A customer-service team began diverting difficult cases just before the monthly close. Nobody had ordered this: the dashboard rewarded completed files and recorded neither complexity, quality nor problems actually solved. Looking at how a workplace metric operates means following the whole path — data, calculation, decision and consequence — before calling a number an objective.
The cost of not looking seldom arrives as a spectacular breakdown. It is distributed across rushed calls, unreported injuries, complex customers nobody wants to help, delayed breaks and people who learn to produce the right signal even when service deteriorates. The useful question is not whether the algorithm is intelligent. It is which behaviour the system makes rational, which work it renders invisible and who can correct it when the result conflicts with reality.
This analysis applies to a simple spreadsheet score as well as to a predictive model supplied by a vendor. Automation increases the scale, but the problem starts earlier: a partial measure becomes a target, the target becomes a ranking, and the ranking shapes schedules, incentives, surveillance or continued employment. Understanding that chain enables a concrete discussion without requiring every worker to become a data engineer.
Taking the score apart without knowing the code#
An understandable audit starts with five columns: stated purpose, input data, calculation rule, linked decision and route to review. For inputs, distinguish observations — orders prepared or time recorded — from inferences such as “commitment”, “attrition risk” or “attitude”. Then list missing context: case difficulty, broken equipment, recent training, support work, disability, customer language or authorised interruptions. If materially different situations produce the same number, that number needs context before it is allowed to produce a consequence.
An understandable audit starts with five columns: stated purpose, input data, calculation rule, linked decision and route to review.
The denominator matters as much as the score. Completing twenty straightforward cases is not comparable to completing twenty investigations; connected time is not the same as useful attention; a recorded pause does not show whether it prevented an accident. Test known examples and manually calculate the score they should receive. Discrepancies expose rounding, duplicated records, incomplete periods or rules that penalise people who accept difficult work. The aim is not to obtain a vendor’s trade secret. It is to determine whether the available explanation lets a person identify and correct an error.
Look for adaptation as well. Once a metric becomes a target, people change behaviour: they close and reopen files, select quick tasks, postpone incident reports or stop helping colleagues. Those responses do not necessarily prove individual dishonesty; they may show that the system rewards conduct opposed to its stated purpose. A serious review compares the indicator with independent outcomes — complaints, returns, accidents, sampled quality and the experiences of customers and workers — and examines differences between shifts, sites, contracts and roles.
What an affected worker can do#
For two weeks, keep a factual record: date, task, incident, displayed data, decision received and concrete effect. Do not remove customers’ personal data or unnecessary confidential information. An effective challenge identifies something verifiable: “the system counted the 34 minutes of a reported equipment failure as inactivity” is more actionable than “the algorithm is unfair”. The request should seek correction of the data, an explanation of the relevant factors, review by a person with authority and confirmation that raising the issue will not trigger an automated penalty.
In Spain, workers’ legal representatives have a right to information about the parameters, rules and instructions on which algorithms or artificial-intelligence systems are based when they influence decisions capable of affecting working conditions or access to and continuation in employment. This appears in Article 64.4(d) of the consolidated Workers’ Statute published by the BOE. Its application depends on the case: it does not make every formula public or replace union and legal advice, but it prevents “the machine decided” from serving as a sufficient explanation.
If a metric uses location, biometrics, health, communications or digital activity, further data-protection questions arise: necessity, proportionality, prior information, retention, access and correction. The Spanish Data Protection Agency’s guide to data protection in employment helps structure that examination. Not every score meets the narrow conditions governing solely automated decisions, so the organisation should document whether a human reviewer genuinely assesses the case or routinely confirms the output.
The employer’s control duties#
Before linking a score to pay, shifts, promotion or discipline, an employer should appoint an accountable owner, document versions and test false positives and false negatives. The review should involve people who perform the job and worker representatives, because they understand exceptions that a vendor cannot see. It also needs stopping thresholds: rising accidents, complaints, unexplained differences between groups, degraded quality or inability to reconstruct a decision. If nobody can explain why a result changed, the system is not ready to produce a serious consequence.
The European Union Artificial Intelligence Act treats certain AI uses in employment and worker management as high-risk, with duties depending on the specific use and implementation timetable. The General Data Protection Regulation establishes principles including transparency, data minimisation and accuracy. No legal label replaces the operational question: can the organisation reconstruct the data, hear relevant context and change the decision before harm becomes irreversible?
A useful quality exercise selects twenty decisions, including five exceptions, and asks two reviewers to reach an outcome using the available documentation. Measure agreement, time and their ability to spot errors. Then ask affected people whether they understood the criterion and found the review route. Publishing an internal summary — what the metric measures, what it does not measure, which decisions it never makes alone and where to challenge it — reduces rumours and improves the data itself because faults can be reported before they accumulate.
A calm family conversation#
When a workplace score reaches home through anxiety, worse shifts or lost income, a family can help without promising a legal result. Listen first and separate facts from interpretations. Then organise evidence, calculate the household impact and prepare three questions for the meeting. If health, discrimination, discipline or dismissal is at stake, seek occupational-safety, union or professional advice early. The aim is to restore room for action, not turn every evening meal into a permanent investigation.
Looking at how a metric works takes time, but not looking has a price too: mistaken decisions that appear objective, essential work disappearing from the indicator, and people forced to optimise an incomplete representation of their craft. A sound metric supports learning; a bad one demands obedience. The difference is whether it accepts context, challenge and correction. When those three doors exist, measurement can improve work. When they do not, the number ceases to be a tool and begins to govern.
An example of a repairable metric#
Imagine a warehouse that scores only units picked per hour. Ana receives a nearby zone and light products; Samir covers a distant zone, resolves missing inventory and helps a new starter. The dashboard puts Ana in green and Samir in red. The comparison looks precise, but it mixes distance, weight, availability and cooperation with personal speed. A sensible correction is not to add a hundred secret variables. It is to separate productive time from recorded interruptions, group comparable tasks, check picking errors and reserve a visible part of the assessment for safety and assistance to the team.
During a four-week pilot, the employer can compare the old and revised rankings without using either for discipline. It should examine who changes position and why, interview a sample of workers and publish the adjustments. If the new model improves speed but errors, musculoskeletal pain or unreported incidents rise, it has not improved work: it has moved the cost. A responsible metric includes balancing indicators and says which one prevails when productivity and safety conflict. The correct answer must always be safety.
At the review meeting, three questions prevent a circular conversation: “Which specific data produced this result?”, “Which relevant context could the system not record?” and “Who has authority to change the decision and document the change?” The affected person should receive an understandable answer and a date, not a technical download that is impossible to interpret. If the individual case is corrected but the cause remains, open a preventive action: change data capture, adjust the comparison or suspend that consequence while a solution is tested.
Before accepting the next dashboard, workers and management can apply one final test: state its purpose in one sentence; name three things the number does not prove; reconstruct a real case; identify the accountable person; and follow the review channel from beginning to end. If one of those tasks is impossible, the organisation does not yet understand the system well enough. Looking does not mean distrusting all technology. It means requiring the same traceability from a workplace tool that we would expect from any decision capable of changing income, health, time or dignity.
