When software manages the shift: AI, job quality and the right to question a score

Workplace AI includes scheduling, allocation and performance monitoring—not only writing assistants. Its effects depend on who can inspect and challenge decisions.

A worker can experience AI without ever opening a chatbot. Software may allocate jobs, predict demand, recommend a shift or turn recorded interactions into a performance score. The experience changes when the system becomes the route through which a person receives work and is judged. A polished dashboard does not make those decisions understandable or fair.

## The evidence covers management, not just generation The OECD's 6 February 2025 algorithmic-management study draws on an employer survey of more than 6000 firms in France, Germany, Italy, Japan, Spain and the United States. That scope matters: it describes employer reports, not every worker's experience. ILO and Joint Research Centre cases in logistics and healthcare examine job-quality and surveillance concerns. The ILO's 2025 social-dialogue cases add examples of worker participation. None establishes that every management algorithm causes the same outcome.

## Ask what the system is deciding Start with the decision, not the label AI. Does software suggest a schedule or impose it? Does a score inform a manager or trigger an automatic consequence? Which data enter the calculation, how long are they retained and who can inspect them? A tool with a narrow operational purpose should not quietly become a broader disciplinary system.

The difference between a recommendation and a decision must be visible in the workflow. A human who always accepts the score without understanding it is not meaningful review. Give the reviewer authority to correct the record, consider context and change the outcome. Keep an explanation of what happened that the worker can actually use.

## Look beyond average productivity A faster average can coexist with worse work for a particular group. Examine unpredictable hours, interrupted breaks, pressure to hurry, correction workload and the distribution of difficult assignments. Check whether language, disability, unusual cases or different equipment affect the recorded measure. A speed score may reflect an inaccessible interface rather than effort.

Do not collect every available signal simply because storage is inexpensive. Define a purpose and the minimum information necessary for it. Monitoring private messages, off-duty behavior or unrelated activity raises a different set of questions from measuring a work queue. Privacy and employment duties depend on jurisdiction and context; this article is not a legal compliance checklist.

## An illustrative scheduling problem Imagine a fictional logistics team receiving computer-generated shifts. One worker's availability was recorded incorrectly, so the system repeatedly assigns impossible hours and interprets rejected shifts as poor reliability. A useful response is to correct the input, review affected decisions and provide a safe appeal path. Asking the worker to improve their score without revealing the error would compound the problem.

This example is illustrative. It does not allege misconduct by any employer in the cited studies. It shows why an input correction process and a decision challenge process must exist together: fixing tomorrow's schedule does not necessarily undo yesterday's adverse assessment.

## Make the challenge route practical Tell people where to ask for an explanation, what information to provide and how quickly they can expect a response. Preserve the relevant decision record so the reviewer is not reconstructing it from memory. Offer a channel usable during the worker's shift and in an accessible format. A complicated appeal that requires unpaid time can exist on paper while failing in practice.

Review disputes collectively as well as individually. Several similar corrections may reveal a flawed metric or policy. Worker representatives and managers can use that pattern to change the deployment, rather than repeatedly treating each person as an exception. Record agreed improvements and revisit them after software or workload changes.

## Keep responsibility with the organization The employer chooses the task, targets and consequences. A vendor's model description does not transfer that responsibility. Before expanding the system, assess who benefits, who carries errors and whether people can still exercise judgment. The useful standard is not whether the score looks objective; it is whether the organization can explain, correct and take responsibility for what the score does.