A company announces software that will assign shifts using artificial intelligence. Management promises fewer mistakes; the supplier presents a flawless dashboard; workers know a reality missing from the demonstration: last-minute changes, family care, journeys without public transport and tasks that only particular people know how to solve. If the system is purchased first and discussed afterwards, that experience arrives too late. Democratising workplace technology means involving those who will live with its decisions while objectives, data, rules and contracts can still be changed.
Democracy does not mean voting on every click or requiring everybody to understand the code. It means distributing the ability to intervene. Workers should know which problem is being addressed, what data will be used, what consequences a score may cause and who can correct it. They also need a collective channel for identifying patterns: one error resembles bad luck; twenty errors concentrated among mothers working reduced hours reveal a rule. Participation uncovers operational knowledge that neither a vendor nor senior management possesses.
The technological decision starts before purchase#
The moment of greatest influence is not launch day but problem definition. “We need artificial intelligence” does not describe a need. “We publish shifts late and make too many involuntary changes” does. The second statement allows comparison between options: reorganising responsibilities, improving the existing system, hiring support or automating one component. Before seeking bids, record the baseline: average notice, changes, errors, correction time, overtime and differences between sites or groups. Without that reference, any dashboard can declare itself successful.
The moment of greatest influence is not launch day but problem definition. “We need artificial intelligence” does not describe a need.
Early consultation asks workers where the process gets stuck and which exceptions matter. The person who opens a site knows when the first bus arrives; customer-service staff know that a long call may indicate quality; safety specialists know which tasks conflict with particular restrictions. These are not anecdotes to be cleaned away before training a system. They are requirements. Turning them into test cases prevents a tool from optimising an imaginary version of work and punishing those who keep the real service functioning.
The purchasing team should translate this knowledge into verifiable conditions. For example: publish schedules fourteen days ahead; never reduce hours automatically; respect rest and stated availability; explain every change; permit review before discipline; never use location outside working time. Every condition needs an owner, indicator and contractual remedy. A marketing promise such as “fairer decisions” is insufficient. Ask what evidence the supplier will provide, what audit access it will permit and what happens if the product fails to meet the requirement.
Participation is not receiving a presentation#
Many organisations call a briefing consultation after the contract, timetable and configuration have already been settled. At that stage, people can only agree or appear hostile to innovation. Authentic participation begins with open alternatives, understandable documentation and time to deliberate. It includes legal worker representatives, process owners, safety and data-protection specialists, and affected people with varied circumstances. Management retains its responsibilities, but it should answer proposals and explain why each significant safeguard is accepted or rejected.
In Spain, Article 64.4(d) of the Workers’ Statute published by the BOE recognises legal worker representatives’ right to information about the parameters, rules and instructions behind algorithms or artificial-intelligence systems influencing decisions capable of affecting working conditions. That information should not become an incomprehensible technical dump. It must allow people to identify purpose, relevant factors, consequences, controls and the scope for human intervention.
The EU Directive 2002/14/EC on informing and consulting employees supplies another useful principle: information and consultation require appropriate content and timing. Applied to automation, that means acting before a choice becomes irreversible. Participation does not replace specific legal duties or professional advice, but it improves the process: it records disagreement, forces a comparison between benefits and harm, and creates institutional memory when managers or suppliers change.
A map of power for understanding the system#
Before testing the tool, draw a page with six questions: who defines the objective, who supplies data, who configures thresholds, who receives the recommendation, who bears the consequence and who can stop it? These are often different actors. Human resources signs the contract; the vendor updates the model; a manager clicks accept; a worker loses a shift. The map prevents each party from claiming to be merely an intermediary. Assign an obligation to explain and an authority to correct to every connection.
Next, describe every decision in ordinary language. Do not write “predictive optimisation module”; write “proposes the weekly schedule”. Replace “productivity analytics” with “compares task times and sends alerts to the supervisor”. Separate recommendation from consequence. A score may be auxiliary information, but it becomes a decision when it determines access to shifts, training, bonuses or dismissal. That clarity helps calibrate safeguards: the greater the impact on income, safety, reputation or continued employment, the stronger testing, explanation and review must be.
Identify people who may be poorly represented: recent recruits, disabled or pregnant workers, part-time staff, carers, speakers of another language, night workers, people with weak connectivity or unconventional careers. The aim is not to create unnecessary sensitive labels, but to devise test situations and hear experience. The average can improve while a small group absorbs all the harm. A participatory committee must examine distribution, not only total productivity.
Minimum data and enforceable boundaries#
Technological democracy includes deciding which information will not be collected. Continuous location, keystrokes, pictures, voice or keyboard activity may be easy to capture and difficult to justify. The Spanish Data Protection Agency’s guide to data protection in employment relationships explains principles including information, necessity and proportionality. For every item, write down its purpose, applicable basis, access, retention and deletion. If one of those questions has no answer, the data should not enter yet.
Boundaries belong in configuration and contract, not merely in a friendly policy. Workplace location is switched off outside working hours; planning data do not quietly become disciplinary evidence; a new purpose triggers a new assessment; the supplier does not retain information to train other products without a valid basis. Workers need a short explanation with examples: what a manager can see, what cannot be seen and for how long. Practical transparency reduces rumours and lets people flag drift before it becomes normal.
Quality and provenance require attention. Historical records contain earlier decisions, not neutral truth. If promotion was offered to one profile for years, learning from the past may reproduce that selection. Participatory review asks what each field represents, who was missing, how records are corrected and whether one variable acts as a proxy for another sensitive one. Removing names is insufficient. Postcode, uninterrupted availability or career gaps can reconstruct inequality without naming it.
Testing with the right to say no#
A responsible pilot is small, temporary and reversible. It keeps the previous procedure as a fallback, produces no automatic penalties and has agreed stopping criteria. Tests cover missing data, similar identities, connection failure, family emergencies, adjustments and cases contradicting the pattern. Record what the system proposed, what the supervisor understood, how long correction took and how the person experienced the process. A supplier’s prepared demonstration cannot replace a trial involving real work.
The European Union Artificial Intelligence Act classifies certain uses of AI in employment, recruitment and worker management as high-risk, with obligations depending on the system, each organisation’s role and the applicable timetable. Beyond the precise legal classification, it offers a sensible discipline: risk management, documentation, data quality, logging, information and human oversight. An employer must establish its own duties with suitable advice; a vendor badge does not replace that assessment.
The right to stop the pilot must be genuine. Pause it when inexplicable decisions, significant harm, persistent differences between groups, unexpected monitoring or an unmanageable volume of corrections appears. Stopping is not failure; it shows that the pilot did its job. Money already spent does not make a defective tool necessary. The contract should support data export, restoration of the earlier process and termination without locking the organisation into exclusive formats or knowledge.
Human oversight with time and authority#
Placing a person after an algorithm does not guarantee control. If that person confirms hundreds of recommendations under pressure, cannot see the factors or fears damaging their own performance score, they function as a stamp. Oversight requires training about limitations, access to context, time to doubt and authority to change outcomes. Measure how many recommendations are changed, why and with what result. A correction rate permanently at zero may indicate perfection; obedience or missing information is much more likely.
An explanation to the affected person should answer five things: what decision was made, which data or factors influenced it, what consequence follows, who reviews it and within what period. “You did not reach the threshold” allows neither record correction nor context. Nor is a generic address that nobody answers meaningful. The channel should acknowledge receipt, protect against retaliation and pause irreversible consequences where reasonable while the case is examined.
Complaints are not noise to suppress; they are design information. Every quarter, a committee with plural representation can group reasons, times, outcomes and affected populations. If many people correct the same availability field or one site produces most alerts, fix the cause rather than individual files alone. Return a summary to workers: what was detected, what changed and when it will be checked. That feedback closes the democratic loop and proves participation has consequences.
Sharing benefits, not only risks#
Automation can save hours, reduce unpleasant tasks or prevent mistakes. Negotiation should ask where that benefit goes. It can become training, lighter workloads, more predictable schedules, better service or time for skilled tasks. When savings flow upwards while workers receive surveillance and a faster pace, resistance is rational. Explicit sharing gives people a reason to contribute knowledge and prevents “efficiency” from meaning simply more output with less room to exercise judgement.
New work must also be counted: cleaning data, resolving false positives, explaining outcomes, maintaining permissions and answering complaints. It often falls on people invisible in the original budget. The productivity balance must include their time alongside errors, turnover, health, learning and quality. A system saving ten planning hours while generating fifteen hours of correction is not efficient. One accelerating orders at the expense of breaks does not improve the whole either; it transfers cost to the body.
What a young worker and family can ask#
Somebody entering a first job cannot negotiate the entire system alone. They can learn concrete questions: how are my shifts allocated, what does this application record, who sees the score, can I correct a record, and is there a representative or confidential channel? Keep the offer, contract, significant messages and necessary screenshots without taking customer information or trade secrets. When a consequence is serious, seek appropriate union, professional or institutional support.
A family helps more by rehearsing conversations than by speaking for the young worker. It can organise a timeline, review an email and separate facts from suspicions. “On 12 January I entered my availability and the system scheduled me outside it; please identify the data used and review the decision” is more effective than “the algorithm is targeting me”. Respect privacy and autonomy too. Preparation means offering vocabulary, calm and company if unfairness occurs, not monitoring a workplace phone.
A ninety-day protocol#
During the first thirty days, define the problem, baseline, power map, affected groups and alternatives. Between days thirty and sixty, negotiate requirements, review data and contract, and design a pilot without irreversible consequences. Between days sixty and ninety, test exceptions, publish instructions, activate appeal and compare results with the baseline. The final choice may be controlled deployment, modification, a longer trial or abandonment. All four answers are legitimate when supported by evidence.
Review continues after deployment. Name an accountable owner, a timetable, harm indicators and an expiry date for authorisation. Any material change to model, data, purpose or consequence reopens consultation. A living record notes version, incidents, human decisions and measures taken. This prevents a limited trial from quietly becoming permanent infrastructure. Democracy is not an inaugural meeting; it is the continuing capacity to know, debate, correct and withdraw.
Better technology because it permits disagreement#
A dignified workplace system does not promise absolute neutrality. It admits limits, leaves records, explains outcomes and learns from those who know the job. Participation does not necessarily slow innovation: it prevents useless purchases, reveals risks before they become crises and improves requirements. Disagreement may remain, because productivity, privacy, safety and autonomy do not always point in the same direction. The democratic response is to make the choice visible, listen to those bearing its consequence and maintain an effective remedy.
The final question is not whether a person or machine rules. It is how power around a decision is organised: who defines success, who may challenge, who receives the benefit and who repairs harm. When these answers are negotiated early, tested with difficult cases and reviewed with evidence, technology becomes a tool. When they hide behind a supplier and a score, it becomes authority without a mandate. Work needs systems capable of calculation; a democratic society also needs people able to say “this must change” and be heard.
