The AI Dividend Gap: Why the Benefits of AI at Work Are Flowing to Those Who Already Have Good Jobs, and How Leaders Can Share Them More Widely
Download the article here:
Listen to a review of this article:
Abstract: Early evidence on generative artificial intelligence at work suggests that its benefits are real but unevenly shared. Year-two results from the American Job Quality Study show that most employees who use AI report working faster and more creatively, yet only 28% of U.S. employees use AI at least weekly and 54% have never used it in their role. Regular use is concentrated among college graduates, managers, higher earners, and knowledge-intensive industries, the same groups that already hold quality jobs, while about half of all employees report less influence over technology adoption than they would like. This article defines the "AI dividend gap," reviews research on who adopts and who gains from generative AI, and examines organizational and individual consequences. It then synthesizes five evidence-based responses, illustrated with documented examples from entertainment, technology, healthcare, hospitality, retail, and financial services: designing for broad access, closing the use gap, building worker voice into technology decisions, bargaining over how gains and risks are shared, and reinvesting gains in better jobs.
Every major workplace technology arrives with a promise about who will benefit. Spreadsheets were going to free accountants from drudgery. Enterprise software was going to give frontline staff the information they needed to serve customers well. Generative artificial intelligence, the class of tools that can draft text, summarize documents, write code, and answer questions in natural language, came with a particularly generous promise: that it would raise the floor, giving less experienced and less credentialed workers access to expertise that was once locked inside the heads of senior colleagues.
There is evidence that it can do exactly that. There is also growing evidence that, left to its own devices, it mostly does not. In October 2026, Gallup published the first results from year two of the American Job Quality Study, a collaboration between Jobs for the Future, the Families & Workers Fund, and Gallup that surveyed 15,482 U.S. employees between late January and late March 2026 (Gallup, 2026). The headline was blunt. Workers who use AI regularly report substantial benefits, but regular users are disproportionately the people who already hold good jobs: college graduates, managers, higher earners, and employees in professional, financial, and information industries. Workers in healthcare support, manufacturing production, and retail sales mostly do not use AI at all. And more than half of all employees say they have less influence over the adoption of new technology at work than they would like (Gallup, 2026).
In my work with organizations on job design and change, this pattern is depressingly familiar. New tools land first where budgets, autonomy, and discretionary time already exist, while the people closest to customers, patients, and production lines receive technology later, with less training and less say in how it reshapes their work. What is new is the speed, which makes the window for shaping who benefits correspondingly short.
This article calls the resulting pattern the AI dividend gap: the difference between groups of workers in their access to AI, their use of it, the benefits they gain from it, and the share of those benefits they ultimately keep in the form of better pay, more interesting work, manageable workloads, and career growth. The gap is not a law of nature. Economists Acemoglu and Johnson (2023) argue that the distribution of gains from technology has always depended on choices about how technology is designed and deployed, and on whether workers have the bargaining power and institutional voice to claim a share. The research reviewed here suggests the same is true inside organizations. Design choices, training investments, participation structures, and agreements about gain-sharing all shape whether AI widens or narrows existing divides.
Readers who have followed earlier discussions in these pages about entry-level work, manager engagement, and AI-driven work intensification will see connections. The focus here is distributional: who gets the benefits, and how worker voice can change the answer.
The AI Dividend Gap Landscape
Defining the AI Dividend, Job Quality, and Worker Voice
The AI dividend refers to the bundle of benefits that AI can generate for workers and organizations: time saved on routine tasks, better quality output, more creative problem solving, more time for the parts of a job people find meaningful, and, potentially, higher pay or faster advancement. The term borrows deliberately from finance. A dividend is a distribution of value to those who hold a stake. The question this article asks is who holds the stake when AI creates value at work.
The AI dividend gap is the systematic difference between groups of workers in how much of that dividend they receive. It can open at four points along a pathway. The first is access: whether a worker has the tools, licenses, and permission to use AI at all. The second is use: whether a worker who has access actually uses AI regularly enough to become proficient, which depends on training, time, confidence, and managerial support. The third is benefit: whether use translates into meaningful improvements in the work itself. The fourth is reward: whether those improvements turn into better pay, more interesting work, career growth, or a more manageable workload, or whether they are absorbed by the organization as additional responsibilities without corresponding gains for the worker.
Job quality provides the frame for thinking about reward. The American Job Quality Study measures job quality across five dimensions: financial well-being (fair pay, stable employment, and adequate benefits); workplace culture and safety; growth and development; agency and voice, meaning influence over decisions such as pay, working conditions, and technology; and work structure and autonomy, including predictable schedules and manageable workloads (Jobs for the Future, 2025). The first wave of the study, conducted in early 2025, found that 60% of U.S. workers did not have a quality job by this definition, and 55% of employees reported limited input on workplace decisions that affect them (Jobs for the Future, 2025). It matters that agency over technology is built into this definition. Under this framework, a job in which AI is imposed without consultation is, almost by definition, a lower-quality job.
Worker voice is the capacity of employees to express concerns and influence decisions that affect their work, whether through formal channels such as unions, works councils, and joint committees, or through less formal channels such as pilot teams, advisory groups, and direct participation in design. Freeman and Medoff (1984) famously argued that unions affect organizations not only through their monopoly bargaining power but also through a "voice face" that surfaces information management would otherwise miss and reduces costly exit. Later work on technology and employment, including MIT's Task Force on the Work of the Future, concluded that institutions giving workers a voice in how technology is adopted are central to translating productivity gains into broadly shared prosperity (Autor et al., 2022). Participatory design is the practical application of that idea: involving the people who will use a technology in defining the problem it solves, shaping how it works, and evaluating its effects.
Prevalence, Drivers, and Distribution
Start with prevalence. In early 2026, 28% of U.S. employees reported using AI in their job at least weekly, and 54% said they had not used AI in their role at all (Gallup, 2026). Among those who had used AI at least a few times in the past year, the reported benefits were considerable. 63% said AI allowed them to do job tasks faster, 56% said it helped them find more creative solutions, 47% said it let them focus more on the most interesting parts of their job, and 46% said it helped them produce higher-quality work (Gallup, 2026). These are self-assessments rather than objective performance measures, and Gallup is careful to say so, but they are not trivial numbers.
Nearly a third of AI users (31%) said their employer was asking them to take on more responsibilities, and among users who said AI helped them work faster, the figure rose to 37% (Gallup, 2026). The youngest workers, aged 18 to 24, were the most likely to report expanded responsibilities, at 38%. This is the reward leak in action: some of the time AI saves is being converted into more work rather than better work. That dynamic has been discussed elsewhere as a time dividend trap; here it matters because it shows that even workers who do benefit from AI may not keep the dividend.
The distribution of use is where the gap becomes visible. 40% of college graduates used AI daily or weekly, compared with 17% of employees without a degree. Managers were regular users at a rate of 37%, compared with 25% of individual contributors (Gallup, 2026). Industry differences were even sharper. Roughly half of workers in professional and business services, financial activities, and information and media used AI regularly, while between 70% and more than 80% of workers in healthcare support, manufacturing production, and retail sales had never used it (Gallup, 2026).
Among daily or weekly AI users, 52% were employed in what the study classifies as a quality job, compared with 46% of infrequent users and 32% of non-users (Gallup, 2026). Gallup's analysts stress that this is an association, not a causal finding: the groups that hold quality jobs, such as more educated workers and managers, are also the groups more likely to use AI. That caveat cuts both ways. AI may not be creating good jobs, but it is clearly not, so far, reaching the workers whose jobs most need improving.
Benefits also differ among those who do use AI. Frequent users were roughly twice as likely as infrequent users to report each benefit: 79% versus 38% for working faster, 67% versus 39% for more creative solutions, 59% versus 26% for higher-quality work, and 61% versus 27% for more focus on interesting work (Gallup, 2026). Benefits tracked education and pay in some respects but not others. Among AI users, 68% of those with a bachelor's degree or more said AI made them faster, compared with 51% of those with a high school education or less, and 70% of users earning $10,000 or more per month reported faster work, compared with 58% of those earning under $4,000 (Gallup, 2026). Yet reported gains in creativity and quality were similar across education groups. That detail is important. It suggests that once workers without degrees actually use AI, they find it about as useful for improving their work as graduates do; the gap is concentrated in access, frequency, and the kinds of tasks people are given AI for.
Age shows a more complicated pattern. Younger workers were not the most frequent users in the Gallup data, with about one in five using AI weekly or more, and workers aged 55 and older were less likely than younger colleagues to say AI made them faster (54% versus 64%) or let them focus on more interesting work (40% versus 49%) (Gallup, 2026). Pew Research Center's survey of 5,273 employed U.S. adults in October 2024 found that workers aged 18 to 29 were the most likely to use chatbots for work at least a few times a month, and that about half of AI users (51%) held at least a bachelor's degree, compared with 39% of non-users (Lin & Parker, 2025). The same survey found that 52% of workers felt worried about the future impact of AI in the workplace and 36% felt hopeful, with only 16% saying that at least some of their work was currently done with AI (Lin & Parker, 2025).
Gender adds another dimension that the Gallup release does not report but that other research documents in detail. Humlum and Vestergaard (2025) surveyed roughly 18,000 Danish workers in 11 occupations exposed to generative AI and linked their responses to administrative records. Women were 16 percentage points less likely than men to have used ChatGPT for work, users had earned slightly more than non-users even before the tool became available, and employer restrictions and a perceived need for training were common barriers to use. A synthesis of 18 studies covering more than 140,000 people worldwide found that the gender gap in generative AI use holds across nearly all regions, sectors, and occupations, and that it persists even when access is equalized (Otis et al., 2024). The authors warn of a self-reinforcing cycle in which underrepresentation in use leads to systems that reflect women's needs less well.
What drives these patterns? Four mechanisms stand out in the research. The first is task exposure. Analysis of occupational tasks suggests that about four in five U.S. workers could see at least a tenth of their tasks affected by large language models, with higher-income jobs potentially facing greater exposure (Eloundou et al., 2024). Knowledge work is simply where current tools fit most easily. The second is organizational access and permission. Workers in frontline roles are less likely to have a company device, a license, or explicit permission, and Humlum and Vestergaard (2025) found employer restrictions to be a material barrier even among exposed occupations. The third is training and support. Boston Consulting Group's 2025 survey of more than 10,600 leaders, managers, and frontline white-collar employees across 11 countries and regions found that regular use among frontline employees had stalled at 51%, while more than three-quarters of leaders and managers used generative AI several times a week; only about a third of respondents said they had been properly trained, and regular use was sharply higher among those who had received at least five hours of training (Beauchene et al., 2025). The fourth is voice. When workers have little influence over which tools arrive and how they are configured, tools tend to be selected for the priorities of those who do have influence.
Only the first of these mechanisms is primarily technological; the other three are organizational choices.
Organizational and Individual Consequences
Organizational Performance Impacts
The most important organizational consequence of the AI dividend gap is that it leaves value unclaimed. The best causal evidence on generative AI at work suggests that the largest productivity gains often accrue to precisely the workers who are least likely to receive the tools under current adoption patterns.
Brynjolfsson et al. (2025) studied the staggered rollout of a generative AI conversational assistant to 5,179 customer support agents. Access to the tool increased productivity, measured as issues resolved per hour, by 14% on average. The gains were concentrated among novice and lower-skilled workers, at roughly 34%, while experienced and highly skilled agents saw little effect. The authors interpret the tool as spreading the tacit practices of top performers to newer colleagues, helping them move down the learning curve faster. Customer sentiment improved and employee retention rose. A field experiment with 758 Boston Consulting Group consultants found a similar leveling pattern: on tasks within the capabilities of the AI, consultants using GPT-4 completed about 12.2% more tasks, worked about 25.1% faster, and produced work rated more than 40% higher in quality, with below-average performers improving by about 43% and above-average performers by about 17% (Dell'Acqua et al., 2023).
Put these findings next to the Gallup distribution and the organizational problem becomes clear. If AI raises the performance of less experienced and lower-performing workers most, then concentrating access and support among managers, graduates, and already-strong performers is a misallocation. The organization is buying the tool for the people who gain least from it and withholding it, implicitly, from the people who would gain most.
The second consequence is risk. The same consulting experiment found that on a task deliberately chosen to fall outside the AI's capabilities, consultants using AI were 19 percentage points less likely to reach a correct answer than those working without it (Dell'Acqua et al., 2023). The authors call this the jagged technological frontier: AI is very good at some tasks and quietly wrong on others that look similar. Workers need judgment, training, and permission to question outputs in order to navigate that frontier. Where frontline workers receive tools without training or voice, the risk of uncritical reliance grows; where they are excluded entirely, organizations lose the frontline knowledge that often identifies where the tool breaks.
The third consequence concerns adoption itself. Researchers at MIT, drawing on more than 50 interviews with developers, executives, labor leaders, and policymakers, concluded that involving workers in designing technology and redesigning work processes tends to produce better results than top-down, technology-only rollouts (Kochan et al., 2023). They also noted that workers in call centers welcomed tools that helped them serve customers but objected to tools used to monitor them, a distinction that shapes whether a tool is embraced or quietly resisted. The 52% of employees who want more say in technology adoption are not merely expressing a preference (Gallup, 2026). They are signaling a reservoir of knowledge and goodwill that most organizations are not tapping.
Finally, if AI becomes associated with more responsibilities and no corresponding reward, as nearly a third of users already report, the technology becomes a source of grievance, and the productivity benefits of a tool people resent are fragile.
Individual Wellbeing and Stakeholder Impacts
For individual workers, the AI dividend gap compounds existing disadvantage. If AI raises the productivity and satisfaction of workers who already have good jobs while bypassing workers without degrees and in frontline roles, the gap in job quality widens. Humlum and Vestergaard (2025) titled their paper with this concern in mind: unequal adoption exacerbates existing inequalities among workers.
There is also a skills trajectory concern. Because frequent users report roughly double the benefits of infrequent users (Gallup, 2026), proficiency itself appears to be a large part of the dividend. Workers kept on the margins are missing not just today's gains but the learning that unlocks tomorrow's, which makes them less competitive for internal moves and external opportunities.
Worry is a wellbeing issue in its own right. Pew's finding that more workers are worried than hopeful about AI (Lin & Parker, 2025) is not surprising when workers lack information or influence. Uncertainty is harder to bear when people have no way to influence the outcome. A frontline worker who sees AI transforming professional colleagues' jobs, while their own role receives only new monitoring tools or vague promises, has good reason to feel anxious.
The responsibility creep carries its own wellbeing risk. When the youngest workers are the most likely to be asked to take on more as a result of AI (Gallup, 2026), early-career employees may experience the technology primarily as intensification. That is a poor foundation for a healthy relationship with tools they will use for decades.
The gender dimension carries stakeholder consequences beyond the workforce. If women use generative AI less often and less intensively, then the data and feedback that shape these systems will reflect women's needs less well, and products built on them may serve female customers and employees less effectively (Otis et al., 2024). The same logic applies to frontline workers more broadly: tools refined primarily through the feedback of managers and knowledge workers will fit the work of managers and knowledge workers.
Customers are stakeholders too. In service industries, the quality of frontline work is the quality of the service, and the evidence that AI assistance improved customer sentiment when it supported less experienced agents (Brynjolfsson et al., 2025) suggests that closing the gap also serves the people workers serve.
Evidence-Based Organizational Responses
The research points to five complementary responses that organizations can use to narrow the AI dividend gap. Table 1 summarizes the evidence base, key practices, and illustrative examples for each. They are presented roughly in the order of the dividend pathway, from access through reward, but in practice they work best together. Worker voice in particular runs through all five.
Table 1: Evidence-based responses to the AI dividend gap
Response | Evidence Base | Key Practices | Illustrative Examples |
Design for broad access and the novice advantage | Largest gains for less experienced workers in field studies; task exposure is broad | Enterprise-wide access; tools built for frontline tasks; capture top-performer practices | JPMorgan Chase LLM Suite; customer support assistant studied by Brynjolfsson et al. |
Close the use gap with targeted enablement | Training and leadership support strongly predict regular use; gender and education gaps persist | Paid learning time; peer coaches; targeted outreach to under-users | Walmart and OpenAI certification; Microsoft and AFL-CIO education commitments |
Build worker voice into technology decisions | Participatory design improves outcomes; workers want more influence | Joint committees; early notice; co-design with frontline teams; expert support for worker bodies | Kaiser Permanente and the Alliance of Health Care Unions; German works councils; Microsoft labor summits |
Bargain over how gains and risks are shared | Collective agreements set terms for consent, notice, retraining, and compensation | Notice periods; consent and disclosure; retraining and severance; impact bargaining | Writers Guild of America; SAG-AFTRA; Culinary Union and Caesars; ZeniMax and CWA |
Reinvest the dividend in better jobs | Gains can be converted into intensification or into upgraded roles | Redeploy into higher-value work; guard against responsibility creep; keep human paths | IKEA remote design advisers; Klarna's return to human service |
Design for Broad Access and the Novice Advantage
The strongest argument for broad access comes from the field studies already discussed. When generative AI captures and distributes the know-how of experienced workers, the people who benefit most are those who have the least of that know-how. Brynjolfsson et al. (2025) found that the AI assistant in their study raised productivity of novice and lower-skilled agents by roughly 34% while having little effect on the most experienced, and that it improved customer sentiment and retention. Dell'Acqua et al. (2023) found larger proportional gains for below-average consultants than for above-average ones. These findings do not imply that AI always compresses skill differences; much depends on the task and the tool. They do imply that an organization that wants to maximize AI's return should, at a minimum, not systematically withhold it from less experienced and less credentialed staff.
Broad access is not just about licenses. Because current tools fit knowledge work most easily (Eloundou et al., 2024), closing the gap requires deliberate attention to the tasks frontline workers actually do, rather than rolling out general-purpose chat tools and waiting.
Effective approaches:
Default to enterprise-wide access: Make approved AI tools available to all employees with relevant tasks, including those without assigned desks, rather than piloting only among managers and professional staff.
Build for frontline workflows: Identify high-frequency frontline tasks (shift handovers, policy lookups, customer queries, documentation) and configure tools specifically for them, using mobile and voice interfaces where workers do not sit at computers.
Codify top-performer practices: Use AI to make the tacit knowledge of experienced workers available to newer colleagues, as in the customer support study, and credit the experienced workers whose knowledge is captured.
Map the frontier jointly: Work with users to identify tasks where the tool performs poorly, and publish those boundaries so that novices are not misled by plausible but wrong output.
Track access by segment: Report AI access and use by role, education level, pay band, gender, and age, not just in aggregate.
JPMorgan Chase offers an example of scale in financial services. The bank built LLM Suite, an internal AI portal that integrates large language models, and by mid-2025 it was being used by more than 200,000 employees for tasks in legal, sales, and client services such as analyzing contracts, drafting presentations, preparing client emails, and automating reporting (Arizent, 2025). The significance for the dividend gap is that the bank treated generative AI as general workplace infrastructure rather than a specialist tool for a privileged group. Public reporting does not, however, show how use or benefit is distributed across roles, which is precisely the information leaders need to judge whether broad access is translating into broad benefit.
The customer support deployment studied by Brynjolfsson et al. (2025) illustrates the design side. The assistant suggested responses to agents handling live customer chats, drawing on patterns from high-performing agents, and its biggest effects were on the newest and least experienced. Design shaped distribution.
A common pitfall is assuming that providing access is the same as providing benefit. Gallup's data show that frequent users report roughly double the gains of infrequent ones (Gallup, 2026). A license that sits unused does nothing to close the gap, and an access rollout that is not paired with the enablement and voice practices below can deepen it, because the workers best positioned to experiment will pull further ahead.
Close the Use Gap with Targeted Enablement
The step from access to regular use is where many of the demographic gaps open. Boston Consulting Group's 2025 survey found that frontline regular use had stalled at 51% while more than three-quarters of leaders and managers used generative AI several times a week; employees who received at least five hours of training, and who had in-person training and coaching, were far more likely to be regular users, and positive sentiment about generative AI rose from 15% to 55% when employees reported strong leadership support, though only about a quarter of frontline employees reported receiving it (Beauchene et al., 2025). Humlum and Vestergaard (2025) found that a perceived need for training was one of the main barriers to use among Danish workers, and the gender gap they documented is consistent with differences in who receives encouragement and support. Otis et al. (2024) found that the gender gap persisted even when access was equalized, which suggests that enablement has to address confidence, norms, and perceived permission, not just availability.
Effective approaches:
Fund learning time, not just learning content: Provide paid, scheduled time for AI training, especially for hourly workers whose schedules leave no slack for self-directed learning.
Use peer coaches from the frontline: Train respected frontline employees as AI coaches, since workers often learn new tools most readily from colleagues who do similar jobs.
Target under-users explicitly: Use segment data to design outreach for groups with low use, such as women in technical roles, workers without degrees, and older employees, framed around their tasks rather than generic capabilities.
Make permission explicit: Publish clear guidance on what employees may and may not do with AI, since uncertainty about permission suppresses use most among those who feel least secure.
Equip managers to support use: Give frontline supervisors the skills and time to encourage experimentation and to recognize AI-enabled improvements.
Walmart provides a retail example. In September 2025, the company's chief people officer told U.S. associates that Walmart was collaborating with OpenAI on a certification program that would give frontline and office-based associates free access to a tailored version of OpenAI's certification through Walmart Academy, launching in 2026, and that associates could already use the company's Live Better U education benefit for AI courses (Walmart, 2025). The note framed this as part of a nearly $1 billion commitment to skills training through 2026. The design choice worth noting is the explicit inclusion of frontline associates, in an industry where Gallup found most workers have never used AI. The company's announcement does not make commitments about wages or career paths linked to the certification, so its effect on the reward end of the dividend pathway remains to be seen.
Microsoft and the AFL-CIO, the federation of 60 unions representing 12.5 million workers, built education into their December 2023 partnership. Microsoft committed to provide AI learning sessions for labor leaders and workers starting in 2024, along with workshops through 2026 and joint development of courses aimed at sectors most affected by AI (Microsoft, 2023). Channeling training through unions is a distinctive approach: it reaches workers through institutions they already trust, and it pairs skill-building with a voice structure, discussed below.
A common pitfall is designing training around the most enthusiastic users. Voluntary, self-paced programs tend to be taken up by those who already have time, confidence, and encouragement. The finding that gender gaps persist even when access is equalized (Otis et al., 2024) is a reminder that an open invitation is not the same as an equal one. Without targeting, enablement programs can widen the very gaps they are meant to close.
Build Worker Voice into Technology Decisions
The third response addresses the deepest finding in the Gallup data: 52% of employees have less influence over technology adoption than they would like (Gallup, 2026). Worker voice matters for the dividend gap in two ways. First, it improves the fit of technology to work, which raises the benefits workers experience. Second, it gives workers a seat at the table when decisions are made about how gains are used, which affects whether they keep the dividend. Kochan et al. (2023) recommend making participatory design standard practice through standing task forces or advisory bodies, encouraging worker experimentation, and ensuring that adjustment provisions such as retraining are in place. Freeman and Medoff's (1984) argument about the voice face of unions applies equally to nonunion voice mechanisms: organizations gain information they would not otherwise have.
Effective approaches:
Involve workers at inception: Bring frontline representatives into the problem-definition stage, before vendors are chosen, rather than consulting them only at rollout.
Create joint governance bodies: Establish standing labor-management or employee-management committees with real access to technology plans and decision-makers with investment authority.
Resource worker participants: Give worker representatives access to independent technical expertise so that participation is informed rather than symbolic.
Separate support tools from surveillance tools: Make explicit commitments about monitoring uses of AI, since monitoring is where trust is most easily lost.
Close the feedback loop: Report back on how worker input changed decisions, and evaluate deployments jointly.
Kaiser Permanente and the Alliance of Health Care Unions offer the most detailed recent example in healthcare. During their 2025 bargaining round, a joint AI subgroup of 17 union and 17 management members met repeatedly between May and August and spent more than 150 hours in deliberations (Kochan et al., 2026). The resulting agreement created a national AI task force of five labor and five management members, including senior leaders with technology investment authority, charged with sharing information on planned rollouts, bringing in experts, guiding training, and reviewing concerns about uses such as hiring and performance evaluation. It also committed to strengthening Kaiser's more than 3,500 unit-based teams, frontline labor-management groups that already work on improvement projects, as channels for AI ideas, and to employee voice throughout the technology lifecycle (Kochan et al., 2026). The MIT researchers who documented the process are candid about its limits. The parties disagreed about what early involvement should mean, management cited confidentiality and decentralized decision-making, and the preamble does not create clearly enforceable guardrails, leaving much to the task force. The example is valuable precisely because it shows both the promise and the hard work of institutionalizing voice.
Germany shows how voice can be built into the legal framework itself. German works councils have long held co-determination rights over the introduction of technical equipment capable of monitoring employees, and a 2021 reform known as the Works Council Modernization Act added provisions specific to AI. Works councils may now draw on an outside expert whenever the employer plans to introduce or use AI, with that expert's involvement treated as necessary by law, and they gained participation rights over selection guidelines when AI is used in decisions about hiring, transfers, and dismissals (Ogletree Deakins, 2021). Practitioners note that the expert presumption applies where the AI connection is concrete and relevant, and that the co-determination right over technical equipment is one of the main tasks it can attach to (Schüttauf, 2024). The point for leaders outside Germany is not to copy the statute but to recognize the design principle: worker representatives need both a formal role and the technical capacity to exercise it.
Microsoft's partnership with the AFL-CIO also included a voice component. The company committed to hosting labor summits at which union leaders and workers would meet its AI developers and researchers to share experiences and feedback and to help co-design worker-centered technology, beginning with unions in selected sectors (Microsoft, 2023). The partnership included a neutrality framework for future organizing by AFL-CIO affiliates, building on an earlier agreement with the Communications Workers of America. This reaches upstream of any single employer, to the tools themselves.
A common pitfall is consultation theater: inviting workers to comment on decisions that have already been made. Workers recognize this quickly, and it can do more damage to trust than no consultation at all. The Kaiser Permanente case shows that even well-intentioned partners struggle over the timing and depth of involvement; organizations without a partnership history should expect the same struggle and plan for it explicitly.
Bargain over How Gains and Risks Are Shared
Voice shapes design; bargaining shapes distribution. Some of the clearest examples of workers securing a share of the AI dividend, or at least protection from its downside, come from collective agreements. These agreements typically address four issues: notice before new technology is introduced, consent and disclosure when AI uses workers' output or likeness, retraining and severance when jobs change, and the right to bargain over impacts. Acemoglu and Johnson (2023) argue that countervailing power has historically been decisive in determining whether productivity gains reach workers, and these agreements are a contemporary expression of that argument.
Effective approaches:
Commit to advance notice: Provide meaningful notice before introducing AI that changes job content, so that workers and managers can plan training and redesign.
Require consent and disclosure where identity or creative output is involved: Tell workers when they are being given AI-generated material, and obtain specific consent before using their work or likeness to create AI outputs.
Pair change with retraining and transition support: Guarantee retraining for affected workers and fair severance where roles are eliminated.
Preserve choice over tool use where appropriate: In creative and professional roles, allow workers to choose whether to use AI rather than mandating it.
Agree to revisit impacts: Commit to bargaining or consultation over effects as deployments mature, rather than treating initial terms as final.
The Writers Guild of America secured some of the earliest AI provisions in its 2023 agreement with film and television studios after a lengthy strike. Under the contract, AI is not a writer and AI-generated material is not literary material; a company cannot require a writer to use AI tools, though a writer may choose to with the company's consent; companies must disclose when material given to a writer was generated by AI; and the studios agreed to meet with the Guild to discuss their current and planned uses of AI (Writers Guild of America, 2023). The provisions protect writers' credit and compensation, which is how writers capture value, while leaving room for voluntary use.
SAG-AFTRA extended this approach to video game performers in an agreement members ratified in July 2025. The agreement requires clear written consent with a reasonably specific description before a performer's digital replica is used, bars employers from obtaining consent at initial hiring for most uses in future games, requires usage reports after a game's release identifying where replicas were used and how compensation was calculated, and pays for vocal replicas according to the number of lines generated (Frankfurt Kurnit Klein & Selz, 2025). Here, the dividend from AI-generated performance is partly routed back to the performer whose work made it possible.
The Culinary Workers Union Local 226 brought similar principles to hospitality. Its November 2023 tentative agreement with Caesars in Las Vegas required six months' notice before introducing AI, robotics, or other new technology, 30 days' notice before modifying existing technology, retraining for affected workers, and severance and continued benefits for those affected or laid off (Kullgren, 2023). For housekeepers, cooks, and food servers, groups that Gallup's data suggest are among the least likely to use AI themselves, these provisions protect against the downside of technology deployed around them.
ZeniMax Workers United, a union of quality assurance workers organized with the Communications Workers of America, reached a tentative agreement with Microsoft in December 2023 under which ZeniMax committed to uses of AI that augment human capacities and support worker productivity, growth, and satisfaction, to notify the union when AI implementation may affect members' work, and to bargain over those effects on request (Communications Workers of America, 2023). The language is general, but the notice-and-bargain structure creates a durable channel for addressing the dividend question as tools evolve.
A common pitfall for managers in nonunion settings is to treat these agreements as relevant only to unionized industries. Their core elements (notice, consent, disclosure, retraining, and a standing process for revisiting impacts) can be adopted voluntarily as policy, building the trust that makes adoption smoother.
Reinvest the Dividend in Better Jobs
The final response addresses the reward end of the pathway. If AI saves time and raises quality, the organization faces a choice about what to do with the gain. It can take the gain as reduced headcount, as additional work for the same people, or as an opportunity to upgrade jobs. The Gallup finding that 31% of AI users are being asked to take on more responsibilities, rising to 37% among those who say AI makes them faster, suggests that many organizations are defaulting to the second option without deciding to (Gallup, 2026). More responsibility is not inherently bad, particularly when it brings growth and recognition, but when it arrives without pay, development, or workload adjustment, it converts the worker's dividend into the employer's.
Effective approaches:
Decide explicitly where time savings go: Before deployment, agree on how freed capacity will be used, for example for customer relationships, quality improvement, learning, or reduced overtime.
Redeploy rather than discard: When AI absorbs routine tasks, design new roles that use workers' domain knowledge, and train affected workers for them.
Link new responsibilities to recognition: When AI-enabled productivity leads to expanded scope, adjust job grades, pay, or development opportunities accordingly.
Protect a human path for customers and workers: Keep human escalation routes where quality and trust matter, and avoid cutting so deeply that remaining staff are overwhelmed.
Share gains visibly: Consider gain-sharing mechanisms, such as bonuses or reduced hours, tied to AI-enabled improvements.
IKEA provides the most frequently cited retail example of redeployment. Ingka Group, the largest IKEA franchisee, introduced an AI customer service bot called Billie in 2021, and by mid-2023 the bot was reported to be handling 47% of customer queries (CX Today, 2023). Rather than reducing contact center staff in line with that volume, the company trained 8,500 contact center workers as remote interior design advisers since 2021. Remote design sales through phone and video generated €1.3 billion in 2022, about 3.3% of revenue, and the company set a goal of reaching 10% by 2028 (CX Today, 2023). Asked whether Billie would lead to job cuts, an Ingka executive said that was not what the company was seeing at the time. The example shows the dividend being reinvested in a more skilled, more customer-facing role that draws on workers' product knowledge.
Klarna offers the cautionary counterpart in financial technology. In February 2024, the company said its AI assistant was doing the work of 700 customer service agents and had handled 75% of customer chats in its first month, a claim reported as the company's own (Shibu, 2025). Klarna had paused hiring and reduced headcount largely through attrition. By May 2025, its chief executive told Bloomberg that the company was hiring human customer service staff again, acknowledging that the AI-focused approach had produced lower-quality results and that customers should always be able to reach a person if they wish (Shibu, 2025). The lesson for the dividend gap is that capturing AI gains purely as headcount reduction can erode the quality the gains were meant to protect, and that the workers who know the work best are a resource to be redeployed rather than an expense to be eliminated.
A common pitfall is assuming that because workers report AI makes them faster, they are better off. Speed is only a benefit to the worker if some of the time saved comes back to them as more interesting work, development, or relief. Leaders should ask not only whether AI is saving time but where that time is going and who decides.
Building Long-Term AI Dividend Sharing Capability
The five responses above can narrow the gap in a given deployment. Sustaining that progress as tools evolve requires organizational capability: the systems, structures, and norms that make inclusive adoption the default rather than the exception. Three pillars stand out.
Benefits can leak at each stage of the pathway, from uneven access to rewards captured as additional responsibilities, and worker voice across the lifecycle is what holds the dividend in place.
Measure the Dividend by Segment
Most organizations measure AI adoption only in aggregate: licenses issued, prompts submitted, hours saved. These numbers can look healthy while hiding large gaps between managers and frontline staff.
Building a distributional measurement capability means tracking each stage of the dividend pathway by segment: access, frequency of use, self-reported benefits, and reward outcomes such as changes in workload, responsibilities, pay progression, and internal mobility. Segments should include role level, education, pay band, gender, age, and location, with appropriate privacy safeguards. The American Job Quality Study offers a useful template, since it measures not only AI use but job quality dimensions including agency and voice (Jobs for the Future, 2025). Organizations can include questions about influence over technology decisions in their own engagement surveys and compare their results with the national benchmark that 52% of employees want more influence than they have (Gallup, 2026).
Measurement should also be joint where possible. When workers and managers agree on what to track and review results together, as the Kaiser Permanente task force is charged with doing (Kochan et al., 2026), the data become a shared basis for action rather than a management scorecard.
Institutionalize Voice Across the Technology Lifecycle
Voice that depends on the goodwill of a particular executive or the urgency of a particular dispute does not last. Long-term capability requires standing structures that give workers influence at every stage: when problems are defined, when vendors and tools are selected, when workflows are redesigned, when deployments are evaluated, and when decisions are made about how gains are used.
The research and examples suggest several design principles. Structures should have access to decision-makers with real authority over technology investments, as the Kaiser agreement specifies (Kochan et al., 2026). Worker participants should have independent technical support, a principle German law has made explicit (Ogletree Deakins, 2021). Voice should be channeled through existing frontline structures where they exist, such as unit-based teams or quality circles, rather than creating parallel bodies that compete for scarce time. And voice should extend upstream where possible: partnerships like the one between Microsoft and the AFL-CIO, in which workers influence tool design at the source, address a level of the problem no single employer can reach (Microsoft, 2023).
For nonunion organizations, this might mean employee technology councils with rotating frontline membership, published technology roadmaps with comment periods, and formal commitments on notice and monitoring, extending to nonunion workers the kinds of protections Kochan et al. (2023) see in collective agreements.
Tie AI Gains to Job Quality and Career Mobility
The ultimate test of whether the AI dividend is shared is whether jobs improve. That requires connecting AI deployment decisions to the organization's job architecture, compensation, and career systems. When AI absorbs routine work, roles should be redesigned to use the time freed, ideally in ways that draw on workers' judgment and relationships, as IKEA did with its design advisers (CX Today, 2023). When AI-enabled productivity leads to expanded responsibilities, job grades and pay should be reviewed. When frontline workers build AI proficiency, career paths should open to roles that use it.
This pillar also means treating AI proficiency as a mobility asset for those who have historically had the least mobility. Programs like Walmart's certification effort (Walmart, 2025) create a credential; the capability lies in connecting that credential to internal opportunities and to pay. Without that link, training becomes another obligation rather than a path.
Finally, tying gains to job quality means being explicit about where AI should not be used to cut. Klarna's experience (Shibu, 2025) and the frontier risks documented by Dell'Acqua et al. (2023) both point to the value of preserving human capacity in roles where quality, trust, and judgment matter. Leaders who make these commitments publicly signal that the organization intends to share the dividend, which in turn makes workers more willing to help generate it.
Conclusion
Generative AI is producing real benefits for many of the people who use it. Most AI users say it makes them faster and more creative, and frequent users report roughly twice the benefits of infrequent ones (Gallup, 2026). The problem is not that the technology fails to deliver. It is that the benefits are flowing disproportionately to workers who already have good jobs: college graduates, managers, higher earners, and professionals in knowledge-intensive industries. Workers in healthcare support, production, and retail sales mostly have not used AI at all, and more than half of all employees want more say over how technology is adopted than they currently have.
The irony is that the best causal evidence suggests the workers being left out are often the ones who stand to gain most. When AI is designed to spread the know-how of experienced colleagues, its largest effects tend to fall on newer and less skilled workers (Brynjolfsson et al., 2025; Dell'Acqua et al., 2023). Closing the AI dividend gap is therefore not only a fairness imperative. It is a way of capturing value that current adoption patterns leave on the table.
The responses are within leaders' control: broad access, targeted enablement, voice from inception, written terms for notice, consent, and retraining, and reinvestment of gains in better jobs. The examples reviewed here, from Hollywood writers and Las Vegas hotel workers to healthcare unions, retail associates, and bank employees, show these approaches being tested across industries with varying success. Together they suggest that the distribution of AI's benefits is a choice organizations are making now, whether they recognize it or not.
Practical takeaways for leaders:
Measure distribution, not just adoption: Track AI access, use, benefit, and reward by role, education, pay, gender, and age.
Give the tools to those who gain most: Extend AI access and frontline-specific tools to less experienced and less credentialed workers, rather than concentrating them among managers.
Pay for learning time: Fund scheduled training, peer coaching, and targeted outreach to groups with low use.
Bring workers in at the start: Create standing joint structures with real authority and technical support to shape AI decisions from problem definition onward.
Set the terms in writing: Commit to notice, consent, disclosure, retraining, and a process for revisiting impacts, whether or not a union is present.
Decide where the time goes: Agree in advance how AI-enabled capacity will be used, and link expanded responsibilities to recognition, pay, and growth.
Keep a human path: Preserve human roles where quality, trust, and judgment matter, and redeploy experienced workers rather than discarding their knowledge.
Research Infographic
References
Acemoglu, D., & Johnson, S. (2023). Power and progress: Our thousand-year struggle over technology and prosperity. PublicAffairs.
Arizent. (2025, June 3). JPMorganChase's LLM Suite wins American Banker's 2025 Innovation of the Year Award grand prize. Arizent.
Autor, D., Mindell, D. A., & Reynolds, E. B. (2022). The work of the future: Building better jobs in an age of intelligent machines. MIT Press.
Beauchene, V., Duranton, S., Kalra, N., & Martin, D. (2025). AI at work 2025: Momentum builds, but gaps remain. Boston Consulting Group.
Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942.
Communications Workers of America. (2023, December 14). ZeniMax workers reach agreement with Microsoft on contractors and AI. Communications Workers of America.
CX Today. (2023, June 15). IKEA's contact center agents become interior design advisors. CX Today.
Dell'Acqua, F., McFowland, E., III, Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality (Working Paper No. 24-013). Harvard Business School.
Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308.
Frankfurt Kurnit Klein & Selz. (2025). Inside the new SAG-AFTRA Interactive Media Agreement: New standards for AI and digital replicas. Frankfurt Kurnit Klein & Selz.
Freeman, R. B., & Medoff, J. L. (1984). What do unions do? Basic Books.
Gallup. (2026, October 6). AI benefits at work unevenly distributed. Gallup News.
Humlum, A., & Vestergaard, E. (2025). The unequal adoption of ChatGPT exacerbates existing inequalities among workers. Proceedings of the National Academy of Sciences, 122(1), e2414972121.
Jobs for the Future. (2025). The American Job Quality Study. Jobs for the Future.
Kochan, T. A., Armstrong, B., Shah, J., Castilla, E. J., Likis, B., & Mangelsdorf, M. E. (2023). Bringing worker voice into generative AI. Massachusetts Institute of Technology.
Kochan, T. A., Kelly, E. L., & Minster, A. (2026). Negotiating to work in partnership on AI: The 2025 agreement between Kaiser Permanente and the Alliance of Health Care Unions. MIT Institute for Work and Employment Research.
Kullgren, I. (2023, November 9). Las Vegas union scores AI, daily cleaning wins in Caesars pact. Bloomberg Law.
Lin, L., & Parker, K. (2025). U.S. workers are more worried than hopeful about future AI use in the workplace. Pew Research Center.
Microsoft. (2023, December 11). AFL-CIO and Microsoft announce new tech-labor partnership on AI and the future of the workforce. Microsoft.
Ogletree Deakins. (2021, August 20). Germany's Works Constitution Act: Important changes to the law effective June 2021. Ogletree Deakins.
Otis, N. G., Delecourt, S., Cranney, K., & Koning, R. (2024). Global evidence on gender gaps and generative AI (Working Paper No. 25-023). Harvard Business School.
Schüttauf, A. (2024). The necessity of an AI expert for the works council. Luther Employment Law Newsletter, 4/2024. Luther Rechtsanwaltsgesellschaft.
Shibu, S. (2025, May 9). Klarna is hiring customer service agents after AI couldn't cut it on calls, according to the company's CEO. Entrepreneur.
Walmart. (2025, September 4). Donna Morris to associates: Being people-led means investing in you. Walmart.
Writers Guild of America. (2023). Artificial intelligence [Know your rights]. Writers Guild of America.
Jonathan H. Westover, PhD, Chief Research Officer (Nexus Institute for Work and AI); Co-Founder & Chief Workforce and Learning Officer (Future State University); Founder & CEO (Human Capital Innovations); Professor of Organizational Leadership & Change (UVU). Read Jonathan Westover's executive profile here.
Suggested Citation: Westover, J. H. (2026). The AI Dividend Gap: Why the Benefits of AI at Work Are Flowing to Those Who Already Have Good Jobs, and How Leaders Can Share Them More Widely. Human Capital Leadership Review, 40(2). doi.org/10.70175/hclreview.2020.40.2.6
