When AI Raises Productivity, Who Receives the Gains?

The claim sounds simple because it combines two different questions

When a manager says an AI system has made a team more productive, one reading is that workers helped create more value and should share in it. Another is that the organization has reduced the cost or time needed for some tasks, while wages remain governed by existing rules. Both readings are possible. Productivity by itself does not settle the distribution question. Worker involvement can affect how a system is selected, checked, and fitted to real work. Pay-setting determines whether extra value or changed responsibilities translate into higher wages, bonuses, more predictable hours, paid training, advancement, or some other return to workers.

The evidence supports mechanisms, not a promise that consultation produces raises. OECD surveys in finance and manufacturing across seven countries found that workers in firms reporting consultation were more likely to expect AI to increase wages. Those were expectations, not later payroll records. The OECD’s selected workplace cases include instances where more complex work or new skills coincided with higher pay, often under Austrian collective agreements. A separate Finnish study found worker representation rights raised labor productivity in covered firms, but it did not study AI or how value was divided. Together these findings make a practical distinction: voice can influence the production process; a wage rule can allocate a return.

For an early or mid-career knowledge worker, the immediate question is not whether an occupation is safe. It is what changed in the task bundle, who now carries verification and accountability, and which written or customary rule sets pay. A faster first draft, a shorter claims review, or more reports completed in a day may be useful evidence, but it does not by itself show higher quality-adjusted output, profit, or compensation. Nor does it establish that a whole job will disappear. Capability, actual use, employer adoption, task redesign, labor demand, and displacement are separate signals.

This article audits the claim that workers receive productivity gains when they are involved in an AI rollout. The answer is narrower: involvement can expose hidden work and improve implementation; its effect on pay depends on bargaining power and the compensation system. A team can have meaningful consultation yet unchanged wages. A pay scheme can reward an output measure while giving workers little say over how the measure is produced. To understand who benefits, inspect both the workflow and the rule that distributes returns.

The evidence has important limits. OECD’s 2021–2022 case studies were selected examples in eight countries and four sectors, not a representative sample of all deployments. Its survey covered workers and employers in finance and manufacturing in seven countries, not the full labor market. The Finnish study concerns a historical legal change, not generative AI. These sources can help identify routes through which voice and compensation may matter. They cannot calculate an individual’s chance of a raise, predict a firm’s next decision, or establish an AI-specific law of gain sharing.

Sources: The impact of GenAI on jobs, productivity and work organization: A review of the empirical evidence; The impact of AI on the workplace: Evidence from OECD case studies of AI implementation

What counts as a gain, and who is measuring it?

Before asking who receives a gain, ask what the word means. A worker may report that a tool saves time. An employer may report that a process is faster. An operations team may count completed cases. An accountant may measure output per paid hour, operating cost, revenue, or profit. A customer may experience a quicker response. These outcomes overlap, but none is interchangeable with a wage increase. Each describes a different point in the chain from a technical capability to a possible distribution decision.

Suppose a service team uses software to draft routine answers. If an employee now handles twenty cases in a shift instead of sixteen, that could indicate increased volume. But the count alone omits whether cases were comparable, whether more answers needed correction, whether response quality changed, and whether the employee spent extra time checking outputs. If volume stayed constant while backlog fell, the gain might appear as shorter waiting time rather than more measured output. If saved time was filled with escalations, coaching, or detailed documentation, the work may have changed without the headline task becoming faster overall.

The 2026 International Labour Organization research brief reviews experiments, firm-level data, platform studies, and worker and firm surveys across countries including Australia, Denmark, Germany, Korea, Kuwait, the United Kingdom, and the United States. Its abstract describes productivity gains as real but uneven and often unverified. It also says worker-reported time savings of a few percent of working hours had not yet consistently translated into higher measured output, earnings, or employment. That synthesis is a warning against turning self-reported minutes into a verified productivity estimate. It does not establish that time savings never matter; it says the observed link to broader outcomes is not automatic.

The OECD’s 2024 policy paper reports that four in five workers in its surveys said AI improved their performance and three in five said it increased their enjoyment of work. These are survey responses about perceived experience, not an audited measure of output or a pay result. The report also identifies risks such as work intensity, data use, and inequality. A worker can feel more capable and enjoy a task more while their employer sees a different financial return. A firm may use capacity to serve more customers, reduce delays, avoid hiring, or invest in another process. The distribution depends on the market and organizational choices that follow.

A defensible productivity claim therefore needs a baseline and a comparison. Which tasks are included? What was the period before and after introduction? Are cases comparable? Is quality held constant or measured? Are errors, rework, customer outcomes, and review time included? Is the denominator worker hours, total labor cost, or machine time? Is the result a team average that conceals differences among roles? Does the gain persist after training and workflow changes? Without these details, the useful conclusion may be only that a tool appears to help with a bounded task under observed conditions.

The distinction matters for negotiation. Workers should not have to accept a vague claim that productivity rose as proof that their own performance rose, nor should a company’s efficiency statement be assumed to create a divisible pool of cash. More output may be offset by software costs, quality remediation, price competition, or investment. Conversely, a gain may include reduced repetitive burden or improved service rather than a direct profit increase. Ask what the organization actually measured and what decision it intends to make with the result.

Sources: The impact of GenAI on jobs, productivity and work organization: A review of the empirical evidence; The impact of AI on the workplace: Evidence from OECD case studies of AI implementation; The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers

How can involvement change the work before pay is discussed?

Worker involvement matters first because employees often know where a task begins and ends, what exceptions matter, and which errors create downstream costs. A system that performs well on a clean demonstration may collide with incomplete records, unusual customers, policy exceptions, or a final decision that someone must defend. A worker who sees those cases can point out the needed checks, training, escalation path, or limits before a workflow is fixed. This is a practical route from voice to implementation quality. It is not yet a route from voice to higher wages.

In its case studies, the OECD describes a Japanese auto insurer that used image analysis to assess vehicle damage and estimate repair costs. Workers raised a concern that the AI estimate could differ from the estimate produced during repair work. That disagreement matters because the early image is not the final physical inspection. The case illustrates how worker knowledge can identify a timing or information gap: a system may be useful as an initial assessment while requiring a process for later evidence and correction. The case does not show that consultation itself increased pay or that every insurer adopted such a process.

In a Japanese mortgage setting, the report describes discussions around how AI-supported screening results should be interpreted and used in lending decisions. The point is not that staff should veto every tool or that human review automatically solves bias. It is that a score can become an operational rule unless people clarify what it represents, which factors may need scrutiny, and how exceptions are handled. An employee who is accountable for the result needs a way to question inputs and route uncertain cases. Otherwise an apparent time saving may arrive alongside additional invisible checking and responsibility.

The OECD research team assembled 96 case studies based on 325 interviews, conducted in 2021–2022 across Austria, Canada, France, Germany, Ireland, Japan, the United Kingdom, and the United States. The cases span finance, manufacturing, energy, and logistics. Sixty percent of interviews were with management, HR, or AI implementation leads; 26 percent were with workers or their representatives. The report explicitly notes the management-heavy interview mix. That detail should temper how confidently readers generalize its accounts of consultation or job effects. It is evidence about observed mechanisms in selected workplaces, not a balanced census of worker experience.

The report says direct and representative consultation were often credited as important to successful implementation. Direct consultation appeared in all countries, while works-council involvement was nearly ubiquitous in the Austrian and German cases. In Japan, firms often held systematic sessions to build consensus before implementation. The report also notes that consultation could offer only short-lived reassurance about job stability. This is useful counterevidence to an overly positive story: involvement may improve understanding and fit while leaving insecurity or distribution unresolved.

A meaningful process has concrete features. Workers can identify tasks affected, discuss objectives, see what data the system uses, test results on ordinary and exceptional cases, define who verifies outputs, and report problems without being penalized for slowing an unsafe workflow. They can ask what training is paid, what changes to monitoring will occur, and when the process will be reviewed. A meeting after the vendor and management have already settled the design may collect reactions without giving workers influence over the important decisions.

Representation changes the scale of the conversation. A single employee can flag a local issue, but a representative can compare effects across shifts, job grades, or departments and seek consistent protections. Direct consultation can also reach workers who are not represented and those whose tasks are specifically affected. These approaches need not be substitutes. The right channel depends on the workforce, local law, and existing agreements. The distribution question remains separate: who can negotiate a change to pay, hours, classifications, bonuses, training, or redeployment, and how can the terms be enforced?

Worker involvement can also reveal that the gain comes from a different source than management first assumes. A document processing tool may lower typing time but shift the bottleneck to exception handling. A scheduling system may reduce planning effort but increase the number of last-minute changes employees must resolve. A code assistant may speed up routine implementation while increasing review demand for a senior engineer. Those examples are illustrations of task accounting, not claims about measured average effects. They show why workers need a chance to describe the full workflow before a narrow speed metric becomes a target.

Sources: The impact of AI on the workplace: Evidence from OECD case studies of AI implementation; Global case studies of social dialogue on AI and algorithmic management; The impact of GenAI on jobs, productivity and work organization: A review of the empirical evidence

Does worker voice itself raise productivity?

There is some causal evidence that a formal worker-voice institution can affect productivity, but the strongest study in this evidence packet is not about AI. Harju, Jäger, and Schoefer’s 2025 paper in the American Economic Journal: Applied Economics studies Finland’s 1991 introduction of a right to worker representation on boards or advisory councils in firms with at least 150 employees. Using a difference-in-differences design, the authors report that worker voice raised labor productivity. They found no effect on voluntary separations and a reduction in involuntary separations during the recessionary sample period. A separate shop-floor representation change had more limited effects.

The design asks whether outcomes changed differently for eligible firms relative to a comparison group around the policy change. That is stronger evidence for a causal effect than a simple comparison of firms that chose to consult against firms that did not. It supports a plausible information-sharing mechanism: worker representatives may bring operational knowledge into decisions that management would otherwise make with less information. Yet the paper’s abstract does not tell us that AI was introduced, that every firm improved, or that employees received a wage share. The law, institutions, and economic period are specific to Finland.

This study helps separate two arguments that are often bundled together. The first is that giving workers a voice can improve a firm’s operating performance. The second is that workers then capture part of the extra value. The Finnish result supports the first under the studied arrangement. It does not establish the second. A firm could become more productive while following a fixed salary scale, distributing profits to owners, reducing prices, investing in equipment, or changing employment. Productivity and distribution are related economic questions, but one is not a proxy for the other.

The difference-in-differences estimate also should not be used as a universal prediction for consultation. It estimates the effect of a specific representation right in a particular historical institutional setting, among firms above a size threshold. The authors report a different and more limited result for another representation institution. That variation itself matters: the channel, authority, access to information, and relationship to management can shape what voice does. Merely asking employees to submit comments is not equivalent to formal representation in governance.

A practical lesson for a team is to treat voice as part of operational design and governance. Can employees access enough information to judge whether the change is working? Can their concerns affect staffing, workflow, or the criteria for success? Are representatives involved early enough to influence procurement or implementation? Does management return with a response and a review date? These questions test whether the process carries influence. An employee survey that collects concerns but provides no feedback loop may create data without creating voice.

The right conclusion is not that voice always raises productivity, or that productivity proves consultation worked. It is that the Finnish study gives credible, bounded evidence that a representation right can improve labor productivity in an institutional setting, while the AI-specific survey and case evidence supplies mechanisms and associations. To decide what a worker should do, the relevant question is how much influence the actual process grants and what it lets workers negotiate. Do not infer a personal pay outcome from an organization-level productivity finding.

Sources: Voice at Work; Global case studies of social dialogue on AI and algorithmic management

What did AI-related pay changes look like in OECD cases?

The OECD’s AI workplace cases provide unusually concrete examples of how a task change can connect to pay. Interviewees reported wage increases for workers most affected by AI in 15 percent of case studies. This figure is a share of selected case studies, not a share of workers, employers, or all AI deployments. The increases were reported in cases from Austria. They tended to accompany greater task complexity, new skills gained through training, or performance measures that affected wages. Most commonly, the account linked pay growth to more demanding work or new skills. The cases do not support a single ladder from adoption to higher pay: similar technology-related change could meet an existing classification trigger, lead to recruitment for different skills, or be attached to an individual efficiency evaluation. The relevant mechanism is the rule that translated a particular change into compensation, where one did.

One Austrian insurance company introduced a system that reviewed customer inquiries, including contract details and damage claims. Workers who had handled routine questions began to handle more complex cases. The OECD report says the more demanding work supported higher wages, and that the collective labor agreement stipulated a right to wage increases when work became more demanding. The case is important because it identifies both a work change and a rule connecting that change to pay. Without the provision, added complexity might still have been recognized, but the case cannot tell us what would have happened under a different agreement.

Another Austrian auto-parts manufacturer used image processing for quality assurance. Workers moved from measuring random vehicle-body samples to identifying or resolving more complex quality problems after retraining. The report describes a move into technician roles and associated wage increases. The mechanism here is not a bonus for clicking a tool. It is a change in skill and job content, supported by training and reflected in role classification. A worker considering a similar transition should ask whether the learning is paid, whether the new classification has a transparent range, and whether incumbent staff can access the same route as new hires.

The cases also show a less inclusive path. In an Austrian financial services example, the organization hired younger workers with skills in data analysis and complex problem solving at higher pay rather than retraining current staff. Existing manual data maintenance and routine operational tasks had become less necessary. The report says current employees did not receive the same wage increases. The stated managerial view about the skill gap is part of a case account, not proof that retraining was impossible or that older workers were inherently unable to learn. The distributional point is plain: a new skill premium can attach to the roles a firm recruits for, while experienced staff whose tasks have changed are left outside it.

Performance-linked pay appeared in a Japanese financial services case. The firm introduced a search tool that helped employees find internal reference information. Worker bonuses were affected by efficiency evaluations, so effective use of the system could raise pay. That creates a direct link between tool use and an individual compensation metric. It also raises design questions: is the metric under a worker’s control, does it include accuracy and customer outcomes, and can employees appeal a poor rating caused by system quality, case mix, or data gaps? The case describes a possible route to gains; it does not establish that the measure was fair or that every worker benefited.

A UK financial services case connected overall firm efficiency and profitability after AI introduction to slight wage increases, including for affected employees. The report explicitly cautions that average pay changes can reflect hiring of specialized staff rather than higher pay for those whose work changed. This is an essential distinction. A firm’s average wage can rise because it added highly paid technical roles even if the pay of existing customer service or operations workers remained flat. The group that contributed to or experienced the task redesign must be tracked separately from the firm-wide average. This makes attribution more demanding than comparing average payroll before and after adoption: the relevant comparison is between people whose tasks changed and those whose pay did not, with hires and role mix kept visible. The case report’s warning does not resolve the causal contribution of AI, but it prevents a firm-wide wage average from being mistaken for a raise earned by incumbent workers.

There were also cases where wages did not change, and cases where new hires entered at lower pay after certain tasks were automated or skills were no longer required. These outcomes complicate the idea that productivity gains naturally spill over to everyone. A gain can be captured through higher company returns, lower customer prices, investment, or more capacity. A changed role can earn more when the work becomes more complex and the compensation system recognizes it. The evidence contains different routes, not a single AI wage effect. These contrasting accounts also show why the worker group must be named before a gain is described. A pay increase for a newly recruited analyst, a reclassified incumbent, and an employee receiving an efficiency bonus are different outcomes even if each appears in the same organization’s wage data. They differ in access, stability, and the role of the worker in producing the measured result. The OECD cases document examples of these routes, but do not compare them under a common metric or tell us which one produces the largest lasting return.

The case set is useful because it makes the mechanism observable: task complexity, skill acquisition, classification, performance criteria, and collective agreement language. It is weak for estimating prevalence. The 96 cases were selected qualitative studies from eight countries, with only one to seven stakeholder interviews per case and a management-heavy interview distribution. The reported 15 percent cannot be used as a forecast or benchmark for an employer. A worker can use the examples to ask better questions, not to claim that their own work deserves the same increase by statistical entitlement.

Sources: The impact of AI on the workplace: Evidence from OECD case studies of AI implementation

Why can the pay-setting rule matter more than the headline?

A productivity headline describes output or cost. A pay-setting rule determines which changes count when compensation is set. The rule might be an individual salary negotiation, a published pay band, a collective agreement, a job evaluation system, a bonus formula, a profit-sharing arrangement, or a statutory floor. If the rule says nothing about new responsibilities, workload, skill requirements, or measured results, a productivity gain can coexist with unchanged wages. If it recognizes a changed job grade or a negotiated increase, the same task change may produce a different outcome.

Collective bargaining can matter because it establishes terms for groups of workers, including people with limited individual leverage. The OECD’s 2019 comparative report describes firm-level bargaining as a mechanism that can compress pay differences within a firm and sector-level coordination as a mechanism that can reduce differences across a wider group. Those are distributional functions, not guarantees that every employee’s pay tracks the productivity of their own team. A coordinated wage scale may prioritize broad coverage and lower dispersion; it may not mirror every local productivity fluctuation.

The OECD review also describes tradeoffs. Its analysis finds collective bargaining is associated with lower wage dispersion, while some more centralized or coordinated systems may align wages less closely with firm- or sector-level productivity. The report discusses the possibility that firm-level flexibility links wages more closely to performance but may increase inequality. These are comparative patterns and mechanisms, not a direct estimate of how an AI system changes pay. A worker should not infer that a decentralized bonus is inherently fairer, or that collective terms always capture a firm-specific gain.

Different arrangements distribute value differently. A skill-based pay progression may recognize a newly required capability. A job evaluation system may reward greater judgment, complexity, or accountability. A performance bonus can share a measured result with the individuals whose results meet a threshold. Gain sharing can connect a group payout to an agreed productivity or cost measure. Profit sharing can connect returns to company-wide profitability, though employees may have little control over the relevant accounts. Negotiated reductions in hours or workload can return time instead of cash. None is automatically suitable: each depends on reliable measures, eligibility rules, timing, and the worker’s control over the outcome. These mechanisms also operate at different scales. A job grade follows the assessed role, a bonus may follow an individual or team target, and profit sharing follows the firm’s results. A worker can contribute to local efficiency while the company’s overall profits fall, or help a profitable firm without meeting an individual threshold. Naming the unit of measurement avoids treating these arrangements as interchangeable.

The case studies show why agreement language matters. In the Austrian insurance example, increased job demands triggered wage provisions through a collective agreement. That means a worker’s route to higher pay did not rest only on persuading a manager that the work felt harder. A provision supplied a category and procedure. In an unrepresented workplace, an employee may still be able to discuss reclassification or a raise, but the route can be more discretionary and less consistent across similarly affected coworkers. The legal and contractual context varies by country, sector, and status. It also raises a practical timing question: does the provision apply when duties change, when a review confirms sustained change, or only at the next bargaining round? Workers can ask what documentation triggers that process and whether the organization will review incumbent employees as well as new hires.

For an individual contributor, start with the current classification and the duties that define it. Ask whether exception handling, accountability, or specialist skills count in the grade, and whether bonus measures include quality. A title that stays the same can conceal a material change in responsibility; a narrower role can raise a different question about development or redeployment. Keep those discussions separate from a general claim that AI should produce a raise.

Metric design matters when compensation depends on performance. A volume target can reward easy cases or rushed checks; an adoption target can reward use without showing value. If the measure is team-wide but employees cannot influence staffing or assignments, individual payouts may feel arbitrary. Ask for a baseline, quality safeguards, a defined eligible group, and a way to challenge the calculation. These terms determine whether a nominal incentive is interpretable.

The practical test is whether a rule names the eligible workers, tracks a meaningful contribution, protects quality, and can be enforced. Broad standards can prevent arbitrary differences; local flexibility can recognize changes a uniform scale misses. A hybrid can pair shared floors with transparent ways to recognize changed tasks.

Sources: A Hard Day’s Night: Collective bargaining, workers’ voice and job quality; The role of collective bargaining systems for labour market performance; The impact of AI on the workplace: Evidence from OECD case studies of AI implementation

When can performance-linked pay share a gain, and when can it mislead?

A performance-linked bonus can share a gain when the measure corresponds to work that employees can influence, the baseline is credible, and quality or safety does not disappear from the formula. If a customer support team has a verified reduction in repeat contacts while maintaining resolution quality, a team-level pool could recognize collective contribution. If an employee receives a bonus for efficient use of an internal search system, the link may feel direct. But the more a metric depends on case mix, system reliability, staffing, or managerial choices, the less it represents an individual’s own contribution. Gross time saved is not itself a bonus pool: the team may spend some of that time checking outputs, handling exceptions, or serving additional demand. A measure should account for those costs and avoid counting faster production as an improvement if error rates or customer outcomes worsen.

The OECD case from Japanese financial services describes bonuses affected by evaluations of efficiency after an AI-enabled internal search tool was introduced. The report says workers who made effective use of the technology could see wage increases. It is evidence that firms can connect AI-related performance to compensation. It does not report a controlled evaluation showing the bonus increased total worker income, nor does it show what happened to errors, workload, or differences among staff. Treat it as an example of a mechanism, not proof that performance pay reliably shares productivity gains.

A sound bonus design states the outcome, measurement method, eligible workers, quality and safety safeguards, and how outside factors are handled. A team measure can fit interdependent work better than artificial individual attribution, but it can hide differences in contribution or role. The distribution formula should be visible before results are known, including how part-time schedules, leave, new starters, and support roles are treated.

The report’s Austrian examples show one alternative: recognize a shift in job content and skill through classification and negotiated pay, rather than trying to measure each person’s marginal efficiency. This may suit work where the new value comes from more judgment, analysis, or responsibility that is difficult to reduce to a clean count. A grade-based increase may be predictable and portable within an employer; a bonus can be faster to introduce but more volatile. The tradeoff depends on whether workers need stable compensation or a variable share linked to firm performance.

The comparative bargaining evidence frames the tradeoff: firm-level adjustment may track performance more closely while widening wage differences; coordinated bargaining can reduce dispersion while responding less to one team’s productivity. A proposal should therefore state whose gain is measured and whose pay changes under the rule.

Performance pay may also transfer risk from the firm to the worker. If a tool is unreliable, demand fluctuates, or staffing is cut, a worker may lose expected bonus income despite doing the required work. If the system raises throughput targets while preserving base pay, employees may deliver more work without receiving more per hour. If success depends on unpaid learning or extra review, a bonus based on gross volume can misprice the actual labor. Workers with care responsibilities, disabilities, part-time schedules, or less favorable task assignments may face unequal access to the incentive. These are design checks, not predictions about a given employer. Distribution within the eligible group matters too. An equal team payment recognizes shared workflow contributions; a weighted formula may recognize responsibilities or hours but can reproduce differences in access to high-value assignments. The chosen basis should be disclosed and reviewed for who is excluded, not only for the total payout.

A gain-sharing arrangement should define a period, baseline, included costs, quality threshold, eligible workforce, and return form: cash, time, training, staffing, or tools. Workers need a comprehensible calculation and a route to challenge errors. If margins are confidential, verified reductions in rework may offer an operational measure, but should not be presented as a profit share. The formula should also state what happens when the target is missed, the tool changes, or the trial ends.

A worker considering variable pay can compare the guaranteed salary with likely payouts under conservative assumptions. Check whether targets can change mid-period, who controls assignments, whether quality counts, and whether payout is discretionary. Compare the offer with base-pay adjustment, reclassification, paid training, or reduced hours. The evidence does not establish which format pays more in general.

Sources: The impact of AI on the workplace: Evidence from OECD case studies of AI implementation; The role of collective bargaining systems for labour market performance

What do consultation surveys establish, and what do they leave open?

The OECD’s 2022 employer and worker surveys covered finance and manufacturing in seven countries, with more than 2,000 employers and 5,300 workers. In its worker analysis, the comparison is between respondents who said their firm consulted them or a representative and respondents who reported no such consultation. In manufacturing, the consulted group was 11 percentage points more likely to expect AI to increase wages over the following decade; the report also shows a higher expectation among consulted finance workers. These are respondents’ expectations, not realized raises. The report says its analysis could not account for all wage-setting mechanisms, including sectoral agreements in some countries, and the observational comparison cannot isolate consultation from differences between firms. This is an association, not an estimate of consultation’s causal effect on pay.

The employer survey found associations between consultation and reported effects, but the pattern varied by sector and measure. In finance, employers that consulted were statistically more likely to report improved worker productivity and satisfaction, while differences were not significant for other indicators. In manufacturing, statistically significant differences appeared for working conditions measures, including satisfaction, performance measurement, and health and safety, rather than the same productivity pattern. The variation is a reason to resist a single summary such as ‘consultation improves every outcome.’ The worker and employer results also come from separate respondent groups: workers report their expectations and experience, while employers report firm outcomes. They do not follow the same worker from consultation through a later pay decision, nor link an individual’s task contribution to payroll. A reported association between consultation and expectations therefore cannot establish that a particular worker received a raise, or that consultation rather than another workplace practice explains the expectation. Sector differences further caution against treating the manufacturing percentage as a general estimate for knowledge work.

The report notes that its analysis could not account for the full set of wage-setting mechanisms, including sectoral agreements in some countries. Firms that consult may also differ in management, HR practices, representation, or technology use. The survey cannot isolate those factors or show whether consultation was early, influential, or connected to pay authority.

The case studies add process detail, but they are selected and interview based, with management voices outnumbering worker voices. The ILO’s 2025 global case-study synthesis documents social dialogue around AI and algorithmic management across regions; it does not estimate an average wage effect. Both sources illuminate context rather than a universal return from dialogue.

Taken together, these sources support asking what workers can influence and what outcomes will be reviewed. They do not support treating participation as proof that everyone benefits. The relevant evidence is the result for affected roles and the wage-setting mechanism that applies.

Taken together, these sources support asking what workers can influence and what outcomes will be reviewed. Participation is not proof that everyone benefits; actual outcomes for affected roles and the applicable wage-setting mechanism remain the relevant evidence.

Sources: The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers; Global case studies of social dialogue on AI and algorithmic management; The impact of GenAI on jobs, productivity and work organization: A review of the empirical evidence

What should a worker or team ask before a rollout?

Begin with one recurring task that has actually changed, such as drafting routine replies or classifying records. Write down what the system does, where its use stops, and which person handles exceptions or signs off. This keeps the discussion tied to an observed workflow rather than an occupation-wide forecast.

For a short trial, agree who can see the before-and-after record and when it will be reviewed. Include a quality check and the time spent correcting or verifying outputs; otherwise a faster first step can conceal a slower end-to-end process. If no baseline exists, record the current task conditions before rollout and note changes in case mix or staffing that could affect comparison. A review should have an owner who can explain the measure and a route to correct missing or misclassified data. Workers should be able to see how their work was counted, particularly when records are used for performance evaluation or pay.

Track who inherits changed duties and what support they receive. Note whether current staff can access paid training and higher-complexity work, especially if new hires are receiving a skill premium. Record the affected roles and contract groups so the discussion does not treat a team average as everyone’s result. Include staff on different shifts, part-time schedules, remote arrangements, and workers doing review or exception work; those groups may encounter different workflows or fall outside a narrow bonus eligibility rule. Where the change removes routine tasks that helped junior staff learn, ask what supervised work will replace that experience.

Bring the changed duties to the relevant role description, pay band, collective agreement, or bonus terms. Ask what event triggers review, which workers are eligible, and how a calculation or classification can be challenged. If no compensation or hours rule is expected to change, establish that clearly rather than treating a productivity claim as an implied promise. For a variable payment, clarify how leave, new starters, temporary contracts, and reassigned cases are handled before results are known; otherwise the organization can change the eligible group after the work has been done.

Choose the appropriate channel for the workforce: direct discussion for local workflow details, worker representation where available for consistent terms across roles or shifts, or both. The OECD cases show that arrangements differ by country; applicable law and agreements determine the route in a particular workplace. If workers raise an issue, ask who responds, what decision can still change, and when management will report back. A feedback channel with no answer or decision point records concerns but does not establish influence.

Agree what happens when results are mixed. The response might be to narrow the task, add human review, provide training, change staffing, or pause use while a quality problem is addressed. A review date without authority to alter the workflow is only monitoring. Keep a brief record of the task boundary, review date, quality measure, affected group, training access, applicable pay rule, and unresolved questions. That record gives workers a concrete basis for revisiting the decision after implementation, without assuming in advance that the result must be higher pay.

Sources: The impact of AI on the workplace: Evidence from OECD case studies of AI implementation; Global case studies of social dialogue on AI and algorithmic management

Verdict: voice can shape the gain; a rule determines its return

The evidence supports a conditional answer. Worker input can surface task knowledge, error paths, training needs, and accountability issues. OECD survey associations concern selected reported outcomes, not realized wage effects; the Finnish quasi-experiment concerns a 1991 representation right and productivity, not AI or pay distribution. In the selected AI cases, pay increases appeared in a minority and were linked in some instances to changed skills, task complexity, or an existing rule.

Voice can improve implementation without changing distribution. A verified output gain still does not say whether value goes to workers, customers, owners, or reinvestment. Conversely, formal rules can miss local contributions or reward a narrow subset; collective scales and bonuses each carry tradeoffs. The evidence does not identify one best model across workplaces.

For a knowledge worker, record a recurring task bundle—including corrections, exceptions, and downstream responsibility—and compare it with the role description and pay criteria. Bring that record to a conversation about how the change is measured and recognized. The useful next question for a team is whether its baseline, quality costs, workload changes, and distribution rule will be reviewed together.

Sources: Voice at Work; The impact of AI on the workplace: Evidence from OECD case studies of AI implementation; The impact of GenAI on jobs, productivity and work organization: A review of the empirical evidence

Questions readers ask