Inequality is often hidden within the average. In decent work in the age of automation, efficiency was achieved by transferring friction to people with less ability to claim.

An average can describe a real improvement while concealing who pays for it. If the mean time needed to complete a task falls from twelve minutes to eight, a dashboard will show progress. It may not show that part of the workforce finishes in six minutes while another part needs twenty-five because it receives the difficult cases, uses older equipment, or has to correct information the system cannot recognise. Nor will it necessarily show missed breaks, questions answered after a shift, or people who leave before their experience is recorded.

That invisibility is particularly dangerous in automated work. Systems for allocation, forecasting, recruitment, and evaluation produce figures that look precise. Arithmetic precision does not guarantee fair representation. A metric can be calculated without error and answer the wrong question; it can summarise a site's productivity while ignoring how the burden is distributed; it can raise customer satisfaction while hiding the exhaustion of the people who absorb every exception.

Inequality does not always appear as a difference between two clearly defined groups. Sometimes it clusters around circumstances: short contracts, night shifts, caring duties, disability, limited tenure, or the need to use an alternative channel. These circumstances can intersect and create small groups that disappear inside an overall mean. A single round of segmentation is therefore insufficient. Friction has to be heard where it arises, the data must be checked for its ability to represent that friction, and the decision taken when a difference appears must be reviewed.

Measurement by group does not authorise permanent labelling of individuals. The purpose is to test outcomes and correct processes, not to build unnecessary profiles. Relevant categories, proportionate data, and privacy-preserving thresholds are required. When a sample is small, combining periods, using qualitative evidence, or conducting a confidential review may be preferable to publishing a table from which somebody could be identified.

The following three cases are educational composites. They combine plausible situations in recruitment, logistics, and public service, but they do not describe identifiable employers or replace labour, safety, or legal advice. Each reveals the same pattern: an aggregate figure celebrates efficiency while costs move towards those with less time, stability, health, or power to challenge a decision.

Looking beyond the mean without turning anyone into a category#

A distributional assessment begins before a tool is installed. The problem to be solved, the people expected to benefit, and the people exposed to harm must be identified. A baseline needs more than a mean: the distribution of time, percentiles, variation between shifts, errors, dropouts, complaints, corrections, and later outcomes. If only the favourable figure is preserved, any subsequent inequality will look anecdotal.

A distributional assessment begins before a tool is installed. The problem to be solved, the people expected to benefit, and the people exposed to harm must be identified.

Outcome and effort should be separated. Two people may complete the same number of tasks, yet one may face more travel, interruptions, or emotional labour. An algorithm that counts only completed items turns that invisible effort into an apparent performance deficit. Choice and constraint must also be separated. Declining a shift may indicate disinterest, or it may follow from late notice, a poor connection, or a caring responsibility the system never recorded.

Comparison should use groups defined for a legitimate reason, not for statistical curiosity. Shift, role, or exposure may matter in occupational safety. Progression and error rates across protected characteristics may matter in recruitment where there is an appropriate basis for analysis. Language, channel, or support needs may matter in public services. Every breakdown requires a purpose, restricted access, a retention period, and an explanation of the action to be taken if a difference emerges.

Quantitative evidence needs short accounts of experience. A long tail of completion times shows that something is happening, but it does not explain whether permissions are missing, a screen reader fails, or complex cases are always sent to the same team. Interviews, observation of work, and complaint reviews turn a statistical anomaly into a mechanism that can be corrected. Worker voice is not a participatory ornament; it supplies operational knowledge that the database does not contain.

Authority is also required. If a review finds that a group is losing income or rest, somebody must be able to pause the rule, restore the outcome, and require changes from a supplier. A committee that merely receives a quarterly presentation is not supervising; it is witnessing a decision already made. Real accountability combines access to records, knowledge of the process, sufficient independence, and the power to provide remedy.

The decisive question is not whether everyone receives exactly the same treatment. Some differences reflect valid needs or agreements. The question is whether a difference is connected to the purpose, proportionate, explainable, and open to context from those who bear it. Superficial equality can be unfair when it demands an identical pace from different tasks; an adjustment can be fair when it removes a documented barrier without closing opportunities.

Case one: the application filter that improved average response time#

In the first composite case, an employer introduced an automated filter to order applications for entry-level roles. Its purpose was to reduce weeks of waiting and respond to more candidates. After the pilot, the average time to first review fell and the proportion of vacancies filled on schedule rose. Management read both figures as an unequivocal improvement.

The distribution told a different story. Applications with linear histories and familiar qualifications progressed quickly. Candidates who described experience gained through care, informal employment, overseas education, or a career change entered a low-priority queue. The system did not explicitly reject those profiles. It assigned lower scores because their words and sequences departed from the historical examples used as a reference. Mean response time improved while one group waited until the vacancy had already closed.

Friction also fell on those least able to insist. Applicants with internal contacts obtained a manual review; those familiar with industry language rewrote their résumés; everyone else received a generic message. The appeal route requested an application number that was not displayed on the mobile version. A human alternative formally existed, but its design reserved access for people who already knew how to navigate the organisation.

The review went beyond comparing final hires. Progression rates, waiting times, and correction reasons were measured by type of career history; a sample of scores was audited; and recruitment professionals examined false negatives without seeing the initial score. Employment organisations were also asked how capability is expressed outside a conventional résumé. The assessment showed that several signals associated with quality were actually signals of familiarity with a format.

Administrative automation remained, but the score stopped being the only gate. Random samples of lower-scored profiles were reviewed, equivalent experience was recognised, the appeal was simplified, and a maximum response time was set for every application. When a persistent difference appeared, the rule was suspended until its cause could be explained. Improvement was no longer measured through speed alone; it included opportunities that the filter had been losing.

Case two: the most efficient warehouse route and the costliest route for some bodies#

In the second composite case, a logistics centre deployed a system that assigned routes and picking sequences. Average distance per order fell and hourly shipments rose. The supplier presented a stable chart with no striking differences between days. Requests for reassignment and reports of muscular discomfort nevertheless increased on two shifts.

The mean combined tasks that were not equivalent. Some routes contained light products at a comfortable height; others accumulated weight, twisting, and collection from low shelves. The optimiser sought minimum distance, not biomechanical load. Because it also learned from historical completion times, it assigned the most demanding routes to those who had previously completed them quickly. Demonstrated ability became a repeated obligation and erased the need for recovery.

Temporary workers carried a disproportionate share. They knew less about prevention channels, feared losing future offers, and accepted warnings that permanent staff discussed with a supervisor. The dashboard recorded every delay, but not pain mentioned in a conversation or minutes spent helping a colleague. Detailed monitoring coexisted with organised blindness.

The corrective test added cumulative weight, picking height, turns, breaks, rotation, and safety alerts. Results were analysed by shift, tenure, and contract type with confidentiality safeguards. A difficult-day simulation, including a breakdown and staff shortage, tested whether the system again concentrated effort on the same people. Worker safety representatives were given the power to stop a specific configuration.

Overall productivity fell slightly during the adjustment, but corrections, reported fatigue, and turnover also fell. The relevant measure stopped being the maximum number of orders during an ideal hour and became an operation that remained sustainable over weeks. Technology continued to propose routes, but a favourable average could no longer serve as permission to ignore the physical distribution of cost.

Case three: the service assistant that moved exceptions to the last link#

In the third composite case, a public service deployed an assistant to classify enquiries and propose answers. Most simple procedures were resolved sooner and mean handling time fell. Success was communicated as the percentage of conversations closed without intervention. People who repeated a question, switched channels, or visited an office after failing to obtain a useful answer disappeared from that account.

Complex cases reached a small team at the end of the chain. Language barriers, missing documents, unanticipated family circumstances, and decisions with financial consequences accumulated there. The assistant saved time upstream by transferring ambiguity and tension to professionals with less room to respond. Because each automated conversation counted as a closure, a person making three attempts appeared as three successes, while a later office visit appeared as a new matter.

Inequality most affected those unable to devote a morning to repairing the process. Someone with flexible hours could call, insist, and travel to an office. A person combining shifts, care, and travel abandoned the attempt sooner. Dropout did not appear as a denial and therefore triggered no review. The average measured those who managed to stay inside the process, not the full population that needed the entitlement.

The institution linked channels through protected case identifiers, distinguished technical closure from confirmed resolution, and reviewed repeated enquiries. It also created a visible transfer to a person, preserving context instead of requiring a new start. Offices recorded why visitors arrived after a digital attempt. That follow-through located confusing forms and categories that treated common exceptions as individual mistakes.

The assistant retained its usefulness for simple questions. Assessment, however, began to observe first-contact resolution, repetition, dropout, the person's total time, and differences across channels and support needs. A human route stopped being a secret concession and became part of the design. Efficiency no longer meant closing more conversations; it meant resolving a procedure without transferring its cost to the people least able to challenge it.

Four institutional sources for interpreting inequality#

The International Labour Organization paper on algorithmic management defines the phenomenon as the use of programmed procedures to coordinate labour and examines implications for work organisation and job quality in both platforms and conventional workplaces. That perspective prevents a purely technical account: an allocation rule also organises autonomy, intensity, information, and power.

The European Agency for Safety and Health at Work brings together evidence on digitalisation, automation, and algorithmic worker management. Its prevention-centred approach is a reminder that productivity is not the whole outcome. Workload, stress, privacy, autonomy, and participation must be examined as dimensions of the same workplace change.

The OECD report on using artificial intelligence in the workplace presents opportunities alongside concerns about work intensity, data collection, and increasing inequality. Positive aggregate findings from its surveys do not remove those risks. They show why it is necessary to ask who obtains the improvement, in which occupation, and under what conditions.

The EU Platform Work Directive 2024/2831 sets specific rules on transparency, data protection, and human oversight for automated systems in that field. It should not be extended by analogy to every workplace, and it does not replace appropriate legal analysis. It does provide a concrete institutional reference for explanation, review, and participation when algorithmic decisions affect platform work.

The four sources have different scopes. The ILO supplies an international labour framework; EU-OSHA focuses on safety and health; the OECD synthesises evidence and policy options; and the Directive establishes obligations in a defined field. None offers a universal fairness formula. Together, they support better questions and prevent a single mean from being presented as sufficient proof of decent work.

Sources must be used precisely. A reference does not automatically validate a tool, and an obligation addressed to platform work should not be represented as covering every workplace. A concrete assessment depends on jurisdiction, employment relationship, purpose, data, and consequence. Honesty about those limits is part of institutional rigour.

A distributional test before automation is expanded#

Before a system is scaled, a test can be organised in six movements. First, the purpose is fixed and the decision that will change is identified. Next, a baseline records distributions rather than the mean alone. Third, relevant groups and situations are agreed with workers and safety functions. Ordinary cases, exceptions, and a realistic interruption are then tested. Benefits and costs, including those outside the dashboard, are compared. Finally, a named person decides whether to continue, correct, or stop, and sets a review date.

The test needs prior thresholds. If one group loses opportunities, experiences more late changes, or accumulates more errors, the difference that triggers investigation and the harm that requires a pause should be known. There is not always a magic number. A quantitative signal can be combined with severity, persistence, and testimony. What matters is preventing the standard from changing after an uncomfortable result appears.

The remedy must also be tested. A person who did not work on the project can try to understand a decision, provide context, and obtain correction. Total time, number of steps, and the authority of the responder are measured. A route that requires expert language, physical attendance, or internal contacts will reproduce the inequality it was meant to repair. An alternative must reach an equivalent outcome without humiliation or retaliation.

Minimum documentation can fit on one sheet: purpose, responsible people, data, groups examined, distributed results, incidents, voices consulted, changes, time limit, and exit rule. That memory allows reviews to be compared and displacement to be detected. It also makes it harder for an average advantage to be repeated for years while exceptions are treated as personal failures.

A mature organisation does not promise that all differences will disappear. It promises to look for them, explain them, reduce those that are unjustified, and repair consequences. That work requires curiosity about the figure, accountability to intervene, and care for those who carry the exception. These values do not obstruct innovation; they give success a more demanding definition.

Inequality is often hidden within the average because an average removes names, trajectories, and extremes to produce a manageable figure. The response is not to abandon every summary, but to remember what summarisation removes. Work that does not surrender dignity measures improvement and its distribution; listens to those who drop out; gives supervisors authority; and preserves a way out before calling a system efficient.