When measuring better requires not looking at a single figure. An organisation may announce that productivity has risen, deliveries arrive sooner, or an automated system is correct in most cases and still be making working life worse for part of its workforce. An average describes a surface; dignity requires attention to distribution, exceptions, and the invisible labour supporting the result.
Good measurement does not mean accumulating indicators. It means connecting every figure to a question, a population, a period, and a consequence. An average time can hide increasingly unpredictable working days. An overall error rate can coexist with failures concentrated among disabled people, carers, or workers on insecure contracts. High satisfaction can silence those who left before they could answer.
Automation intensifies the problem because it turns partial signals into rapid decisions. A dashboard summarises, ranks, and recommends; an accountable person then accepts the output under pressure. Mathematical appearance makes it easy to forget that someone chose the objective, variables, and threshold. Someone also decided what would not be counted. A metric never arrives alone: it organises attention, authority, and resources.
Decent work does not require rejection of efficiency. It requires efficiency to remain compatible with fair income, safety, reasonable autonomy, participation, and remedy. A professional assessment must therefore examine the total benefit as well as who produces it, who receives it, and who absorbs its costs. If organisational savings depend on lost breaks or unrecognised correction work, the balance is incomplete.
The following three cases are educational composites based on plausible situations, not accounts of identifiable organisations or findings from field research. They are not legal, employment, or occupational safety advice. Their purpose is to show how an aggregate figure can celebrate improvement while concealing a small harm for many people and a serious harm for some.
A useful average begins by declaring what it leaves out#
Every indicator compresses reality. Compression is valuable when it helps detect change, compare alternatives, and direct attention. It becomes dangerous when the summary is presented as the whole reality. Before adopting a metric, an organisation should document its purpose, unit of analysis, source, frequency, margin of error, and the decisions it may trigger.
Every indicator compresses reality. Compression is valuable when it helps detect change, compare alternatives, and direct attention.
An arithmetic mean says nothing about dispersion or extremes. Two teams may record the same average handling time while one maintains stable workloads and the other alternates waiting with episodes of saturation. Percentiles, ranges, and distributions provide a better picture, but they do not explain by themselves why a difference exists. Conversation with affected people completes the diagnosis.
Disaggregation should respond to relevant risks while respecting privacy. The task is not to collect unlimited personal categories but to identify where effects may concentrate. Contract type, shift, site, seniority, or working arrangement can reveal patterns. When small groups could identify an individual, safeguards, sufficient aggregation, and restricted access are needed.
The denominator matters too. A company may report that ninety per cent of requests are resolved within a day if it counts only those that successfully pass through a form. People unable to complete it disappear from the statistic. Measuring abandonment, repeated attempts, and requests through alternative channels prevents the organisation from confusing system convenience with service quality.
Qualitative measures supply context. Interviews, task observation, incident records, and meetings with representatives may reveal fatigue, loss of professional judgement, or fear of challenging an outcome. Such accounts are not anecdotes opposed to data; they are another kind of data. Triangulation tests whether the indicator represents the working experience it claims to summarise.
Mature governance establishes in advance which result would justify continuation, which inequality would require review, and which harm would require a stop. Without action thresholds, monitoring becomes observation without responsibility. Whoever receives an alert needs time, competence, and authority to intervene, while workers need to know the channel through which they can add context or challenge a conclusion.
Case one: the warehouse that gained minutes and lost breaks#
In the first composite case, a logistics centre introduced a planner to allocate routes. For six weeks, the dashboard showed a twelve per cent reduction in average time per order. Management presented the figure as a clear improvement. Volume increased without additional hiring, and the overall number of mistakes remained stable.
The distribution told another story. Simple routes had improved greatly, while one area with high shelving accumulated delays. To recover the target, the system repeatedly sent the same experienced workers there and automatically postponed their breaks. Each change involved only a few minutes, but it occurred several times a week. On two occasions, a worker with diabetes lost the safe window in which she could eat.
The principal indicator did not record displaced breaks, walking distance, recovery between tasks, or concentration of reassignments. Nor did it include the minutes that other people spent providing informal help. The apparent gain combined genuine optimisation with intensified labour and care between colleagues. The average turned both components into a single success story.
The review added time percentiles, distance, breaks actually taken, late changes, and incidents by area. The safety committee examined distribution by shift without exposing individual diagnoses. People could declare functional constraints through a confidential channel. The planner stopped moving agreed breaks, and repeated reassignment triggered an ergonomic review.
The new measurement did not destroy the benefit. It preserved better routes while correcting the hidden cost. Productivity remained above the baseline, though less than the first presentation claimed. The difference represented time required to work safely. The more modest figure was institutionally more honest and allowed improvement to endure without relying on the silent sacrifice of a few people.
Case two: the contact centre that confused silence with satisfaction#
The second composite case concerns a service centre that introduced automated answers and a call-scoring system. Average handling time fell, the follow-up survey returned high ratings, and formal complaints declined. The organisation concluded that technology had simplified the work and improved the experience of customers and staff.
However, the questionnaire reached only people who completed the digital flow. Those requesting an adjustment, using a screen reader, or unable to understand a category became stuck, ended the call, or visited in person. These abandonments did not appear as complaints. Within the team, difficult cases reached a small group of agents because the system rewarded those who kept conversations short.
The favourable average concealed two transfers. First, effort moved from the interface to customers facing greater barriers. It then moved to staff who agreed to solve exceptions and watched their own scores deteriorate. Some people stopped offering additional help to protect the indicator. The metric did not merely describe conduct; it taught people which conduct the institution rewarded.
The redesign widened the measurement funnel. It included attempts, abandonment, repeat contact over seven days, effective resolution, and use of accessible routes. It separated case complexity from agent performance and reviewed conversation samples with workforce participation. The survey offered alternative channels and ceased to be presented as the sole evidence of satisfaction.
The service also reserved time for exceptions and recognised informal mentoring as work. An interdisciplinary team examined the journeys with most abandonment each month and could change wording, categories, or rules. Complaints initially rose again; rather than proving failure, the increase showed that the route for expression was more visible and safe. Better measurement made an existing problem audible.
Case three: the factory with fewer reported incidents#
In the third composite case, a factory installed sensors and cameras to detect approaches to hazardous areas. The system generated immediate alerts and a weekly index for each production line. After one quarter, reported incidents had fallen and two lines had achieved a green rating. The project was described as a data-supported preventive improvement.
A participatory inspection found an uncomfortable explanation. Alerts had gradually become linked to performance assessments. When a person received several, they had to justify their conduct to a manager. Workers began avoiding reports of false positives and resolving obstacles without documenting them. One line with worn markings appeared safe because staff had learned to move outside the camera’s field.
The decline measured silence, not necessarily safety. The individual indicator did not distinguish between reckless conduct, a blocked route, production pressure, or sensor failure. By concentrating on the last visible movement, it erased earlier organisational conditions. The person became the cause of an episode that spatial design and work pace had helped produce.
The committee separated purposes. Individual alerts would prevent immediate danger and would not automatically feed discipline or productivity assessment. Aggregate analysis would look for repeated locations, times, and conditions. Before attributing an action to a person, review would examine design, maintenance, workload, training, and testimony. Staff would have access to the relevant sequence and a human route for correction.
The factory added leading indicators: obstacles reported, repair time, participation in safety walks, and changes completed. Alerts and reports increased for several months, but repetition at the same location fell. The dashboard stopped rewarding the absence of voices and began measuring the capacity to learn. A worse short-term figure represented a more reliable preventive culture.
Four institutional sources for widening the picture#
The International Labour Organization’s framework on decent work connects job creation, rights at work, social protection, and social dialogue. Its definition includes fair income, safety, development, participation, and equality. Applied to automation, this perspective prevents success from being reduced to output or headcount and requires attention to the concrete quality of work.
The OECD Artificial Intelligence Principles connect inclusive benefit, human rights, transparency, robustness, and accountability. For a measurement policy, they support understandable information about operation and limits, proportionate traceability, and identifiable responsible parties. An average without an explanation of uncertainty or a possibility of challenge falls short of that discipline.
The NIST Artificial Intelligence Risk Management Framework organises work around governing, mapping, measuring, and managing. Its approach treats assessment as a continuing cycle: context changes, emerging effects appear, and a metric suitable during a pilot may become unsuitable when population or purpose expands. The framework is voluntary and does not replace legal duties or collective bargaining.
The European Agency for Safety and Health at Work’s resources on digitalisation identify physical, psychosocial, and organisational opportunities and risks in systems that collect data and support decisions about tasks, shifts, and performance. This perspective places health inside design and management rather than treating it as a late check after production objectives have been met.
The four sources do not provide a universal formula or automatically validate a tool. Together, they support a direction: measure purpose as well as impact; include diverse perspectives; maintain clear accountability; review throughout the life cycle; and connect innovation to rights, safety, and participation. Each organisation must translate that direction into its context, agreements, and applicable obligations.
A dashboard that can see distribution, voice, and remedy#
A balanced dashboard can organise indicators into five families. The first examines outcomes: quality, time, safety, and value created. The second examines distribution: extremes, differences between groups, and concentration of changes. The third records conditions: workload, breaks, predictability, and autonomy. The fourth examines voice: participation, consultations, challenges, and abandonment. The fifth follows remedy: time limits, changed decisions, and recurrence.
Every indicator needs a short specification. It should state who produces it, which population it includes, what it excludes, how often it is updated, and which level triggers a response. It should also identify who can see it and how long it is retained. This specification prevents data intended for learning from becoming material for individual sanctions without analysis or safeguards.
The dashboard should present absolute numbers alongside proportions. A stable rate can hide a growing number of affected people when volume increases. Means should appear with percentiles and ranges, and results with an earlier baseline. If the definition changes, the series needs a visible warning; otherwise, a methodological improvement may look like a workplace improvement that never happened.
Periodic review requires a plural forum including management, technical staff, prevention, human resources, and workforce representation. Its task is not to approve colours but to interpret contradictions and commission action. When one measure improves and another deteriorates, the group should investigate the relationship, hear cases, and document why the system continues, changes, pauses, or is withdrawn.
Affected people need an accessible explanation and a channel without retaliation. A challenge should not require expertise in the model or proof of discrimination as an entry condition. Supplying relevant context or identifying an incorrect consequence should be enough. The organisation retains the burden of justifying its decision, checking the data, and providing timely remedy when the outcome cannot be sustained.
Failure should also be rehearsed. A test can remove a data source, simulate supplier outage, or introduce a case that does not fit. The exercise reveals whether an alternative exists, whether someone knows how to stop the flow, and whether recovery includes people. Restoring a system does not by itself return a lost break, a denied opportunity, or a damaged reputation.
Measuring better in order to decide with dignity#
The procedure can begin with a simple question: what claim does each figure support, and what claim does it not support? It then identifies who appears, who is missing, and which conduct it encourages. Finally, it connects the finding to a concrete decision, an accountable person, a deadline, and a possible remedy. This sequence turns data into governance and limits the convenience of an aggregate story.
An organisation must accept that some indicators will deteriorate when a route for voice opens. More complaints, incidents, or exceptions can mean greater trust and better detection. Interpretation depends on examining accessibility, repetition, and response together. Over time, a healthy system does not seek silence; it seeks problems reported early, corrected well, and less likely to recur.
Better measurement also means retiring metrics that have lost their purpose. An indicator may be technically stable yet encourage damaging shortcuts or duplicate surveillance. Review should ask whether it still helps a decision and whether its informational cost is proportionate. What is measured from habit eventually governs without an explicit mandate.
The most defensible success cannot fit into a single figure. It appears when service improves without greater risk, savings are shared, exceptions receive attention, and an ordinary person can understand and question the consequence. Dignity supplies a demanding measure: no aggregate advantage makes a serious, avoidable harm to a minority acceptable.
Curiosity encourages the search for the story an average does not tell; responsibility assigns power to act; care prevents the exception from being treated as noise. Together, these dispositions turn measurement into collective learning. Automation deserves to remain when its benefits withstand a distributive view, its operation permits disagreement, and the institution retains genuine capacity to explain, stop, and repair.
