Management | Human Resources | Artificial Intelligence

Algorithms now recruit candidates, allocate shifts, set targets, score performance and restrict access to work. The decisive management question is no longer whether a person appears somewhere in the process. It is whether that person has the knowledge, time and authority to prevent an automated judgment from becoming an organizational wrong.

By Frank Farnel | Responsible Public Affairs | September 18, 2026

Executive Summary

• Algorithmic management is already mainstream. An OECD survey of more than 6,000 managers found adoption rates of 90% in the United States and an average of 79% across the European countries surveyed. Yet nearly two-thirds of managers using these tools reported at least one concern about their impact on workers.
• The phrase “human in the loop” can describe anything from genuine managerial judgment to a rushed click on an automatically generated recommendation. Presence is not oversight. Effective oversight requires context, competence, capacity, contestability and correction.
• Four cases reveal the consequences of weak and strong governance: Uber’s contested €825 million Dutch data-protection penalty; iTutorGroup’s age-discrimination settlement; California’s warehouse-quota citations against Amazon; and the negotiated controls adopted at Deutsche Telekom and IBM Germany.
• European obligations are arriving on different timetables. Member States must transpose the Platform Work Directive by December 2, 2026. The AI Act’s detailed high-risk rules for employment systems listed in Annex III are now scheduled to apply from December 2, 2027. Existing data-protection, discrimination, labor and health-and-safety duties remain relevant in the meantime.
• Boards and executives should govern the management decision, not merely procure the software. A vendor contract cannot transfer responsibility for the organization’s purpose, data, thresholds, staffing or response when a decision harms a worker.

The Manager Has Not Disappeared. Management Has Moved Into the System.

For years, workplace technology was treated as administrative infrastructure. Payroll calculated. Scheduling software arranged. Applicant-tracking systems stored. The distinction between the tool and the manager seemed reasonably clear.

That distinction is no longer reliable. An automated system may rank candidates before a recruiter sees them, recommend who receives overtime, infer that an employee is underperforming, calculate an individualized target, trigger a warning or restrict a platform worker’s access to income. The manager may encounter only the output. In some organizations, the manager’s role has quietly become the confirmation of a conclusion produced elsewhere.

The scale matters. The OECD defines algorithmic management as the use of technological tools to fully or partly automate functions traditionally performed by managers: instructing, monitoring and evaluating workers. Its survey of more than 6,000 mid-level managers in France, Germany, Italy, Japan, Spain and the United States found the tools already widely adopted. Managers often see benefits; 60% reported improved decision quality in the OECD’s 2026 policy review. But the same evidence records a serious trust problem. Twenty-eight percent identified unclear accountability when a decision is wrong, 27% difficulty understanding algorithmic recommendations, and 27% inadequate protection of physical or mental health.^[1]

This is not an argument against automation. Consistent scheduling can reduce favoritism. Early warnings can reveal unsafe workloads. Properly validated tools may help recruiters examine more applications and help managers identify training needs. The mistake is to confuse computational consistency with organizational fairness. A rule can be applied consistently and still be irrelevant, discriminatory, unsafe or simply wrong.

Analysis: The central governance risk is not that software replaces every manager. It is that software defines the evidence, frames the options and sets the tempo within which managers act. A nominally human decision can therefore remain functionally automated.

The Five Conditions of Meaningful Human Oversight

The European AI Act provides a useful starting principle. For high-risk systems, oversight must enable people to understand capabilities and limitations, recognize automation bias, interpret outputs, disregard or reverse them, and stop the system when necessary. Deployers must assign oversight to people with the necessary competence, training, authority and support.^[2] Those provisions describe more than compliance mechanics. They describe what responsible management should look like.

For executives, the idea can be translated into a five-part test.

1. Context: Can the reviewer see what the system cannot?

An output is produced from selected data. A manager must be able to examine information outside that selection: a disability accommodation, equipment failure, unusual customer behavior, a change in territory, missing records, a disputed rating or a legitimate reason for interrupted work. If the reviewer sees only the score and the score’s supporting variables, the system has defined the entire reality of the decision.

2. Competence: Does the reviewer understand the tool well enough to disagree?

Training cannot stop at navigation. Reviewers need to understand the purpose, known limitations, relevant error rates, data provenance, affected groups and circumstances in which the output should not be trusted. A manager who knows how to open a dashboard but not how its recommendation can fail is an operator, not an overseer.

3. Capacity: Is there enough time and staffing for judgment?

A review right is fictitious when one person must clear hundreds of alerts per hour, when disagreement requires several screens of justification, or when the system’s recommendation takes effect before the human sees it. Oversight has an operating cost. Organizations unwilling to fund it have chosen automation, whatever their policy says.

4. Contestability: Can the affected person reach someone with real authority?

A generic help desk, chatbot or appeal form is not necessarily human review. The worker or candidate needs a comprehensible reason, a route to present contrary facts and a reviewer who can change the outcome. The EU Platform Work Directive makes this unusually concrete: it requires a human being to take decisions that restrict, suspend or terminate a platform-work relationship or account, and gives people a right to an explanation and a substantiated review.^[3]

5. Correction: Does one bad decision improve the system?

Reversing an individual error is necessary but insufficient. Organizations must look for patterns. Which groups appeal most often? Which managers override the system? Are reversals clustered around a location, job or data source? A complaint is also diagnostic information. If it ends with the individual remedy and never reaches the model owner or process designer, the organization will reproduce the same error at scale.

Case One: Uber and the Difference Between Review and Authority

On August 21, 2026, the Dutch Data Protection Authority confirmed a penalty of €824.99 million against Uber over automated suspensions and deactivations of driver accounts during the period examined, 2018–2022. The regulator concluded that drivers had not been adequately informed and that significant decisions had occurred without sufficient human involvement. The case originated in complaints involving European drivers and was handled in the Netherlands because Uber’s European headquarters are located there.^[4]

Uber disputes the decision and has announced an appeal. The company says its policies provide human review and opportunities to contest suspensions, and argues that the penalty is disproportionate. Those points matter: the decision is not final, and an appeal may alter either the legal findings or the amount.

The governance lesson does not depend on treating every contested allegation as established. The case forces a precise question: what counts as human involvement when access to work is at stake? A reviewer who enters after the account is blocked, sees only the system’s file and lacks practical authority to restore access may be performing customer service rather than management oversight.

The Platform Work Directive, which Member States must transpose by December 2, 2026, points toward a demanding answer. It requires sufficient human resources, competent and trained reviewers with authority to override decisions, regular evaluation of impacts, involvement of worker representatives and a right to a reasoned reply. Its protections are specific to platform work, but its managerial logic travels further. Any organization using automated discipline should know who can halt the process before harm, who can repair it afterward and who owns the pattern revealed by repeated appeals.

Case Two: iTutorGroup and the Efficient Execution of a Discriminatory Rule

The simplest algorithmic failures are sometimes the most revealing. In 2022, the U.S. Equal Employment Opportunity Commission alleged that iTutorGroup had programmed recruitment software to reject female applicants aged 55 or older and male applicants aged 60 or older. More than 200 qualified applicants were rejected, according to the agency.^[5]

The litigation settled in 2023 for $365,000, without a trial on the merits. The consent decree also required anti-discrimination measures and, should U.S. hiring resume, training and monitoring obligations.^[6]

This was not a mysterious machine-learning model that discovered an obscure proxy in millions of data points. According to the EEOC, the age thresholds were programmed. The system did precisely what it had been configured to do. That is why “the algorithm made the decision” is rarely a useful managerial explanation. Someone defined the rule, approved the data field, selected the threshold, accepted the workflow and failed to detect the outcomes.

Human review at the end would have helped only if reviewers saw the rejected population rather than merely the applicants who passed. This is a recurring design flaw. Organizations audit the people selected by a system while the excluded candidates, shifts, assignments or promotions become invisible. Responsible oversight therefore examines the decision funnel, including who disappears before a human becomes involved.

Case Three: Amazon and the Management Target That Became a Legal Disclosure

California’s Warehouse Quotas Law does not prohibit productivity targets. It requires employers to provide workers with written descriptions of applicable quotas and bars quotas that prevent lawful meal or rest periods, bathroom use or compliance with occupational health and safety law.

In June 2024, the California Labor Commissioner cited Amazon nearly $5.9 million for 59,017 alleged violations at two warehouses. The agency said Amazon had failed to provide written notice of quotas. Amazon disputed the citations, arguing that it did not use fixed quotas at the facilities and that performance was evaluated over time in relation to site-level expectations.^[7]

The dispute illustrates why definitions matter. A company may regard a dynamic benchmark, comparative rate or continuously updated expectation as something other than a quota. A worker may still experience it as a target that shapes pace, breaks, evaluation and job security. Management systems acquire power through effects, not labels.

This is also where health and safety enter algorithmic governance. The EU Platform Work Directive explicitly recognizes that algorithmic direction, evaluation and discipline can increase monitoring, accelerate work and heighten stress and anxiety. An efficiency metric is therefore not merely an operational number. It may be an input into workload, psychosocial risk and employee behavior. HR, operations and occupational safety cannot govern it in separate rooms.

Case Four: Deutsche Telekom and IBM Germany Built Governance Before the Complaint

The most instructive examples are not enforcement actions. A 2025 International Labour Organization working paper documents negotiated systems at Deutsche Telekom and IBM Germany that place worker voice inside technology governance.^[8]

At Deutsche Telekom, works agreements described by the ILO require consultation before new technology is purchased. A joint committee evaluates risks, including whether software can track individual employees. The workforce-analytics agreement prohibits automated decisions without human oversight and the use of analytics to monitor individual performance or behavior; data are generally aggregated so that conclusions cannot be drawn about individuals. An expert group with equal labor-management representation reviews the use of employee data and analytics tools.

IBM Germany’s 2020 AI framework classifies applications by risk. According to the ILO case study, AI used for personnel measures or immediate decisions affecting employees without human oversight is treated as high risk and prohibited. Lower-risk applications, such as training recommendations visible to the employee and supervisor, are permitted. An AI Ethics Council of experts and employer and employee representatives evaluates applications, hears objections and can correct recommendations.

These arrangements do not prove that every outcome is fair or that collective governance is frictionless. They do demonstrate a stronger operating model. The organization does not wait for a worker to identify a hidden harm. It creates a venue where purpose, data access, risk classification, permissible uses and dispute resolution can be negotiated before deployment.

The ILO researchers report that works council representatives believed these agreements improved health and safety, trust and acceptance while reducing opposition to expensive systems. That finding should interest executives who treat consultation as delay. Early scrutiny can expose a poor use case before procurement costs, employee resistance and legal risk are locked in.

A Decision Matrix for Algorithmic Management

Decision type	Illustrative use	Minimum governance	Human authority required	Red flag
Low-consequence assistance	Suggesting training content or drafting a routine schedule	Purpose definition, data controls, periodic accuracy check	User can ignore or change output without penalty	Recommendation quietly becomes a performance expectation
Resource allocation	Assigning shifts, routes, leads, overtime or cases	Impact testing, exception rules, worker notice, outcome monitoring	Manager can alter allocation and document context	Override is technically possible but operationally punished
Performance evaluation	Scoring productivity, quality, behavior or potential	Validated job relevance, bias analysis, health-and-safety review, representative consultation	Trained reviewer sees source data and contrary evidence	Only the score is visible; underlying evidence cannot be challenged
High-consequence decision	Rejecting a candidate, denying pay, suspending, demoting or terminating	Pre-decision human review, written reasons, appeal, logs, senior accountability	Named person has time and power to stop, reverse and remedy	Human contact occurs only after the outcome takes effect
Prohibited or unacceptable use	Workplace emotion inference, covert sensitive-trait inference or retaliation prediction	Do not deploy; legal review and technical blocking	Executive owner terminates the use case	Vendor relabels the tool without changing its function

The Regulatory Calendar Is Not a Management Calendar

The European legal timetable has become more complex. The consolidated AI Act published after the July 2026 amendments states that the detailed requirements for Annex III high-risk systems—including AI used in recruitment, promotion, termination, task allocation and performance monitoring—will apply from December 2, 2027. The Act nevertheless already contains applicable prohibitions, including the general prohibition on workplace emotion recognition except for limited medical or safety purposes, and it sits alongside the GDPR, national labor law, discrimination law and collective-bargaining obligations.

The Platform Work Directive operates on a nearer horizon. Member States must transpose it by December 2, 2026. Its algorithmic-management chapter is limited to digital labor platforms, yet it establishes a visible European benchmark: disclose the systems, restrict certain data uses, evaluate impacts, staff oversight adequately, require human decisions for serious adverse actions and provide explanations and review.

Meanwhile, on July 20, 2026, the European Commission opened a second-stage social-partner consultation on possible EU action concerning fair telework and the right to disconnect. The consultation explicitly includes algorithmic management in the broader workplace.^[9] The direction of travel is therefore broader than platform work, even if the eventual form and scope of any initiative remain undecided.

Waiting for each final rule would produce a fragmented operating model: one control for candidates, another for employees, a third for contractors and a fourth for platform workers. The better approach is a common decision standard based on consequence. If an automated output can materially affect income, opportunity, health, reputation or continued employment, it deserves meaningful review regardless of the person’s contractual category.

What Leaders Should Do Now

Inventory decisions, not products

A software inventory is necessary but not sufficient. Map every consequential people decision and identify where automation supplies evidence, ranking, recommendations, thresholds or triggers. One platform may contain dozens of distinct use cases with different risks.

Name an accountable business owner

Procurement may buy the tool; IT may integrate it; HR may administer it. None of those facts answers who is responsible for its impact. Each high-consequence use requires a named executive owner who can suspend the process, fund remediation and report to the board.

Design the override before deployment

Ask practical questions. Which screen shows the evidence? Who can reverse the result? Does reversal require permission from the model owner? What happens to pay or access while the appeal is pending? How is the worker notified? A principle that cannot be executed in the workflow is not a control.

Measure the reviewers

Track volume, review time, agreement rates, reversals, appeal outcomes and differences across demographic groups, locations and managers. A 99.9% agreement rate may show excellent accuracy—or automation bias and a powerless review function. Metrics need investigation, not automatic celebration.

Contract for evidence and change

Vendor terms should provide access to documentation, performance metrics, change notices, logs, audit cooperation, incident reporting and the ability to suspend a feature. The organization also needs clarity on who bears the cost of remediation and how rapidly a harmful rule can be changed.

Consult the people being managed

The OECD reports that consultation is associated with better outcomes, while the ILO cases show how structured worker voice can turn resistance into governance.^[10] Employees know where data are incomplete, where targets collide with safety and where customers or local conditions distort a metric. Their experience is risk intelligence.

Give the board an impact view

Boards do not need model-level detail for every application. They do need a view of consequential uses, exposed populations, material incidents, appeal patterns, significant vendor dependencies and whether management can stop the system. Algorithmic management belongs within workforce, legal, technology and operational-risk oversight—not in an innovation appendix.

Conclusion: Authority Cannot Be Automated Away

Organizations will continue to automate managerial work because the benefits are real. Software can handle complexity, detect patterns and make operations more consistent. It can also accelerate a bad assumption, obscure responsibility and turn a local error into a system-wide practice.

The answer is not to insist that every decision remain manual. Nor is it to place a person at the end of an automated pipeline and call the process human. Meaningful oversight is an institutional capability. The reviewer must understand the context, possess relevant competence, have enough capacity, hear a challenge and correct both the decision and the system.

The manager is still accountable when the recommendation comes from software. The board is still accountable when the software comes from a vendor. And the company is still accountable when efficiency makes the decision difficult to see. A human in the loop is a diagram. Human authority is a management system.

Key Evidence

• 90% in the United States and 79% across the surveyed European countries: reported adoption rates for at least one algorithmic-management tool in the OECD’s survey of more than 6,000 managers. OECD, 2025.
• Nearly two-thirds: share of managers using such tools who reported at least one concern about their impact on workers; unclear accountability was the most frequent individual concern. OECD, December 19, 2025.
• €824.99 million: Dutch regulator’s August 2026 penalty against Uber concerning automated account suspensions and deactivations. Uber disputes the decision and is appealing. Reuters, August 21, 2026.
• More than 200 applicants and $365,000: applicants the EEOC said were rejected because of programmed age thresholds, and the later settlement amount. EEOC, September 11, 2023.
• 59,017 alleged violations and nearly $5.9 million: California Labor Commissioner citations concerning warehouse-quota disclosure at two Amazon facilities. Amazon disputed the citations. California DIR, June 18, 2024.
• December 2, 2026 / December 2, 2027: deadline for Member States to transpose the Platform Work Directive; application date for the AI Act’s detailed Annex III high-risk requirements after the 2026 amendments. Directive (EU) 2024/2831; AI Act consolidated text.

Glossary

Algorithmic managementUse of software to instruct, monitor, evaluate or otherwise coordinate workers and work-related decisions.Automated decision-making systemA system that takes or materially supports decisions, often by processing personal or behavioral data.Automation biasThe tendency to accept a machine-generated output too readily, even when contrary evidence is available.DeployerUnder the EU AI Act, an organization or person using an AI system under its authority in a professional context, subject to specified exceptions.Human oversightOperational arrangements enabling competent people to monitor, interpret, disregard, reverse or stop an automated system.ContestabilityThe ability of an affected person to understand, challenge and obtain meaningful review of a decision.Works councilAn employee-representation body with information, consultation or co-determination rights under national law.

References and Further Reading

European and International Sources

1. European Union. Regulation (EU) 2024/1689—Artificial Intelligence Act, consolidated text. Consolidated July 27, 2026.
2. European Parliament and Council. Directive (EU) 2024/2831 on Improving Working Conditions in Platform Work. October 23, 2024.
3. European Commission. Commission Launches Second-Stage Consultation of Social Partners on Fair Telework and the Right to Disconnect. July 20, 2026.
4. OECD. Algorithmic Management in the Workplace: New Evidence From an OECD Employer Survey. OECD Artificial Intelligence Papers, February 6, 2025.
5. OECD. How Widespread Is Algorithmic Management in Workplaces? December 19, 2025.
6. OECD. Recent Policy Developments on AI in the Labour Market. July 2026.
7. Virginia Doellgast, Shruti Appalla, Dina Ginzburg, Jeonghun Kim and Wen Li Thian. Global Case Studies of Social Dialogue on AI and Algorithmic Management. ILO Working Paper 144, International Labour Organization, 2025.
8. Janine Berg and Hannah Johnston. AI in Human Resource Management: The Limits of Empiricism. ILO Working Paper 154, International Labour Organization, 2025.
9. OECD. Exploring Win-Win Outcomes of Algorithmic Management. 2025.

Official Enforcement and Case Sources

10. U.S. Equal Employment Opportunity Commission. EEOC Sues iTutorGroup for Age Discrimination. May 5, 2022.
11. U.S. Equal Employment Opportunity Commission. iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit. September 11, 2023.
12. California Department of Industrial Relations. Labor Commissioner Cites Amazon Nearly $6 Million for Violating California’s Warehouse Quotas Law. June 18, 2024.
13. Kate Abnett and Toby Sterling. Dutch Regulator Fines Uber $966 Million for Automating Driver Suspensions. Reuters, August 21, 2026.

Source and Methodology Note

Research was completed on September 18, 2026. The article prioritizes enacted European legislation, official enforcement releases, OECD and ILO research, and Reuters reporting for the recent Dutch regulatory decision whose complete decision was not readily accessible in English through the regulator’s public site. Enforcement allegations and contested findings are not presented as final judicial determinations. Uber has announced an appeal of the Dutch penalty. Amazon disputed the California citations. The iTutorGroup matter ended in a consent decree rather than a merits judgment. The ILO’s Deutsche Telekom and IBM findings are qualitative case studies based on agreements, interviews and documentary research; statements about perceived benefits are attributed to the interviewed works council representatives. The OECD adoption figures reflect a 2024 survey of managers in six countries and should not be generalized mechanically to every sector or economy. The five-condition framework and conclusions identified as analysis are the author’s synthesis.

Suggested Internal Links

• Ethics and Compliance — culture, responsibility and organizational controls.
• Issues Management — early detection, stakeholder intelligence and escalation.
• Expertise — Responsible Public Affairs frameworks for ethical decision-making, stakeholder trust and governance.

#AlgorithmicManagement #ResponsibleAI #FutureOfWork