Blog · Review Essay · Last reviewed June 25, 2026

The Algorithm and the Workplace Control System

Hilke Schellmann's The Algorithm is strongest when read as a book about institutional power: workplace AI does not need consciousness to shape a life. It only needs a score, a workflow, a vendor contract, and an employer willing to treat the output as authority.

For this review, workplace AI means automated, algorithmic, or AI-assisted systems that materially shape access to work, pay, scheduling, promotion, monitoring, discipline, accommodation, or termination. The governing question is not whether a model makes the final decision, but whether its score, summary, ranking, alert, or dashboard changes what a human decision-maker can see and justify.

The book's deeper warning is that the workplace becomes a control system when selection, surveillance, direction, and record-making are joined. A hiring score decides who enters; sensors and dashboards define performance; generated summaries become personnel records; and the worker must contest a machine-shaped file before contesting the manager.

The Book

The Algorithm: How AI Decides Who Gets Hired, Monitored, Promoted, and Fired and Why We Need to Fight Back Now was published by Da Capo on January 2, 2024. The publisher lists Hilke Schellmann as the author, gives the book as 336 pages, and identifies ISBN-13 9780306827341. Amazon lists the same title and author on its product page and uses 0306827344 as the hardcover ASIN and ISBN-10.

Schellmann is reporting on a present workplace in which employers buy systems that screen applicants, evaluate speech or video, rank workers, monitor activity, predict attrition, coach performance, and make managerial judgment appear technical. That makes the book a useful companion to Weapons of Math Destruction, The Eye of the Master, The Quantified Worker, and Data-Driven Truckers: the issue is not only bad math, but the social location of the math.

Current Context

As of June 25, 2026, the workplace AI problem has moved from novelty to employment infrastructure. Employers use tools for recruiting, resume parsing, job-ad targeting, assessments, interview scoring, scheduling, productivity monitoring, attrition prediction, call coaching, performance summaries, and manager dashboards. Some tools are marketed as AI; others are ordinary rules, analytics, or workflow software. The important boundary is material influence over work, not the vendor label.

The International Labour Organization's definition of algorithmic management is useful here because it is functional rather than mystical: algorithmic systems use tracked data and other information to organize, assign, monitor, supervise, and evaluate work, and they can be AI-based or rules-based. Schellmann's book belongs in that frame. The workplace control system is not one product category; it is the stack of selection, sensing, scoring, scheduling, documentation, discipline, and appeal.

Current law and policy now recognize that boundary, but unevenly. Federal U.S. civil-rights agencies have said automated systems remain subject to existing civil-rights, consumer-protection, fair-competition, and equal-opportunity laws. The EEOC's AI materials address disability discrimination, adverse impact, worker guidance, and automated systems. The Department of Labor's 2024 roadmap is nonbinding guidance, but it names a worker-centered baseline: governance structures, meaningful human oversight for significant employment decisions, transparency, worker input, labor-rights protection, training, and worker-data security.

The enforcement lesson is concrete. In 2023, the EEOC announced a $365,000 settlement with iTutorGroup after alleging that online application software automatically rejected female applicants age 55 or older and male applicants age 60 or older. The case matters here because the alleged discrimination did not require a sentient system or a science-fiction machine. It required a rule embedded in hiring software, an employer using it, and applicants who were screened out before ordinary judgment could begin.

The state and local map is fragmented. New York City's Local Law 144 covers certain automated employment decision tools and requires a recent bias audit, public audit information, and notice, while the New York State Comptroller's 2025 audit found enforcement and complaint-routing gaps. California's employment regulations on automated-decision systems took effect on October 1, 2025. Illinois Public Act 103-0804 took effect on January 1, 2026 and makes discriminatory AI use, and failure to provide required notice for covered AI use, civil-rights issues. Colorado's SB26-189 was signed on May 14, 2026 and creates duties for automated decision-making technology used to materially influence consequential decisions, including employment, beginning January 1, 2027.

In the European Union, the AI Act treats many employment, worker-management, recruitment, task-allocation, and performance-monitoring systems as high-risk. Article 5 also prohibits workplace emotion-inference AI except for medical or safety reasons, and Article 26 requires deployers of high-risk systems to inform worker representatives and affected workers before use. The Commission's implementation page now says rules for high-risk systems in areas including employment will apply from December 2, 2027 after the May 7, 2026 political agreement on the AI omnibus; prohibited practices and AI-literacy duties already entered application on February 2, 2025. That timing matters because employers are deploying systems before the full high-risk compliance machinery is in force.

Hiring as Interface

The book's sharpest chapters treat hiring as an interface problem. A job seeker thinks she is submitting a resume, recording an interview, taking an assessment, or playing a workplace game. The employer sees a structured stream of features, scores, flags, and rankings. Between those two worlds sits a vendor product that converts a person into a set of machine-readable traces.

This is where Schellmann's reporting matters. She shows that workplace AI can be consequential even when it is mundane. A system does not need to be general, autonomous, or conscious to reshape opportunity. It only has to sit upstream of a human conversation, decide who gets advanced, and make rejection feel like the neutral result of a process no one in the room can fully explain.

That is the practical form of algorithmic belief. The score becomes persuasive because it arrives with the aura of scale: many applicants, many data points, many comparisons. But scale is not evidence. If the measured construct is weak, if training data reflect past discrimination, or if the output has not been validated for the actual job, the interface can launder guesswork into procedure.

The sharper test is construct validity under power. What is the tool claiming to measure: job skill, personality, communication style, culture fit, future performance, rule compliance, attrition risk, or something else? Who chose that construct? What evidence shows it matters for the job? Which groups bear false negatives? What accommodation path exists for disabled applicants? Without those answers, the applicant is not being evaluated by a neutral interface. They are being sorted by an unexamined theory of work.

A bias audit label does not answer those questions by itself. An audit may measure selection-rate differences for a defined tool, population, and period while leaving the job analysis, accommodation design, vendor data, scoring construct, and human-review workflow mostly untouched. Hiring accountability requires the employer to connect the audit to the actual job and to the actual decision point where someone is rejected, ranked, or routed away.

The Workplace Becomes a Sensor

The workplace side of the book is darker because it follows the logic after hiring. Once employment is treated as a continuous data environment, the worker becomes easier to measure than to hear. Activity logs, productivity dashboards, communication analytics, location traces, and attrition predictions can all become management inputs. Some may help allocate work or catch problems. Others turn ordinary labor into a permanent audition.

That is why The Algorithm belongs in a labor archive, not only an AI archive. The tools Schellmann investigates shift discretion toward buyers, vendors, and dashboards. They also shift risk downward. A worker can be harmed by a false signal without knowing the source, the evidence behind it, or the channel for correction. The employer may call the system advisory, but the worker experiences the advice as pressure, suspicion, or denial.

This is the same lesson that runs through Ghost Work: automation often changes who must absorb ambiguity. In hiring and management, the ambiguity lands on the person being evaluated. They must perform for systems they cannot inspect, optimize for criteria they cannot see, and contest decisions that may be distributed across policy, software, vendor contract, and managerial habit.

The recurring site theme is a feedback loop, not a takeover story. A dashboard defines performance, workers adapt to the dashboard, adapted behavior becomes new evidence, and management treats the new evidence as proof that the dashboard sees reality. That is why employment AI belongs beside algorithmic management and The Tyranny of Metrics: a proxy becomes dangerous when an institution rewards the proxy and then forgets it was a proxy.

Generative AI adds a quieter sensor: the record-maker. Meeting summaries, call transcripts, ticket notes, coaching suggestions, code reviews, and manager drafts can become evidence about diligence, tone, judgment, or future employability. If the worker cannot inspect and correct the generated record, the system has not merely helped management write faster. It has changed the personnel file.

That makes source separation a safety requirement. A personnel record should distinguish raw observation, worker statement, manager judgment, vendor score, model-generated summary, and final employment action. When those layers collapse into one fluent paragraph, the worker is forced to dispute a narrative without knowing which part came from the job, the manager, the model, or the dashboard.

The Governance Reading

Read in 2026, the book is also a map of a regulatory problem that has become explicit. The EEOC's publications page groups artificial-intelligence materials under employment discrimination, including adverse-impact, disability, worker, and automated-systems resources. The EEOC, DOJ, CFPB, and FTC joint statement says automated systems remain subject to existing civil-rights, consumer-protection, fair-competition, and equal-opportunity laws. That matters because an employer cannot make a discriminatory screen lawful by buying it from a vendor.

The policy surface is no longer only federal guidance. New York City's AEDT page says covered employers and employment agencies may not use an automated employment decision tool unless it has had a bias audit within one year, public audit information is available, and required notices have been provided. A 2025 New York State Comptroller audit then showed why notice and audit rules still need enforcement capacity: complaint routing was ineffective, outreach had stalled, and the auditors found more potential noncompliance than the city review had identified.

NIST's AI Risk Management Framework gives the broader risk-management grammar: govern, map, measure, and manage, carried across the system lifecycle rather than treated as a one-time checklist. The U.S. Department of Labor's 2024 AI best-practices roadmap adds the labor frame: meaningful human oversight for significant employment decisions, transparency to workers, worker input, protection of labor and employment rights, training, and worker-data security. The European Commission's AI Act page identifies employment, worker management, and access to self-employment as high-risk use cases, with rules for certain high-risk areas including employment set to apply from December 2, 2027.

Those sources do not settle the problem. They do clarify its shape. Workplace AI is a rights, evidence, procurement, and accountability problem. The question is not "can the model predict something?" It is "what decision will this prediction influence, what proof makes it job-related, what groups bear the error, what records are kept, and what recourse exists when the system is wrong?"

That makes procurement part of labor governance. An employer buying a screening tool, productivity dashboard, or performance-summary system should require validation evidence, accessibility evidence, subgroup testing, update notices, audit cooperation, data-use limits, appeal support, retention rules, and a right to suspend use. A vendor cannot become the hidden employer while the legal employer points to the contract.

The minimum governance file should name the employment decision, system owner, vendor, model or scoring version, data sources, measured construct, validation evidence, adverse-impact testing, accessibility review, worker notice, human-review role, appeal path, retention rule, update trigger, and stop condition. That file should live in an AI system inventory and procurement record, not in a slide deck. The employer remains responsible for the work system even when the scoring function arrives as a cloud service.

The file should also include a worker-facing receipt for consequential use: what system materially influenced the result, what data category mattered, what human reviewed it, what record can be corrected, what deadline applies, and who has authority to change the outcome. Without that receipt, recourse is mostly a demand that the worker guess which machine-shaped claim harmed them.

The Evidence Test

The book is most useful when it turns AI governance into an evidentiary discipline. A serious employer should be able to name the decision point, the affected population, the job analysis, the measured construct, the validation study, the subgroup error patterns, the accessibility and accommodation process, the data-retention rule, the vendor's update history, the audit scope, and the person with authority to pause use. Without those records, the system is not merely under-documented. It is asking applicants and workers to accept a classification whose basis they cannot inspect.

Source discipline matters because workplace AI fails through handoffs. A vendor claims a tool is validated; HR treats the claim as procurement evidence; a manager treats a score as neutral advice; a worker experiences the advice as a lost interview, a worse schedule, a discipline file, or a termination. The paper trail has to separate marketing language from primary evidence: validation reports, adverse-impact tables, model and prompt versions, notice text, appeal outcomes, incident logs, and records of human overrides. Otherwise "human in the loop" becomes a signature after the system has already shaped the choice.

The safety implication is practical. For low-stakes triage, a cautious tool may be acceptable if it is monitored and easy to override. For hiring, promotion, discipline, scheduling, disability accommodation, pay, or termination, the standard should be higher: documented job relevance, representative testing, accessible alternatives, meaningful notice, worker or representative input, independent audit access, logs fit for later review, and a remedy path that can change the outcome.

For generative systems, the evidence test also has to separate assistance from assessment. A tool used to help workers draft, summarize, coach, code, translate, or search records should not quietly become a performance-scoring system without a new review. If prompts, meeting summaries, tickets, call transcripts, or manager notes become personnel evidence, the employer needs a source trail and retention rule before the generated wording hardens into record.

The human-review test is whether the reviewer can change the result. A manager needs time, source access, accommodation information, authority to override, and protection from retaliation or productivity pressure when rejecting the tool. If the interface hides the basis for the score or makes override exceptional, human review is theater. If appeal outcomes never feed back into procurement, validation, and model retirement, recourse is only customer service.

Safety also means checking the work system, not only the model. A scheduling model can create fatigue. A productivity dashboard can make lawful breaks look like deviance. A call-coaching tool can penalize accent, disability, or care work. A generated performance note can overwrite context. The relevant evidence therefore includes health and safety reports, accommodation outcomes, complaint records, override logs, appeal reversals, and worker or representative consultation, not only benchmark accuracy.

Where the Book Needs Care

Schellmann's method is investigative and case-driven. That is the book's strength: it gives readers concrete encounters with systems that otherwise hide behind marketing language. The limitation is that exposure alone can make the answer look simpler than it is. Bad products should be named, but the deeper pattern is a market in which employers want cheap certainty, vendors sell confidence, and workers have little power to demand evidence before the system acts on them.

The reviewer's caution is that audits can become their own ritual. A bias audit without access to the right data, a notice without meaningful explanation, a complaint process people do not know how to use, or a human review that simply ratifies the machine can preserve the same power relation in cleaner paperwork. The Algorithm is most useful when it pushes readers beyond "which tool failed?" toward "why was this tool allowed to mediate the employment relationship at all?"

The book should also not be read as a ban on every workplace tool. Scheduling software, accessibility tools, safety alerts, translation, training support, and administrative automation can help workers when they are narrow, transparent, contestable, and governed with worker participation. The line is crossed when assistance becomes surveillance, when measurement becomes discipline, or when convenience becomes a reason to weaken rights.

What This Changes

The practical reading is direct. Before adopting an employment AI system, ask who selected it, what job-related evidence supports it, what population it was tested on, how error differs across groups, whether disabled applicants and workers can request accommodation, how long data is retained, whether the vendor can be audited, and what path exists for appeal. If those questions cannot be answered, the system is not ready to decide anything important.

Schellmann's contribution is to keep the analysis grounded in work. AI governance often drifts toward abstract capability debate. The Algorithm returns the issue to the applicant waiting for an answer, the worker watched by a dashboard, the union or works council asking for notice before deployment, and the manager tempted to confuse a score with judgment. The book's central lesson is not that machines are taking over. It is that institutions keep building machines into places where accountability was already weak.

That is the bridge to AI in Employment, Algorithmic Management, Algorithmic Recourse, and Vendor and Platform Governance. The work system does not become accountable because a tool is accurate in a demo or audited once. It becomes accountable when workers can see the system, challenge the record, receive accommodation, preserve evidence, and reach a human with authority to change the outcome.

Source Discipline

This review treats Schellmann's book as investigative reporting about workplace AI practices, not as a legal treatise. Publisher and author pages are used for bibliographic claims. EEOC, DOJ, DOL, NIST, ILO, NYC, California, Illinois, Colorado, and European Commission materials are used for current legal and policy context. A New York State Comptroller audit is used as evidence about enforcement capacity, not as proof that every AEDT user violates the law.

Legal claims are dated because this field is moving. "Bias audit," "notice," "human review," "high-risk," "employment AI," and "automated decision-making technology" do not mean the same thing across the EEOC, NYC Local Law 144, California civil-rights regulations, Illinois Public Act 103-0804, Colorado SB26-189, the EU AI Act, and voluntary federal guidance.

For deployment claims, vendor marketing is not enough. The stronger evidence is operational: job analysis, validation studies, subgroup error analysis, accessibility testing, notices, worker-consultation records, model or system versions, data-retention policies, appeal logs, override records, incident reports, and contract terms that let the employer audit, pause, or exit the system. This review does not claim that any AI system is conscious, divine, or AGI; it treats workplace AI as delegated managerial power that must remain contestable.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books