How AI changes tasks, jobs, authority, skill, surveillance, and the distribution of economic gains, and how we can design better forms of human-AI work.
WHO DECIDES?
TASKSDATAPAYRIGHTSAUTHORITY
Opening question
If AI makes a team twice as productive, who receives the benefit?
THE PRODUCTIVITY STORYMore output, better quality, faster service, fewer hazards, and new capabilities.
AI can remove drudgery, spread expertise, and make scarce skills more available.
?
THE POWER STORYHigher profit, fewer workers, tighter monitoring, faster pace, or less professional discretion.
The same technical gain can produce very different working lives.
Technology creates possibilities. Organizations and institutions allocate the gains.
The central claim
AI changes work through capability and control.
CAPABILITYWhat can the combined human-AI system observe, predict, generate, decide, or physically accomplish?
CONTROLWho chooses the goal, owns the data, sets the pace, reviews the output, receives the gain, and can stop the system?
A powerful tool can strengthen worker agency or strengthen managerial control.
Work has several layers
A model rarely replaces a job. It enters a layered work system.
1TASKa bounded activity such as classify, draft, inspect, route, lift, or diagnose
2WORKFLOWthe sequence, handoffs, exceptions, tools, and dependencies that produce an outcome
3JOBa bundle of tasks, responsibilities, relationships, identity, judgment, and pay
4ORGANIZATIONauthority, incentives, staffing, knowledge, culture, procurement, and accountability
5LABOR MARKEToccupations, skills, wages, bargaining, mobility, education, law, and social protection
Five possible outcomes
The same AI capability can push work in different directions.
01AUTOMATEmove a task from a person to software or a machine
02AUGMENThelp a person perform a task faster, better, or more safely
03RECOMBINEredistribute tasks across roles and redesign the workflow
04CREATEmake a new product, service, occupation, or task economically possible
05CONTROLincrease monitoring, standardization, pace, or managerial authority
This is not only a generative-AI story
Workplace AI is a stack of technical families with different forms of power.
FEEDBACKfast and reliable ground truthdelayed, strategic, or unknowable
CONTEXTbounded and repetitiveopen world with rare exceptions
ERROR COSTlow, reversible, easy to detecthigh, hidden, cumulative, irreversible
INTEGRATIONone system and standard processmany handoffs, institutions, and incentives
Substitution and complementarity
AI can compete with a worker on one task while increasing the value of another.
SUBSTITUTIONThe system performs a task previously purchased from labor. Demand for that task can fall, especially when quality is sufficient and integration is cheap.
+
COMPLEMENTARITYThe system makes human judgment, relationships, domain expertise, physical action, or accountability more productive and more valuable.
The employment result depends on task mix, demand, prices, new work, and who reorganizes the workflow.
The workflow is the real product
A model output creates value only when the organization redesigns everything around it.
CAPTUREcollect the right data with consent, quality controls, and context
INFERpredict, generate, rank, detect, optimize, or plan with uncertainty
REVIEWcheck quality, plausibility, policy, exceptions, and affected people
ACTchange a schedule, decision, service, machine, or physical process
RECOVERdetect failure, escalate, reverse harm, repair the process, and learn
Buying a model is not the same as building a reliable human-AI work system.
Algorithmic management
AI can perform parts of management even when it performs none of the worker's core craft.
ORGANIZEbreak work into units, queues, routes, and targets
ASSIGNmatch workers, shifts, customers, locations, and tasks
MONITORcapture time, movement, communication, output, and behavior
EVALUATEscore quality, risk, productivity, compliance, and predicted fit
The system becomes a manager when its output changes opportunity or consequence.
Simple rules, optimization, and dashboards can exercise as much authority as machine learning.
The automated hiring pipeline
A candidate can be filtered by several models before a person ever sees the application.
SOURCEtarget ads and recommend possible candidates
PARSEextract education, experience, skills, and keywords
SCREENapply rules, assessments, rankings, or predicted fit
INTERVIEWschedule, transcribe, summarize, or score responses
SELECTcombine signals into a recommendation or automated cutoff
MONITORcompare later performance with predicted success and retrain
Scheduling and allocation
Optimization can coordinate complex work, but its objective becomes working conditions.
WHAT THE SYSTEM CAN OPTIMIZE
coverage, demand, travel, throughput, and equipment
skills, certifications, availability, and fatigue constraints
service levels, cost, reliability, and response time
WHAT WORKERS EXPERIENCE
predictable or constantly changing schedules
fair or opaque access to desirable shifts and income
safer pacing or intensified productivity targets
flexibility chosen by workers or imposed on them
The objective function is a policy about whose constraints count.
Monitoring and performance
Measurement can reveal hazards and erase the invisible work that makes a job succeed.
DATA ≠ WORKTHE MEASURE IS A PARTIAL VIEW
VISIBLEkeystrokes, location, calls, items, speed, errors, response time
OFTEN HIDDENmentoring, emotional labor, prevention, coordination, judgment, recovery
RISKmetric gaming, pace pressure, stress, privacy loss, false discipline
DESIGN TESTwould a worker agree that the measure represents good work?
Robots and cobots
Physical AI can remove hazards, extend capability, and create new failure modes around people.
542,000industrial robots were installed in 2024, the fourth straight year above half a million
ISOLATED AUTOMATIONrobots perform repetitive or hazardous work inside guarded cells
COLLABORATIVE ROBOTSpeople and robots share a space or sequence under controlled conditions
WEARABLE ROBOTICSexoskeletons and assistive devices change strength, fatigue, and ergonomics
AUTONOMOUS MOBILE SYSTEMSrobots navigate warehouses, hospitals, farms, construction, and public space
The goal is not to put a robot near a person. It is to design a safe joint task.
Agents change the unit of automation
An agent does not only produce an answer. It can pursue a sequence across tools.
01receive a goal and operating constraints
02plan subtasks and choose tools
03read files, query systems, write, calculate, or communicate
04observe results and revise the plan
05request approval, act, document, and stop
Longer task horizons are not the same as reliable workplace autonomy.METR measures a 50% success horizon on software tasks. Workplaces often need much higher reliability, realistic task diversity, security, and accountable recovery.
The ironies of automation
The more routine work automation handles, the harder the remaining human work can become.
MONITOR WITHOUT PRACTICEThe person watches normal operation but is expected to intervene during a rare abnormal event.
LOSE THE LEARNING PATHNovices no longer perform the simpler tasks that once built the knowledge needed for expert judgment.
INHERIT THE EXCEPTIONSThe system handles easy cases and leaves people a queue of ambiguous, emotional, high-risk, or broken cases.
Automation can reduce workload while increasing the skill and support required for what remains.
Skill formation and apprenticeship
AI can spread expert guidance or break the path by which novices become experts.
AI AS A COACHExplain reasoning, surface examples, give feedback, adapt practice, and make expert knowledge available at the moment of need.
↔
AI AS A CRUTCHComplete the formative task, hide uncertainty, reduce retrieval practice, and leave the person accountable for work they no longer understand.
A good work system optimizes both today's output and tomorrow's human capability.
02
What the evidence says
AI productivity is real in some settings, absent or negative in others, and still only one step in understanding wages, job quality, and employment.
An evidence ladder for AI and work
A benchmark win is several steps away from a better job or a stronger economy.
1MODEL CAPABILITYCan the system perform a bounded task under test conditions?
2HUMAN PERFORMANCEDoes a person using it become faster, more accurate, safer, or more capable?
3WORKFLOW PERFORMANCEDoes the full team improve after verification, coordination, and recovery costs?
4ORGANIZATIONAL OUTCOMEDo quality, output, cost, reliability, and customer outcomes improve?
5WORKER OUTCOMEWhat happens to pay, hours, autonomy, learning, health, safety, and security?
6LABOR-MARKET EFFECTHow do hiring, occupations, wages, mobility, inequality, and new work change?
Field evidence · customer support
AI assistance raised output most where workers had less experience.
~14%average increase in issues resolved per hour after access to an AI conversation assistant
SETTINGmore than 5,000 customer-support agents in a real company
MECHANISMthe system recommended responses using patterns from high-performing interactions
DISTRIBUTIONgains were largest among less-skilled and less-experienced workers
LIMITone firm, one workflow, and a tool designed around available conversation data
Experimental evidence · professional writing
For bounded writing tasks, participants finished much faster and produced better-rated work.
40%less time on average, alongside an 18% increase in output quality
DESIGNparticipants were randomly assigned access to ChatGPT
RESULTperformance became less unequal across participants
LIMITshort, self-contained tasks with externally rated outputs, not entire jobs
The jagged frontier
The same professionals improved sharply on some tasks and became less correct on another.
INSIDE THE FRONTIERConsultants completed more tasks, worked about 25% faster, and produced outputs rated more than 40% higher in quality.
AI fit the task
OUTSIDE THE FRONTIEROn a task designed beyond the model's capability, AI users were 19 percentage points less likely to reach the correct answer.
Confidence exceeded competence
Human expertise includes recognizing when the tool is on the wrong side of the frontier.
Field experiment · experienced developers
Experienced developers expected AI to make them faster. In this study, it made them slower.
BEFORE THE TASKSDevelopers predicted AI would reduce completion time by 24%.
The tools scored well on benchmarks and participants already had moderate AI experience.
≠
OBSERVED RESULTAI access increased completion time by 19%.
Sixteen developers completed 246 real tasks in mature repositories they knew well.
A strong benchmark and a positive belief can both miss workflow cost.
Why the results differ
AI productivity is conditional, not a property of the model alone.
TASK FITHow structured, verifiable, familiar, and well represented is the work?
WORKER FITWhat expertise, trust calibration, prompting, domain knowledge, and recovery skill does the person bring?
WORKFLOW FITHow much time goes to context loading, checking, editing, integration, coordination, and repair?
INCENTIVESAre people rewarded for quality, volume, learning, customer outcomes, or visible activity?
TIME HORIZONDoes the study measure one task, repeated use, organizational redesign, or long-run skill?
OUTCOMEAre we counting speed, quality, safety, value, job quality, profit, wages, or employment?
Adoption in the United States · 2026
Workers report broad use, while firms still show uneven organizational adoption.
55%of workers reported using AI for at least one surveyed job task. Among people who used it in the prior week, 31% said it saved one to two hours, while 3% said it added time.
MARCH 2026 HOUSEHOLD SURVEY
18%of firms reported AI use during a November 2025 to January 2026 supplement. Employment-weighted adoption was 32%, showing greater use in larger firms.
U.S. CENSUS BUSINESS DATA
Occupational exposure · global
Generative AI exposure is highest where work is digitized, and it is uneven across income and gender.
24%of global employment is in occupations with some degree of generative-AI exposure
3.3%falls in the ILO's highest exposure category, not a measured job-loss rate
34% vs 11%some exposure in high-income countries compared with low-income countries
4.7% vs 2.4%female and male employment in the highest global exposure category
Exposure follows occupational structure. It does not measure who has the infrastructure or power to benefit.
Employment evidence · August 2026
There is no sign of mass displacement, but entry-level hiring in exposed occupations is a real warning signal.
19%shortfall for workers ages 22 to 25 in AI-exposed occupations compared with the path of less-exposed peers
AGGREGATEno evidence of widespread economy-wide displacement through June 2026
MECHANISMthe divergence is driven mainly by lower hiring, not greater separation
USAGEdeclines concentrate where AI use substitutes for tasks; complementary use is flatter or positive
EXPERIENCEolder workers show no comparable exposure-related gap
CAUTIONthe evidence is descriptive and does not prove AI caused the entire divergence
Employer expectations are scenarios, not destiny
Employers expect large job churn by 2030, but the numbers combine AI with several other forces.
170Mjobs employers expect to be created by structural labor-market change14% OF CURRENT EMPLOYMENT
92Mjobs employers expect to be displaced over the same period8% OF CURRENT EMPLOYMENT
78Mnet increase implied by those employer expectationsNOT AN AI-ONLY FORECAST
The report surveys more than 1,000 large employers and includes technology, demographics, trade, and the green transition.
The economy of a productivity gain
Productivity tells us that more value is possible. It does not tell us who captures it.
LOWER PRICEScustomers receive some of the gain through cheaper or more accessible services
MORE OUTPUTthe organization serves more people, raises quality, or creates new products
HIGHER PROFITowners retain the gain when prices, pay, and staffing remain stable
BETTER WORKwages rise, hours fall, safety improves, or workers gain learning and autonomy
LESS LABORhiring slows, staffing falls, work is outsourced, or remaining jobs intensify
Markets, bargaining, competition, ownership, and public policy determine the distribution.
03
Power and governance
Governance is the design of authority, evidence, rights, accountability, and recovery around the system, not a policy document added after deployment.
A workplace power map
Power comes from controlling a different part of the system.
EMPLOYERdefines jobs and objectivesowns or licenses the systemsets staffing, pace, pay, and consequences
WORKERcontributes labor and situated knowledgecan comply, adapt, contest, organize, or exitbears many errors directly
VENDORcontrols models, updates, interfaces, and technical accessshapes what can be inspected or changedmay hold critical operational data
PUBLIC + INSTITUTIONScreate rights, standards, professional duties, procurement rules, and enforcementfund transitions and social protection
Algorithmic management evidence
Managers report real benefits and real trust problems in the same systems.
60%of surveyed managers said algorithmic management improved the quality of their decision-makingOECD EMPLOYER SURVEY
27%reported difficulty understanding the decisions or recommendations produced by the toolsOECD EMPLOYER SURVEY
56%of workers in OECD finance and manufacturing surveys felt too much data was being collectedOECD WORKER SURVEYS
A useful recommendation can still be opaque, invasive, or difficult to challenge.
Digital surveillance evidence
The purpose and use of monitoring matter more than the sensor alone.
122STUDIES MET GAO'S METHOD STANDARDS
SAFETY USEwearables and sensors can detect hazards, fatigue, cardiac issues, heat, and unsafe conditions
PACE USEproductivity targets can push workers to move faster and increase physical risk
MENTAL HEALTHcontinuous monitoring and poor transparency can increase stress, anxiety, and demoralization
EMPLOYMENTflawed benchmarks can miss important work and affect evaluation, pay, discipline, or termination
DESIGNlimit collection, state purpose, validate the measure, and prohibit harmful secondary use
A hiring failure can be simple
The first EEOC settlement involving automated hiring was not mysterious machine learning. It was an explicit rule.
WHAT HAPPENED
According to the EEOC, iTutorGroup programmed application software to reject female applicants aged 55 or older and male applicants aged 60 or older.
more than 200 qualified applicants were affected
the 2023 settlement required $365,000 and other relief
WHY IT MATTERS
Automation can scale discrimination even when the code is easy to understand.
the employer remains responsible for the employment process
existing civil-rights law still applies to automated decisions
audit must inspect rules, data, outcomes, and workflow
A governance stack
Reliable workplace AI needs controls before procurement, during use, and after harm.
01PURPOSE + NECESSITYDefine the job outcome, affected people, baseline, forbidden uses, and whether AI is needed.
02DATA + LABORDocument sources, consent, privacy, data work, representativeness, access, and retention.
04AUTHORITYSpecify who may recommend, approve, act, override, pause, and remain accountable.
05NOTICE + REDRESSTell people what is used, provide reasons and evidence, enable review, and repair harm.
06MONITOR + RETIRETrack drift, incidents, overrides, job quality, and outcomes. Suspend or remove systems that fail.
Law and policy snapshot · August 2026
Employment AI is moving from general principles toward specific duties, but coverage and timing still vary.
UNITED STATESExisting anti-discrimination and disability law applies when automated tools affect applicants or employees. Employers must consider accommodations and alternative assessment methods.FEDERAL RIGHTS
NEW YORK CITYLocal Law 144 requires a recent independent bias audit, a public summary, and advance notice before covered automated employment decision tools are used.ENFORCED SINCE 2023
EUROPEAN UNIONThe AI Act treats employment systems as high risk, with those rules now scheduled for December 2027. The Platform Work Directive adds protections around automated monitoring and decisions.IMPLEMENTATION IN PROGRESS
Law is a minimum floor. A legal system can still create poor work.
An audit is not a verdict
A selection-rate calculation can reveal a disparity and still miss whether the system should exist.
CONSTRUCTDoes the score measure something job-related, valid, and worth optimizing?
DATAWho is represented, missing, mislabeled, surveilled, or used outside the original purpose?
PERFORMANCEDoes the system work under real conditions, uncertainty, drift, and important exceptions?
GROUP OUTCOMESWho advances, is rejected, receives work, is disciplined, or is burdened by error?
HUMAN WORKFLOWHow do people interpret, defer to, override, repair, and game the system?
IMPACT + REDRESSDid job quality and opportunity improve, and can affected people obtain correction?
Appeal is a system capability
Human review is meaningful only when the reviewer has information, time, skill, and authority.
RUBBER-STAMP REVIEWA person sees the same score, lacks access to evidence, has no time to investigate, fears overriding the system, and cannot pause the consequence.
≠
EFFECTIVE REDRESSThe person receives notice and reasons, can submit relevant evidence, reaches an independent trained reviewer, pauses serious harm, obtains correction, and triggers system learning.
The right to appeal must include the practical ability to change the outcome.
Authority, emergency powers, and shutdown
The level of control should rise with the system's authority and the cost of error.
0INFORMsearch, summarize, retrieve, or explaincite and verify
1DRAFTproduce a proposal that a person authors and ownsreview before use
2RECOMMENDrank, score, or suggest an actionindependent judgment
3GATED ACTIONexecute only after explicit approvalapproval and logging
5HIGH AUTONOMYpursue goals across systems over timeindependent protection and stop
Shutdown means revoke access, pause consequences, preserve a safe fallback, investigate, repair, and restore deliberately.
Worker participation changes implementation
People who do the work know which data, exceptions, incentives, and failure costs the design team cannot see.
LATE CONSULTATION
The system is selected, configured, and measured before workers are asked for feedback.
participation becomes acceptance testing
local knowledge arrives after key choices
workers bear change without real decision rights
CO-DESIGN + SOCIAL DIALOGUE
Workers and representatives shape goals, data, measures, workflow, training, monitoring, and benefit sharing.
better exception and hazard knowledge
more legitimate and usable controls
clearer commitments on jobs and job quality
04
Building better work
The most valuable future is not humans competing against machines. It is institutions using AI to expand capability, dignity, safety, learning, and shared prosperity.
Near-term opportunities · already emerging
AI can make work more capable and more humane when that is the explicit objective.
EXPERTISE AT THE EDGEgive technicians, caregivers, teachers, farmers, and public workers timely access to specialized guidance
ACCESSIBILITY BY DEFAULTadapt interfaces, communication, assessment, and physical tools to a wider range of abilities
SAFER PHYSICAL WORKuse robots, wearables, and vision to remove hazards, reduce strain, and support aging workers
WORKER-CONTROLLED SCHEDULINGoptimize coverage while treating stability, caregiving, rest, and worker preferences as real constraints
SKILL-BUILDING COPILOTScoach reasoning, give feedback, simulate difficult cases, and preserve unaided practice
SMALL-TEAM LEVERAGEhelp small firms, nonprofits, laboratories, and communities perform work that once required a large organization
PUBLIC-SERVICE CAPACITYreduce administrative load so professionals can spend more time on judgment, relationships, and direct service
EARLY HAZARD DETECTIONidentify equipment failure, unsafe environments, workload risk, and system stress before people are harmed
Farther possibilities · plausible, not promised
Future work could distribute AI capability and economic power very differently than today's systems do.
WORKER-OWNED AGENTSpersonal systems that represent a worker's goals, protect boundaries, document contributions, and negotiate routine terms
PORTABLE SKILL MODELSevidence of capability that workers control and can carry across employers without surrendering private data
HUMAN SKILL COMMONSshared, governed knowledge infrastructures that keep professional expertise accessible beyond one vendor
TEAM DIGITAL TWINSsimulate workload, staffing, training, safety, and policy changes before imposing them on real people
ADAPTIVE APPRENTICESHIPcontinuous pathways that preserve novice work, rotate responsibility, and build expertise as automation changes
SHARED PRODUCTIVITY DIVIDENDStranslate gains into shorter hours, broader ownership, stronger benefits, public services, or transition funding
HUMAN-ROBOT CRAFTnew trades where embodied judgment, local improvisation, and flexible machines create things neither could alone
DEMOCRATIC ALGORITHMIC INSTITUTIONSworkers and affected communities gain standing to inspect, challenge, govern, and retire consequential systems
The closing test
When someone proposes AI for work, ask what kind of work system it creates.
01Which task, workflow, and human outcome are we trying to improve?
02Who defines the objective, owns the data, and receives new authority?
03What evidence shows the full human-AI workflow works in this setting?
04How are productivity gains, risks, learning, and job quality distributed?
05Who can understand, contest, override, pause, repair, and retire the system?
CAPABILITY EXPANDSOPTIONSPOWER SELECTSOUTCOMESGOVERNANCE MAKES THEMCONTESTABLE