across multiple samples, followed by a multisource, three-wave study of 359 app workers with beneficial and detrimental pathways through job embeddedness and anxiety. [4] Effect sizes and analytical robustness have not been checked against the complete paper.

Standardized guidance and rapid feedback can be resources that reduce role ambiguity. If the same mechanism involves constant monitoring and high demands, it may be associated with greater anxiety. Since average satisfaction may cancel out the two pathways, we measure role clarity, discretion, monitoring frequency, anxiety, and service quality separately.

The public participants-and-procedure excerpt describes surveys of Chinese food-delivery and express-delivery workers, using self-reports and platform service scores. We treat this as a report of observational mediation analysis, not evidence of causal effects from randomizing management functions. [4] The complete methods, results, and limitations remain unread. Because pay satisfaction is also reported as a contextual condition, Catalysphere proposes separate approval of guidance, monitoring, and sanctions.

In Safety Science 2025, Nilsson et al. compared two Swedish logistics companies through observation and semi-structured interviews. Both companies reported low autonomy and meaning in work, but physical and psychological burdens were stronger in the EC company, suggesting a relationship with economic pressure and differences in organizational practices. [5]

The improvement candidates indicated by the study are: involving workers in technology design, integrating them into occupational safety and health management, creating space for workers to exercise control, and adding managers' qualitative judgments to quantitative evaluations. [5] As a practical design hypothesis, consider the people and procedures that handle appeals alongside algorithmic accuracy.

This is a comparative case study of two companies and is not a study that estimates the causal effect of interventions. It relies on researchers' interpretations and participants' reports and cannot be generalized to other countries or industries. [5] Nevertheless, it can be used as evidence to deeply understand how the impact of similar technologies varies by organizational conditions.

• PARTICIPATE: Involve workers in the target business in requirements, testing, and change approval.
• CONTEXT: Allow adding on-site conditions and reasons for exceptions to quantitative scores.
• REVIEW: Have humans re-review sanctions and low evaluations, explaining the basis and remedies.

For efficiency metrics, place assignment time, wait time, deadlines, rework, and customer outcomes. For work quality, place discretion, role clarity, load, connectivity pressure, anxiety, meaning, voice, safety, and appeals. We set dual minimum standards so that improvements in efficiency do not offset deteriorations in health or remedies.

Comparisons should be made not only before and after implementation but also by stage: proposal only, proposal plus optional change, with monitoring, linked to evaluation, and linked to sanctions. We look at at which stage benefits increase and at which stage burdens jump. We also check for cycles where monitoring intensifies only for those with low evaluations.

We look at distributions not just by averages, but by occupation, employment type, skill level, and pre-approved conditions such as disabilities or caregiving. We do not treat insufficient samples as "no difference," but leave the path where sanction automation is stopped. We do not treat a low number of appeals as success, but measure availability and remedy completion time.

• PERFORMANCE: Measure not just volume, but also quality, rework, and customer outcomes.
• AGENCY: Measure the range of choices for order, method, breaks, exceptions, and rejection.
• HEALTH: Continuously measure load, anxiety, fatigue, and connectivity pressure.
• REMEDY: Measure the time and completion rate from explanation, appeal, correction, to compensation.

Initially, limit to proposals and visualization, starting from a state where workers can change with reasons. Before using for evaluation, audit the distribution of incorrect assignments and exceptions. When linking to rewards or sanctions, have worker representatives, managers, safety and health, and legal departments approve it as a separate change.

The reasons for algorithmic judgments should be shown not only by model explanations but also by used data, rules, thresholds, application period, and change history. Individuals must be able to correct original data, add on-site conditions, and request human review. We also test that making an appeal does not lead to unfavorable future assignments.

Suspension conditions include not only failure to meet efficiency minimums, but also reduced autonomy, health signs, hindered breaks, pending appeals, and disadvantage to specific groups. To resume, separate the problematic function, restore discretion and human review, and correct affected judgments. AI that manages people is approved as a broader work design than model quality.

• Adopt: Ensure the proposal function protects both performance and work quality minimums.
• Limitation: Phase and approve monitoring, evaluation, rewards, and sanctions separately.
• Suspension: Suspend if any of autonomy, health, breaks, or appeals breach the minimums.
• Resumption: Restore discretion and human review, and correct past judgments.