When the manager is a model

A warehouse worker opens the scheduling app on Sunday evening. The coming week contains fourteen hours. The week before contained twenty-seven.

No manager has called. A notice says that shifts are assigned according to demand, availability, and performance. She presses the link marked “Why this schedule?” and reaches a page explaining that hours vary. Her supervisor can see the same timetable she can, though not the weighting that produced it.

Several explanations seem possible. One packing station jammed during her shift and lowered the team’s rate. She declined an evening at short notice because her son was ill. A bus delay made her seven minutes late on Tuesday. Customer orders may have fallen. Perhaps the schedule has nothing to do with her record at all.

The uncertainty matters because fourteen hours will not cover the month. She has been managed, in the practical sense: her time, income, and place in the organisation have been decided. Yet she cannot find anyone who recognises themselves as the author of the decision.

This is what changes when the manager is a model. Management becomes less like an encounter with a person and more like a climate surrounding the working day. Tasks arrive, targets adjust, routes tighten, and ratings accumulate. The system need not issue an order in a human voice. It shapes which hours are available, which pace counts as normal, and which worker appears to be falling behind.

A decision without a meeting

Managers have always relied on representations. Rotas, sales figures, attendance records, case totals, and performance reviews make a large organisation possible. No supervisor can witness every action or understand every circumstance.

Algorithmic management extends this old distance. A system can gather far more signals, compare workers continuously, and act on the result before anyone has decided that a meeting is required. It may distribute delivery jobs, predict staffing needs, flag an unusual pause, rank service agents, or recommend who should receive another contract.

Each action can appear small. A route is reassigned. A break is recorded. A lower-priority task reaches one person rather than another. Together they determine the texture of work: whether a person can plan a week, learn a valuable task, recover after an error, or remain visible as someone worth investing in.

Human managers exercise this power too, often inconsistently. They play favourites, miss good work, punish candour, and remember an old mistake long after it should have faded. A worker may prefer a clear rule to the private moods of a supervisor. Data can reveal that opportunities have been distributed unfairly, and automated scheduling can spare people hours of avoidable administration.

The difference lies in the reach and rhythm of the judgement. A supervisor’s view is partial, but a worker usually knows when that view has entered the day. A model can alter conditions continuously while each adjustment presents itself as an operational fact. Power becomes harder to recognise because no single moment looks large enough to contest.

The worker the record can see

A management system encounters work through what has been made measurable.

It can see how many items moved, how long a call lasted, when a vehicle stopped, whether a task was reopened, and how a customer scored the exchange. These signals may correct flattering self-assessments and expose problems that informal supervision overlooked. They also divide the job into events the system can register and events that remain outside its account.

The experienced employee who steadies a frightened colleague may lose time against a target. The driver who waits because a street feels unsafe may appear inefficient. The adviser who notices that a customer’s repeated questions conceal another problem may lengthen a call that the dashboard expected to end. Even when such acts are praised in principle, the measure applied every hour can make them expensive in practice.

Workers learn the shape of the record. They discover which contribution leaves evidence and which disappears. Soon the job may be performed for two audiences: the person, object, or situation in front of them, and the system assembling a judgement elsewhere.

This divided attention changes professional judgement. If a nurse, teacher, courier, or customer adviser must repeatedly choose between what the moment requires and what the metric rewards, the organisation has already made an ethical decision. It has told the worker which reality will count when performance is reviewed.

A measurement need not be complete to be useful. Trouble begins when its boundaries cannot be seen by the people governed through it. A worker should be able to know which part of the job a score represents, how uncertainty enters it, and where important work has no corresponding field in the data.

The appeal of consistency

There are good reasons to place some managerial decisions in systems. A fair rota cannot depend on who feels comfortable asking the boss for a better shift. Safety rules should survive a supervisor’s impatience. Similar work deserves similar standards, and patterns of discrimination are easier to challenge when decisions leave a record.

Consistency, however, can preserve a bad rule with great discipline. It can also hide unequal conditions behind identical treatment. Two workers may record the same delay even though one chose a long lunch and the other waited for faulty equipment to be repaired. A scheduling model may treat complete availability as neutral while rewarding people whose care, health, or transport allows them to offer it.

The system does not remove discretion. It relocates it. Someone chooses the objective, the signals, the tolerance for error, and the point at which a score changes a person’s opportunities. Someone decides whether a missed target triggers support, scrutiny, or fewer hours. These are managerial judgements even when software applies them at scale.

Calling the result efficient tells us that a process reached its chosen target with fewer resources. It does not tell us whether the target was fair, whether the burden moved onto workers, or whether the organisation preserved the knowledge needed to notice harm.

For that, people must be able to examine more than the output. They need an account of the policy expressed through the model: what kind of worker the organisation is rewarding, which risks it asks employees to carry, and what it believes a good day’s work consists of.

Responsibility behind the screen

An algorithm cannot be embarrassed by an unjust rota. It cannot recognise that an explanation has failed, make an exception because the rule met a circumstance its designers did not anticipate, or accept a worker’s anger as a claim on the relationship. The institution using it can do all of these things.

Responsibility therefore cannot end with access to a generic explanation. “Your score reflects multiple factors” may describe the system accurately while answering none of the worker’s concerns. Nor is a nominal human review enough when the reviewer can see only the score, lacks authority to restore the hours, or is punished for departing from the recommendation.

A real appeal changes the balance of power. It lets a person correct the record, introduce relevant circumstances, question the rule itself, and reach someone authorised to alter the outcome. It creates evidence for the organisation as well. Repeated appeals may show that a metric is misreading a task, that a schedule is pushing risk onto a particular group, or that the model’s apparent precision exceeds the knowledge beneath it.

This route must be safe to use. A right to challenge means little if the worker expects fewer opportunities for being difficult. Collective representation matters here because many patterns become visible only when separate workers compare what has happened to them. One unexplained reduction in hours can look like chance. Twenty may reveal a policy no individual was able to see.

The responsible manager may therefore be a team rather than a single person: supervisors who understand the work, technical staff who can examine the system, leaders who control its objectives, and workers with recognised power in its design and review. What matters is that responsibility has an address and that someone at that address can act.

Power that can be answered

Workplaces will continue to use models to coordinate complex activity. The humane response is not to recreate an age of intimate supervision that many workers never experienced. It is to prevent organisational authority from becoming unreachable as it becomes computational.

That begins before a system is installed. Workers can help identify which signals misdescribe the job, which decisions carry serious consequences, and where human discretion protects safety or care. An organisation can test whether the tool distributes undesirable shifts, development, scrutiny, and income fairly, not only whether it predicts demand.

During use, consequential changes should arrive with an intelligible reason and a path to timely review. Managers need the time and authority to depart from a recommendation. Workers should know when they are being evaluated, which data travels between contexts, and when an inference will expire. No one should have to become a data scientist to ask why their working life changed.

The warehouse worker may learn that orders really did fall. A good process cannot promise her hours the work does not support. It can tell her whether everyone shared the reduction, whether her performance affected the result, when the next schedule will be decided, and who is accountable for the policy. If the data was wrong, someone can correct it. If the rule was unfair, the organisation can change more than her individual score.

Management has always contained an unequal relationship. One side can allocate time, income, opportunity, and judgement; the other must live inside those decisions. Technology can make that power more consistent and better informed. It can also allow those who hold it to disappear behind the process.

When the manager is a model, the central human requirement is answerability. A worker needs more than a result produced without bias in a technical sense. They need to encounter an institution capable of hearing a reason, explaining its own, and changing course when the people inside the record reveal what the record could not see.