The Time Dividend Trap: Why AI's Time Savings Are Intensifying Work, and How Leaders Can Reinvest Them Wisely
Download the article here:
Listen to a review of this article:
Abstract: Generative AI is widely promoted as a way to give workers time back, yet a growing body of evidence suggests that the hours it saves rarely become slack, learning, or recovery. Survey, administrative, ethnographic, and field-experimental studies show modest average time savings that are quickly absorbed by broader task portfolios, faster pacing, higher output expectations, and new coordination and review work. This article examines that pattern, described here as the time dividend trap, in which organizations capture AI-generated time as additional output by default rather than by deliberate choice. Drawing on research on work intensification, the job demands–resources model, recovery, and four-day-week trials, alongside documented examples from technology, financial services, healthcare, professional services, and the public sector, it outlines five evidence-based responses: making the dividend visible, agreeing explicit reinvestment rules, converting time into recovery, investing in learning and judgment, and redesigning the rhythm of AI-enabled work. It closes with guidance on building durable governance of the time dividend.
In the autumn of 2026, two large surveys landed within a week of each other and told a strikingly similar story. The Jabra Research Institute's WorkLife 2026 study, based on a May 2026 survey of more than 6,000 knowledge workers in France, Germany, India, Poland, the United Kingdom, and the United States, found that AI was not shrinking work so much as broadening it. About half of respondents reported spending more time than a year earlier on strategic planning and analysis (51%), creative work (50%), and even administrative tasks (49%), and 85% said they were doing more of at least one kind of work (Jabra, 2026). A few days earlier, Workday's global workforce report showed that 40% of business leaders expected AI to help them get more out of the employees they already had, compared with 28% who expected it to reduce headcount (Workday, 2026).
Put those findings side by side and a pattern comes into focus. Workers are being told, and often genuinely experience, that AI saves them time. Leaders, for their part, are planning to convert that saved time into more output from the same people. The result is a quiet bargain that nobody explicitly negotiated: the minutes AI frees up are reabsorbed into the job almost as soon as they appear. I call this the time dividend trap. The time dividend is real, but it is captured by default, as additional volume and scope, rather than allocated by design.
This matters for at least three reasons. First, the dividend is smaller than the hype suggests. Nationally representative US data put self-reported AI time savings at around 2% of all work hours (Federal Reserve Bank of St. Louis, 2026), and a large Danish study linking surveys to administrative records found average savings of about 3% of work hours among users, with no detectable effect on earnings or recorded hours (Humlum & Vestergaard, 2025). When the dividend is that thin, how it is spent becomes a strategic question rather than an afterthought. Second, the way it is currently spent appears to raise job demands. An eight-month ethnography at a US technology company found that generative AI led employees to take on more tasks, work at a faster pace, and let work seep into breaks and evenings, often without anyone asking them to (Ranganathan & Ye, 2026). Third, the organizations that are treating the dividend deliberately, whether by reinvesting it in higher-value work, in learning, or in shorter hours, are generating some of the most interesting evidence about what sustainable AI-enabled work could look like.
The argument of this article is straightforward. AI does not determine where saved time goes; organizational choices do. Those choices are currently being made implicitly, through performance targets, headcount freezes, and cultural signals about responsiveness. Leaders who want AI to improve both performance and wellbeing need to make the choices explicit. The sections that follow define the core concepts, review what we know about the prevalence and drivers of the trap, examine its consequences for organizations and individuals, and then set out five evidence-based responses illustrated with documented organizational examples. A final section looks at the longer-term capabilities organizations need to govern the time dividend well.
A word about evidence is in order. Much of the most current data on AI and work comes from vendor surveys, including the Jabra, Workday, and Upwork studies cited here. These are useful for tracking sentiment and reported behavior, but they are produced by firms with commercial interests in the future of work, rely on self-report, and rarely publish full questionnaires. Wherever possible I triangulate them with peer-reviewed research, government evaluations, and independent working papers. The overall picture is consistent across sources, which gives some confidence in the direction of the findings even where precise magnitudes remain uncertain.
The Time Dividend Landscape
Defining Key Terms
The time dividend refers to the working time that an individual or team no longer needs to spend on a given task because AI now performs part or all of it. It is a gross figure: the minutes saved on drafting an email, summarizing a meeting, writing boilerplate code, or completing a clinical note. The net dividend is what remains after accounting for the new work AI creates, such as prompting, checking outputs, correcting errors, and coordinating with colleagues who are also using AI. Survey evidence suggests this offset is substantial. In a 2024 survey of 2,500 executives, employees, and freelancers across four countries, 77% of employees who used AI said it had added to their workload, with 39% citing time spent reviewing or moderating AI-generated content and 23% citing time spent learning the tools (Upwork Research Institute, 2024).
Time dividend capture describes the process by which saved time is converted into additional output, usually without explicit agreement. Capture can be organizational, as when targets rise or headcount is held flat because AI is expected to fill the gap, or self-imposed, as when workers voluntarily take on more because the tools make it feel possible. Time dividend reinvestment, by contrast, means deliberately allocating some portion of saved time to activities that build long-term capacity, such as learning, quality improvement, relationship work, innovation, or recovery.
Work intensification is the established term for an increase in the effort required per hour of work, as distinct from extensification, which refers to working longer hours (Green, 2004). Green's analysis of British data showed that work effort rose markedly during the 1980s and 1990s and traced much of the increase to effort-biased technological change, the spread of technologies that raise the pace and density of work rather than simply replacing it (Green, 2004). Kelliher and Anderson (2010) later found that employees given flexible working arrangements often responded by working harder, partly out of a sense of reciprocity toward an employer that had granted them flexibility. Both strands of research are directly relevant to AI, which combines the effort-biased qualities of earlier information technologies with a strong cultural narrative of personal empowerment.
Economists have a related concept. The rebound effect, first described by Jevons (1865) in his observation that more efficient steam engines increased rather than reduced Britain's total coal consumption, captures the tendency for efficiency gains to stimulate additional demand. Applied to work, it suggests that when a task becomes cheaper in time, organizations and individuals will often do more of it, or add adjacent tasks, rather than bank the savings. The Jabra finding that time spent on every category of work grew, including administrative work that AI is supposed to absorb, is consistent with this rebound logic (Jabra, 2026).
Finally, the job demands–resources model provides the theoretical frame for understanding consequences. In this model, job demands such as workload, time pressure, and emotional strain drive exhaustion, while job resources such as autonomy, support, and opportunities for learning drive engagement and buffer the effects of demands (Bakker & Demerouti, 2007). The model builds on Karasek's (1979) insight that high demands are most harmful when combined with low decision latitude. The time dividend can strengthen either side of the ledger. Captured as more volume, it adds to demands. Reinvested in learning, autonomy, or recovery, it adds to resources.
Prevalence, Drivers, and Distribution
How large is the dividend in the first place? The best nationally representative evidence comes from the Real-Time Population Survey tracked by Bick, Blandin, and Deming. Their data show the share of employed US adults using generative AI for work in the previous week rising from 28.2% in the third quarter of 2024 to 39.2% in the second quarter of 2026, and self-reported time savings rising from 1.6% to 2.2% of all work hours over the same period (Federal Reserve Bank of St. Louis, 2026). Their earlier analysis of the same survey documented that generative AI was being adopted faster than personal computers or the internet had been at a comparable stage (Bick et al., 2024). Spread across a whole labor force, 2.2% of hours is economically meaningful, but for an individual it is roughly one hour in a 45-hour week.
The Danish evidence points in a similar direction. Humlum and Vestergaard (2025) surveyed about 25,000 workers in 11 AI-exposed occupations at around 7,000 workplaces and linked their responses to administrative records. Users reported average time savings of about 2.8% of work hours. Crucially, the study also asked what workers did with the time. Roughly 80% said they reallocated it to other work tasks, while fewer than one in ten used it for breaks or leisure. The researchers found no significant effects on earnings or recorded hours and estimated that only a small fraction of productivity gains passed through to workers' pay (Humlum & Vestergaard, 2025). AI chatbots also created new job tasks for some workers, including some who did not use the tools themselves.
Experimental studies show larger effects on specific tasks. In a field study of 5,172 customer support agents, access to a generative AI assistant raised the number of issues resolved per hour by about 15% on average, with the largest gains for novice and lower-skilled workers (Brynjolfsson et al., 2025). In a preregistered experiment with 758 Boston Consulting Group consultants, those using GPT-4 completed 12.2% more tasks and finished 25.1% more quickly on tasks within AI's capabilities, though they were 19% less likely to reach a correct answer on a task deliberately designed to fall outside them (Dell'Acqua et al., 2023). Public sector evidence is similar in scale: in a three-month UK government experiment involving 20,000 civil servants across 12 organizations, users reported saving an average of 26 minutes a day (Government Digital Service, 2025).
Three drivers help explain why this dividend so rarely shows up as reduced hours or lighter workloads.
The first is leader expectations. The Workday data showing that 40% of leaders expect AI to get more out of existing employees is one indicator (Workday, 2026). The Upwork survey is another: 96% of C-suite leaders said they expected AI to increase their company's productivity, and 81% acknowledged they had increased demands on workers in the previous year (Upwork Research Institute, 2024). When productivity gains are already written into plans, saved time is spoken for before it is saved.
The second is task expansion by the worker. Ranganathan and Ye (2026) observed employees using AI to take on work that previously belonged to other roles or would not have been attempted at all. Because AI made unfamiliar tasks feel achievable, people stretched into them voluntarily. This is the rebound effect operating at the level of the individual job.
The third is the collapse of natural boundaries. Microsoft's analysis of anonymized Microsoft 365 telemetry describes what it calls the infinite workday, in which heavily pinged employees are interrupted about every two minutes during core hours, meetings starting after 8 p.m. have risen 16% year over year, and the average worker receives 117 emails a day (Microsoft, 2025). AI did not create this pattern, but it amplifies it by making it trivially easy to continue a task during a lunch break or at the end of the day, a dynamic Ranganathan and Ye (2026) documented directly.
The dividend and its burdens are not evenly distributed. The Jabra data show that heavy AI use is far more common among senior leaders (30%) than among employees without management responsibilities (8%), and that 64% of non-managers had received no AI training or only a brief introduction, compared with 26% of senior leaders (Jabra, 2026). Leaders are therefore more likely to experience AI as a source of leverage and to extrapolate their own experience to the workforce, while front-line employees experience the expectations without the same support. In healthcare, the AI scribe evidence shows that benefits were concentrated among primary care physicians, advanced practice clinicians, female clinicians, and frequent users (Rotenstein et al., 2026). Those distributional patterns matter because a dividend that accrues mainly to some groups, while expectations rise for everyone, is a recipe for strain and resentment.
Organizational and Individual Consequences
Organizational Performance Impacts
At first glance, capturing the time dividend as additional output looks like straightforward good management. If a team can produce the same work in less time, why not ask it to produce more? The evidence suggests three reasons for caution.
The first is that captured gains may be smaller and more fragile than they look. Self-reported time savings are, as the St. Louis Fed researchers note, inherently approximate (Federal Reserve Bank of St. Louis, 2026). The UK government's Copilot experiment was candid that its time-savings estimates were self-reported, that it was not possible to identify how saved time was spent, and that the tool struggled with complex or nuanced work that still required human oversight (Government Digital Service, 2025). When leaders bank an estimated gain by raising targets, they may be banking a number that partly reflects enthusiasm rather than measured change.
The second is that output gains can come at the expense of quality. The BCG experiment showed that the same tool that sped up some tasks led to worse answers on a task outside its capabilities, a pattern the authors described as a jagged technological frontier (Dell'Acqua et al., 2023). Brynjolfsson et al. (2025) found that the most experienced and skilled agents gained little in speed and saw small declines in quality. If saved time is immediately converted into more volume, there is less capacity left for the checking, judgment, and reflection that keep quality high in exactly those situations where AI is weakest.
The third, and perhaps the most important, is that capture can erode the very human capacity organizations need to adapt. Organizations implementing AI need people with the time and energy to redesign workflows, experiment, share what works, and learn new skills. Workday's data show a 13-point gap between the share of workers who know what skills they need (79%) and the share who say their employer actually helps them develop them (66%), along with falling rates of internal mobility at a majority of employers (Workday, 2026). When the time dividend is consumed entirely by current output, there is no slack left for the learning and reconfiguration that sustain future performance.
Klarna offers a cautionary illustration. In early 2024 the company announced that its AI assistant was handling two-thirds of customer service chats, doing the equivalent work of 700 full-time agents, and was expected to improve profits by an estimated $40 million that year (CX Today, 2024). Barely a year later, chief executive Sebastian Siemiatkowski told Bloomberg that cost had been too predominant a factor in the decision and that the result was lower quality, and the company began recruiting human agents again under a model that routes customers to AI first with the option of reaching a person (Computing, 2025). Klarna's case is about headcount rather than time, but it captures the underlying risk. Treating the efficiency dividend purely as a cost saving can trade away quality and customer trust that later have to be rebuilt.
The healthcare evidence adds nuance. In a multisite study of 8,581 ambulatory clinicians at five US academic health systems, adoption of AI scribes was associated with 16 fewer minutes of documentation time per eight scheduled patient hours, a modest reduction in total electronic health record time, and about half an additional patient visit per week, with no statistically significant change in work outside scheduled hours (Rotenstein et al., 2026). Part of the time dividend, in other words, was captured as visit volume and an estimated modest increase in revenue, while after-hours work was largely unchanged. That may be a reasonable trade in a system with access shortages, but it illustrates how quickly saved time is reabsorbed into throughput unless someone decides otherwise.
Individual Wellbeing and Stakeholder Impacts
For individuals, the central risk of the time dividend trap is that AI raises job demands while appearing to reduce them. The job demands–resources model predicts that sustained increases in workload and time pressure, without matching increases in resources, lead to exhaustion and disengagement (Bakker & Demerouti, 2007). The survey evidence is consistent with that prediction. In the Upwork study, 71% of full-time employees reported burnout, 65% said they struggled with their employer's productivity demands, and 47% of AI-using employees said they did not know how to achieve the productivity gains their employers expected (Upwork Research Institute, 2024). These are vendor survey figures and should be read as indicators rather than population estimates, but they align with qualitative findings.
Ranganathan and Ye's (2026) ethnography is especially revealing because it captures how intensification feels from the inside. Employees described short bursts of AI-assisted work as energizing, but on reflection felt busier and less able to disconnect. Constant switching between parallel AI processes, reviews, and meetings left little room for recovery, and the authors cautioned that self-driven extra effort can quietly become a new baseline expectation. This echoes earlier research on mobile email. Mazmanian et al. (2013) described an autonomy paradox, in which professionals who used mobile devices to gain flexibility ended up feeling obliged to be available at all times. Barley et al. (2011) similarly found that email functioned as both a source and a symbol of overload. AI can play the same role: a tool experienced as empowering at the level of individual tasks that collectively erodes boundaries.
Recovery research explains why this matters. Sonnentag and Fritz (2007) showed that psychological detachment from work during off-job time, along with relaxation, mastery experiences, and control, is central to recovering from work demands. When AI makes it frictionless to continue working during breaks and evenings, it attacks exactly the detachment that protects wellbeing. Perlow's (1999) classic study of a software engineering team showed how a "time famine" of constant interruption reduced both wellbeing and productivity, and how a modest restructuring of quiet time improved both. The parallels with AI-enabled multitasking are hard to miss.
There is a more hopeful side. Where the dividend has been deliberately directed toward relieving the most draining tasks, as with ambient AI scribes in clinical documentation, studies have reported meaningful reductions in burnout and improvements in work satisfaction (Olson et al., 2025; Permanente Medicine, 2025). These results, discussed further below, suggest that the same technology can reduce strain when it targets the work people find most depleting and when the saved effort is not immediately reallocated.
The effects also extend beyond employees. Patients in the Permanente evaluation reported that their physicians spent less time looking at a computer and more time speaking directly to them (Permanente Medicine, 2025). In customer-facing settings, Brynjolfsson et al. (2025) found that customers were more polite and less likely to escalate when agents had AI support. Conversely, Klarna's experience suggests that customers notice when efficiency is pursued at the expense of quality (Computing, 2025). How the time dividend is spent is therefore not only an internal human resources question. It shapes the experience of patients, customers, and citizens.
Evidence-Based Organizational Responses
The research reviewed above suggests that escaping the time dividend trap requires more than encouraging people to "use the time well." It requires organizations to make deliberate decisions about how saved time is measured, divided, protected, and used. Table 1 summarizes five responses, the evidence behind them, key practices, and illustrative organizational examples. Each is discussed in turn.
Table 1: Evidence-based responses to the time dividend trap
Response | Evidence Base | Key Practices | Illustrative Examples |
Make the dividend visible | Self-reported savings are approximate; saved time is usually reallocated invisibly (Humlum & Vestergaard, 2025; Government Digital Service, 2025) | Measure net, not gross, savings; track where time goes; report quality alongside speed | UK civil service Copilot experiment; five-system AI scribe study |
Agree explicit reinvestment rules | Leaders plan to capture gains; workers lack clarity on expectations (Workday, 2026; Upwork Research Institute, 2024) | Set a capture/reinvest split; link headcount decisions to measured gains; communicate openly | JPMorgan Chase; Shopify; Klarna |
Convert time into recovery | Reduced-hours trials improve wellbeing without performance loss (Fan et al., 2025; Lewis et al., 2023) | Shorter weeks paired with work redesign; meeting limits; protected detachment | South Cambridgeshire District Council; Microsoft Japan; Convictional; Peak PEO; The Ross Firm |
Reinvest in learning and judgment | AI accelerates learning but quality falls outside its frontier (Brynjolfsson et al., 2025; Dell'Acqua et al., 2023) | Protected learning time; peer sharing; time for review on high-stakes work | Boston Consulting Group experiment; customer support deployment |
Redesign the rhythm of AI-enabled work | AI-driven multitasking erodes boundaries and recovery (Ranganathan & Ye, 2026; Sonnentag & Fritz, 2007) | Intentional pauses, sequencing, focus windows, human check-ins; target draining tasks first | Permanente Medical Group; six-system ambient scribe study; Berkeley Haas field site |
Make the Time Dividend Visible
The first step in escaping the trap is simply knowing how big the dividend is and where it goes. Most organizations currently know neither. They have adoption metrics (licenses issued, prompts sent) and occasional self-reported time-savings surveys, but rarely measure net savings after accounting for review and rework, and almost never track how the saved time is used. Humlum and Vestergaard's (2025) finding that about 80% of saved time is reallocated to other tasks is a population-level estimate. Inside a given organization, the reallocation is usually invisible, which makes it impossible to manage.
The evidence also suggests caution about the measures themselves. Self-reported savings can diverge from measured savings, and gross savings on one task can be offset by new work elsewhere (Federal Reserve Bank of St. Louis, 2026; Upwork Research Institute, 2024). The multisite AI scribe study is a useful model because it used objective electronic health record logs rather than self-report and compared adopters with nonadopters at the same institutions (Rotenstein et al., 2026). Not every organization can replicate that design, but the principle of measuring time with behavioral data where possible, and triangulating it with experience surveys, applies widely.
Effective approaches:
Net, not gross, savings: Estimate time saved after subtracting time spent prompting, reviewing, correcting, and coordinating AI-generated work.
Time-use sampling: Use short, periodic pulse surveys or diary studies asking where recovered time actually went, using categories such as more of the same work, new tasks, learning, collaboration, and breaks.
Behavioral data where available: Use system logs, ticket data, or calendar analytics to complement self-report, with clear privacy safeguards and aggregation.
Quality alongside speed: Track error rates, rework, customer or patient outcomes, and escalations so that speed gains are not mistaken for productivity gains.
Distributional reporting: Break results down by role, level, and team so that uneven dividends and uneven burdens are visible.
The UK Government Digital Service ran one of the largest public evaluations of an AI productivity tool to date. Between September and December 2024, 20,000 civil servants across 12 organizations received Microsoft 365 Copilot licenses, and more than 7,000 responded to surveys. Users reported saving an average of 26 minutes a day, which the report extrapolated to roughly 13 working days a year, and 82% said they would not want to return to working without it (Government Digital Service, 2025). The report is notable for its honesty about limitations. It acknowledged that the savings were self-reported, that it was not possible to identify how the saved time was spent, and that more than 70% of users felt the tool reduced time on mundane tasks and increased time on strategic work, without objective data to confirm that shift. The experiment shows both the value of measuring the dividend and the gap that remains when organizations stop at measuring time saved rather than time reallocated.
The five academic health systems in the Ambient Clinical Documentation Collaborative, co-led by Mass General Brigham and UCSF, took a more rigorous approach. Rather than relying on clinicians' estimates, the researchers used electronic health record audit logs for more than 1,800 AI scribe adopters and more than 6,700 comparison clinicians. They found modest reductions in documentation time, a small increase in visits, and no significant change in after-hours work, with larger benefits for frequent users (Rotenstein et al., 2026). The senior author noted publicly that such modest documentation reductions were unlikely to fully account for the burnout improvements seen in other studies, and that further research would examine where the freed time went (Mass General Brigham, 2026). For health system leaders, the study provides an evidence base for deciding deliberately whether the dividend should go to access, to clinician recovery, or to both.
A common pitfall is treating vendor-reported or self-reported time savings as a budgetable asset. When estimated minutes are converted into targets or headcount plans before net savings and quality effects are understood, organizations lock in assumptions that may not hold and remove the slack needed to discover where AI actually helps.
Agree Explicit Reinvestment Rules
Once the dividend is visible, the next question is who gets it. At present the answer is usually decided implicitly by targets, budget cycles, and headcount freezes. The Workday and Upwork data suggest that leaders often intend to capture gains as additional output, while employees are unclear about what is expected of them (Workday, 2026; Upwork Research Institute, 2024). The gap between those perspectives is itself a source of strain. In the Upwork survey, 40% of AI-using employees felt their company was asking too much of them regarding AI (Upwork Research Institute, 2024).
An explicit reinvestment rule is a simple, public commitment about how the dividend will be shared. It might specify that a portion of measured time savings will be reinvested in learning or improvement projects, that targets will not be raised until gains are verified, or that headcount decisions will be based on demonstrated rather than projected savings. The logic draws on research on the psychological contract and on Kelliher and Anderson's (2010) finding that employees often reciprocate perceived generosity with extra effort. If organizations make the bargain explicit, they can harness that reciprocity rather than relying on silent intensification.
Effective approaches:
A declared split: Announce a default division of verified time savings among output, improvement work, learning, and recovery, and revisit it periodically.
Verified before banked: Raise targets or adjust staffing plans only after net savings have been measured over a meaningful period.
Redeployment over reduction: Where roles shrink, prioritize moving people toward higher-value work, supported by reskilling.
Shared upside: Consider linking some portion of verified productivity gains to pay, development, or time, given evidence of very low pass-through to earnings (Humlum & Vestergaard, 2025).
Transparent communication: Explain plainly what AI is expected to change about workloads, roles, and hiring, rather than leaving employees to infer it.
JPMorgan Chase illustrates a redeployment-oriented approach. In March 2025, the bank's global chief information officer, Lori Beer, told Reuters that its in-house AI coding assistant had made tens of thousands of software engineers 10% to 20% more productive. Rather than framing the gain as a headcount reduction, she described the freed capacity as an opportunity to reassign engineers to other projects and to spend more time on high-value work in AI and data (Reuters, 2025). Redeployment is still a form of capture, since the time goes back into work, but it is an explicit choice to direct the dividend toward strategic priorities rather than simply more of the same output.
Shopify took a different, more output-oriented position. In a memo to staff in spring 2025, chief executive Tobi Lütke described effective AI use as a fundamental expectation of everyone at the company, said that teams must demonstrate why they cannot get what they want done using AI before requesting more headcount or resources, and added questions about AI use to performance and peer reviews (Fortune, 2025). Whatever one thinks of the policy, it has the virtue of clarity. Employees know that the dividend is expected to fund growth without proportional hiring. The risk, from a job demands–resources perspective, is that clarity about expectations is not matched by clarity about resources, such as time to learn and limits on scope creep.
Klarna, discussed earlier, shows what can happen when the capture decision is driven primarily by cost. Its later acknowledgment that cost had been too predominant a factor, and its decision to rebuild human customer service capacity, is a reminder that reinvestment rules should be evaluated against quality and customer outcomes, not only against savings (Computing, 2025).
A common pitfall is announcing an AI strategy that is explicit about expected gains but silent about how employees will share in them. Silence is not neutral. In the absence of a stated rule, workers reasonably assume the default is full capture, and many respond by intensifying their own effort to demonstrate value.
Convert Time into Recovery
The most radical way to spend the time dividend is to give some of it back as time off. Until recently, reduced-hours experiments were a separate conversation from AI adoption, but the two are converging. The evidence from four-day-week trials is relevant because it shows what happens when organizations deliberately bank productivity gains as shorter hours rather than more output, provided they first redesign work.
The strongest evidence comes from Fan et al. (2025), who analyzed 2,896 employees in 141 organizations across six countries that moved to a four-day week without loss of pay, compared with 12 control companies. Organizations spent about two months reorganizing work before a six-month trial. Employees in trial organizations showed improvements in burnout, job satisfaction, and mental and physical health, and those whose individual hours fell the most showed the largest gains. Improved work ability, fewer sleep problems, and less fatigue helped explain the effects (Fan et al., 2025). The earlier UK pilot involving 61 companies and around 2,900 workers reached similar conclusions: 56 companies continued with the four-day week after the trial, 71% of employees reported lower burnout, and revenue was broadly stable over the trial period (Lewis et al., 2023).
These trials were not about AI, and their samples were self-selected. But they demonstrate the core mechanism that matters for the time dividend: when organizations commit in advance to converting efficiency into time, they redesign work to make it possible, and wellbeing improves without obvious performance costs.
Effective approaches:
Redesign first: Pair any reduction in hours with a structured period of work redesign, as in the Fan et al. (2025) trials, rather than simply compressing existing work.
Use AI to fund the change: Identify tasks that AI can reliably absorb and earmark those savings for reduced hours or protected time.
Meeting discipline: Shorten and cap meetings, which reduced-hours pioneers consistently cite as an early source of time.
Coverage models: Use rotating on-call or staggered schedules so customers and colleagues are not left without support.
Protect detachment: Set norms limiting after-hours messaging and AI-enabled "just one more task" work, consistent with recovery research (Sonnentag & Fritz, 2007).
South Cambridgeshire District Council provides the most substantial public sector example. Beginning in January 2023, the council trialed a 32-hour week, evaluated independently by the Universities of Salford and Cambridge. Over 15 months, 11 of 24 performance indicators improved, 11 showed no significant difference, and two declined. Staff turnover fell by 39%, job advertisements attracted 53% more applications on average, and the council estimated annual savings of £371,500 from reduced turnover and staffing costs (Civil Service World, 2024). The trial continued despite a formal warning from the central government department responsible for local government. It was not an AI initiative, but it shows that a public body can convert efficiency into time without degrading services when it measures carefully.
Microsoft Japan offered an earlier private sector illustration. In August 2019 it closed its offices on Fridays as part of a work-life choice experiment, gave full-time staff paid leave, and encouraged 30-minute meetings and remote communication. Compared with the same month in 2018, labor productivity rose 39.9%, electricity use fell 23.1%, and pages printed fell 58.7% (World Economic Forum, 2019). The single-month design limits what can be concluded, but the experiment showed how a fixed reduction in time can force the elimination of low-value work.
More recently, small firms have begun explicitly linking AI to shorter weeks. Convictional, a 12-person software startup, moved to a four-day week without cutting pay, with one employee on call each week on a rotating basis, and its chief executive attributed part of the change to productivity gains from AI-assisted work (Peck, 2025). Peak PEO, a 20-person London professional employer organization that uses AI for invoices, document creation, and sales analysis, moved from flexible Fridays to a four-day week and reported that job openings that had attracted two applications later drew 350 (Anchorage Daily News, 2025). The Ross Firm, a 50-person Ontario law firm, combined a four-day, eight-hour-day schedule with tools for faster research, billable-hour tracking, and call summaries (Anchorage Daily News, 2025). These are journalistic accounts of small firms rather than controlled studies, but they show the time dividend being consciously directed toward recovery.
A common pitfall is compressing five days of unchanged work into four. Without redesign, shorter weeks simply intensify the remaining days, recreating the trap in a new form. The research consistently shows that the benefits depend on removing work, not squeezing it.
Reinvest in Learning and Judgment
A second form of reinvestment directs saved time toward building human capability. This is especially important because the evidence shows that AI's benefits are uneven and that quality can suffer when people over-rely on it in areas where it is weak. The BCG experiment found that consultants using AI were less likely to reach correct answers on a task outside AI's frontier (Dell'Acqua et al., 2023). Learning where that frontier lies, and building the judgment to recognize it, takes time that the trap removes.
There is also evidence that AI can accelerate learning when people have room to use it that way. Brynjolfsson et al. (2025) found that AI assistance helped novice customer support agents improve more quickly, with gains concentrated on less common problems where agents had less experience, and with improvements in agents' English fluency. The tool functioned partly as a vehicle for spreading the tacit knowledge of top performers. Yet the Jabra data suggest that most front-line employees have received little structured AI training, with 64% of non-managers reporting no training or only a brief introduction (Jabra, 2026). Workday's finding of a 13-point gap between knowing what skills are needed and receiving employer support to develop them points in the same direction (Workday, 2026).
Effective approaches:
Protected learning time: Earmark a fixed share of verified time savings for structured learning, experimentation, and practice.
Peer knowledge sharing: Create regular forums where employees share effective uses, failures, and frontier boundaries, so learning compounds across teams.
Review capacity for high-stakes work: Preserve time for human checking on complex, ambiguous, or consequential tasks where AI is most likely to err.
Developmental use of AI: Encourage people to use AI as a tutor or coach on unfamiliar tasks, not only as a production tool.
Learning metrics: Track skill growth, internal mobility, and quality outcomes as returns on the time dividend, not just output volume.
Boston Consulting Group provided the setting for one of the most influential studies of AI and knowledge work. In a preregistered field experiment run with academic researchers, 758 of the firm's consultants were randomly assigned to work without AI, with GPT-4, or with GPT-4 plus a prompt engineering overview on realistic consulting tasks. The results showed large speed and quality gains on tasks within AI's capabilities but worse performance on a task outside them (Dell'Acqua et al., 2023). The lesson for reinvestment is direct. Organizations that use part of the time dividend to help people map the frontier, by testing, comparing, and sharing results, reduce the risk that efficiency gains are offset by errors in exactly the situations that matter most.
The customer support deployment studied by Brynjolfsson et al. (2025) shows the learning potential on the other side of the frontier. Novice agents with AI access progressed faster than those without it, in effect compressing part of the learning curve. The organizational implication is that some of the time dividend for experienced workers could be reinvested in coaching and in refining the knowledge that AI systems distribute to newer colleagues, rather than simply in handling more tickets.
A common pitfall is assuming that people will learn on the job in the gaps AI creates. In practice, as the Danish and Jabra data suggest, those gaps are filled by more work almost immediately. Learning that is not protected in the calendar tends not to happen.
Redesign the Rhythm of AI-Enabled Work
The final response addresses the texture of work rather than its volume. Even when total hours are stable, AI can intensify work by increasing switching, multitasking, and boundary erosion. Ranganathan and Ye (2026) propose that organizations develop what they call an AI practice: a set of deliberate norms covering the rhythm and boundaries of AI-enabled work. Their recommendations include intentional pauses before major decisions, sequencing work to batch non-urgent updates and protect focus windows, and human grounding through protected time for check-ins and shared reflection.
These recommendations align with longstanding evidence. Perlow (1999) showed that coordinated quiet time improved software engineers' productivity. Recovery research emphasizes detachment and control over one's time (Sonnentag & Fritz, 2007). Microsoft's telemetry data on interruptions and after-hours work suggest that many organizations are moving in the opposite direction (Microsoft, 2025). Redesigning rhythm also means being selective about where AI is applied first. The clearest wellbeing gains have come where AI absorbed the most depleting tasks, as with clinical documentation, rather than where it simply sped up everything.
Effective approaches:
Target draining work first: Prioritize AI use cases that remove tasks employees experience as most depleting or least meaningful.
Intentional pauses: Build brief reflection points into AI-assisted decisions to surface counterarguments and check alignment with goals (Ranganathan & Ye, 2026).
Sequencing and batching: Group AI-generated updates and reviews into defined windows rather than allowing continuous streams.
Focus protection: Establish shared focus blocks and limits on the number of parallel AI processes or agents people supervise at once.
Human grounding: Protect time for conversation, reflection, and relationship-building, which AI-enabled speed tends to crowd out.
The Permanente Medical Group in Northern California offers a large-scale example of targeting draining work. Between October 2023 and December 2024, 7,260 of its physicians used ambient AI scribes across more than 2.5 million patient encounters, saving an estimated nearly 16,000 hours of documentation time (Kaiser Permanente Division of Research, 2025). In surveys, 82% of physicians reported a positive effect on overall work satisfaction and 84% on patient interactions, while 47% of patients said their physician spent less time looking at the computer and 39% said the physician spent more time speaking directly to them (Permanente Medicine, 2025). The dividend here was reinvested in the relational core of clinical work rather than in additional throughput alone.
The six health systems in the study led by Olson et al. (2025) showed how quickly targeted relief can affect wellbeing. After 30 days of using an ambient scribe, the share of physicians and advanced practice practitioners reporting burnout fell from 51.9% to 38.8%, and participants reported lower cognitive load for note-related tasks and less after-hours documentation. The study's pre-post design limits causal claims, and the multisite findings of Rotenstein et al. (2026) suggest that measured time savings were modest. Taken together, the two studies imply that reducing the cognitive burden of a draining task may matter as much as the minutes saved.
The 200-person US technology company where Ranganathan and Ye (2026) conducted their ethnography illustrates the opposite dynamic. With broad access to generative AI and no explicit norms about its use, employees expanded their scope, ran several AI processes in parallel, and let work spill into breaks and evenings. No one mandated this. It emerged from a combination of enthusiasm, ambiguity about expectations, and tools that made continuing always easier than stopping.
A common pitfall is assuming that because AI use is voluntary, the intensification it produces is benign. Self-initiated overwork is still overwork, and once it becomes visible to colleagues and managers, it tends to reset everyone's expectations.
Building Long-Term Time Dividend Governance
The five responses above address immediate choices. Over the longer term, organizations need durable capabilities to govern the time dividend as AI becomes more capable and more pervasive.
Treating the Time Dividend as a Strategic Asset
Organizations routinely govern financial capital through budgets, investment committees, and return metrics. Time freed by AI is a form of capital, yet it is rarely governed at all. Treating it as a strategic asset means assigning ownership, setting allocation principles, and tracking returns. A senior leader or cross-functional group might be responsible for an annual time dividend review that examines measured savings, how they were allocated, and what each allocation produced in output, quality, capability, and wellbeing.
This approach also helps resolve the tension between short-term and long-term returns. Capturing the dividend as output produces immediate, visible gains. Reinvesting it in learning, improvement, or recovery produces returns that are slower and harder to measure but that underpin future adaptability. The rebound logic Jevons (1865) identified applies here: unless someone deliberately decides to bank efficiency gains, they will be consumed by expanded demand. Governance is how that decision gets made on purpose.
Building Sustainable Performance Norms
The time dividend trap is ultimately a problem of norms. When AI makes it possible to do more, the definition of "enough" shifts upward. Green's (2004) analysis of effort-biased technological change and Kelliher and Anderson's (2010) research on reciprocal intensification both show how performance expectations can ratchet upward without any formal decision. Sustainable performance norms counter this by defining what good work looks like in terms of outcomes and quality rather than volume and responsiveness.
Building such norms involves several practices. Performance frameworks should reward judgment, collaboration, and improvement work, not just throughput. Leaders should model boundaries by not sending AI-assisted messages at all hours and by visibly using time for learning and recovery. Teams can agree on explicit service levels for response times so that availability expectations do not expand to fill every gap AI creates. The Fan et al. (2025) finding that individual reductions in hours, rather than company-level policies alone, drove wellbeing gains is a reminder that norms must change in daily practice, not just on paper.
The psychological contract also needs attention. Employees who see their organization capture every efficiency gain while offering little in return are unlikely to sustain discretionary effort or to share innovations that might make their own roles more efficient. Conversely, organizations that share the dividend visibly build trust that encourages people to surface better uses of AI. The low pass-through of productivity gains to earnings in the Danish data suggests that, at least so far, most of the dividend is being retained by employers (Humlum & Vestergaard, 2025). Whether that balance is sustainable as AI becomes more capable is a question leaders should be asking now.
Distributing Stewardship Across the Organization
Finally, time dividend governance cannot be purely top-down. The most consequential allocation decisions happen in teams, in the daily choices about whether to take on another project, run another analysis, or leave at a reasonable hour. Equipping team leaders and employees to make those choices well is essential.
At the team level, this might involve regular retrospectives that ask not only what AI helped accomplish but also what happened to the time it saved, and whether that allocation matched the team's priorities. Teams can set their own reinvestment targets within organizational principles, deciding for example to use some savings for a quarterly improvement project or for protected focus time. The Jabra evidence that workers with AI-capable managers are far more likely to use AI daily and report greater career optimism suggests that local leadership shapes how AI is experienced (Jabra, 2026), and the same is likely true of how its dividends are spent.
At the individual level, people benefit from frameworks for recognizing intensification in their own work: noticing when AI-enabled scope expansion is crowding out recovery, when parallel tasks are degrading attention, or when saved time is disappearing without a conscious choice. Ranganathan and Ye's (2026) observation that AI-assisted bursts felt energizing in the moment but exhausting in aggregate is a pattern many workers will recognize. Naming it is the first step toward managing it.
Conclusion
The promise that AI will give people time back is not false, but it is incomplete. The best available evidence suggests that AI saves a modest share of working time on average, with larger gains on specific tasks, and that this time is overwhelmingly reabsorbed into more work. Workers broaden their roles, leaders raise expectations, boundaries erode, and new tasks emerge to fill the space. Hours do not fall, and in some settings the pace and density of work rise. This is the time dividend trap: a pattern of default capture that turns a potential resource into an additional demand.
The trap is not inevitable. The time dividend is an organizational choice, even when it is made by omission. Organizations that measure the dividend honestly, agree openly on how it will be shared, protect some of it for recovery and learning, and redesign the rhythm of AI-enabled work are showing that AI can reduce strain while sustaining performance. Health systems that have directed AI toward the most draining clinical tasks have seen meaningful reductions in burnout. Public and private organizations that have deliberately converted efficiency into time have maintained performance while improving wellbeing and retention. And firms that have redeployed rather than simply eliminated capacity are positioning themselves for the adaptation that AI will continue to demand.
The broader lesson is that technology does not decide how its benefits are distributed. People do. Leaders who want AI to deliver durable value need to treat freed time with the same seriousness they bring to freed capital, making allocation decisions explicitly, evaluating their returns, and recognizing that the most valuable returns may come from investments in people rather than in volume.
Measure net, not gross: Estimate time savings after accounting for review, correction, and coordination, and track where saved time actually goes.
Decide before you bank: Agree and communicate an explicit rule for how verified savings will be split among output, improvement, learning, and recovery.
Redesign before you compress: Pair any reduction in hours with genuine work redesign, so that shorter weeks remove work rather than intensify it.
Protect learning time: Reserve a fixed share of the dividend for structured learning and frontier-mapping, since unprotected learning time disappears.
Target the draining work: Direct AI first at the tasks people find most depleting, where wellbeing gains are largest.
Guard the boundaries: Set norms for pacing, multitasking, and after-hours AI use that preserve detachment and recovery.
Govern the dividend: Assign ownership, report allocations and returns, and give teams a voice in how saved time is used.
Research Infographic
References
Anchorage Daily News. (2025, December 31). These companies say AI is key to their four-day workweeks. Anchorage Daily News.
Bakker, A. B., & Demerouti, E. (2007). The job demands-resources model: State of the art. Journal of Managerial Psychology, 22(3), 309–328.
Barley, S. R., Meyerson, D. E., & Grodal, S. (2011). E-mail as a source and symbol of stress. Organization Science, 22(4), 887–906.
Bick, A., Blandin, A., & Deming, D. J. (2024). The rapid adoption of generative AI (NBER Working Paper No. 32966). National Bureau of Economic Research.
Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942.
Civil Service World. (2024, July 9). Public sector's biggest four-day work week trial a "win-win-win". Civil Service World.
Computing. (2025, May 13). Meat back on the menu as Klarna rows back on AI. Computing.
CX Today. (2024, February 29). Klarna claims its new AI assistant does the work of 700 full-time agents. CX Today.
Dell'Acqua, F., McFowland, E., III, Mollick, E., Lifshitz-Assaf, H., Kellogg, K. C., 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 (Harvard Business School Working Paper No. 24-013). Harvard Business School.
Fan, W., Schor, J. B., Kelly, O., & Gu, G. (2025). Work time reduction via a 4-day workweek finds improvements in workers' well-being. Nature Human Behaviour, 9(10), 2153–2168.
Federal Reserve Bank of St. Louis. (2026, August 27). Does generative AI save time at work? FRED Blog.
Fortune. (2025, April 8). Shopify CEO says no new hires without proof AI can't do the job. Fortune.
Government Digital Service. (2025). Microsoft 365 Copilot experiment: Cross-government findings report. UK Government.
Green, F. (2004). Why has work effort become more intense? Industrial Relations, 43(4), 709–741.
Humlum, A., & Vestergaard, E. (2025). Large language models, small labor market effects (Becker Friedman Institute Working Paper No. 2025-56). University of Chicago.
Jabra. (2026). WorkLife 2026: The new shape of work. Jabra Research Institute.
Jevons, W. S. (1865). The coal question: An inquiry concerning the progress of the nation, and the probable exhaustion of our coal-mines. Macmillan.
Kaiser Permanente Division of Research. (2025). AI-assisted notetaking gains steady support from Kaiser Permanente physicians. Kaiser Permanente.
Karasek, R. A. (1979). Job demands, job decision latitude, and mental strain: Implications for job redesign. Administrative Science Quarterly, 24(2), 285–308.
Kelliher, C., & Anderson, D. (2010). Doing more with less? Flexible working practices and the intensification of work. Human Relations, 63(1), 83–106.
Lewis, K., Stronge, W., Kellam, J., & Kikuchi, L. (2023). The results are in: The UK's four-day week pilot. Autonomy.
Mass General Brigham. (2026, April 1). AI scribes linked to modest reductions in EHR documentation time [Press release].
Mazmanian, M., Orlikowski, W. J., & Yates, J. (2013). The autonomy paradox: The implications of mobile email devices for knowledge professionals. Organization Science, 24(5), 1337–1357.
Microsoft. (2025). Breaking down the infinite workday (Work Trend Index special report). Microsoft WorkLab.
Olson, K. D., Meeker, D., Troup, M., Barker, T. D., Nguyen, V. H., Manders, J. B., Stults, C. D., Jones, V. G., Shah, S. D., Shah, T., & Schwamm, L. H. (2025). Use of ambient AI scribes to reduce administrative burden and professional burnout. JAMA Network Open, 8(10), e2534976.
Peck, E. (2025, June 29). The four-day workweek gets a new booster: AI. Axios.
Perlow, L. A. (1999). The time famine: Toward a sociology of work time. Administrative Science Quarterly, 44(1), 57–81.
Permanente Medicine. (2025, April 7). Analysis: AI scribes save physicians time, improve patient interactions and work satisfaction. The Permanente Federation.
Ranganathan, A., & Ye, X. M. (2026, February 9). AI doesn't reduce work: It intensifies it. Harvard Business Review.
Reuters. (2025, March 13). JPMorgan engineers' efficiency jumps as much as 20% from using coding assistant. Reuters.
Rotenstein, L. S., Holmgren, A. J., Thombley, R., Sriram, A., Dbouk, R. H., Jost, M., Aizenberg, D., MacDonald, S., Kanaparthy, N., Williams, B., Hsiao, A., Schwamm, L., Murray, S., Byron, M., You, J. G., Centi, A. J., Iannaccone, C., Frits, M., Landman, A. B., . . . Mishuris, R. G. (2026). Changes in clinician time expenditure and visit quantity with adoption of artificial intelligence-powered scribes: A multisite study. JAMA.
Sonnentag, S., & Fritz, C. (2007). The Recovery Experience Questionnaire: Development and validation of a measure for assessing recuperation and unwinding from work. Journal of Occupational Health Psychology, 12(3), 204–221.
Upwork Research Institute. (2024). From burnout to balance: AI-enhanced work models. Upwork.
Workday. (2026, October 5). Workday global workforce report: AI is rewriting jobs more than it's cutting them [Press release].
World Economic Forum. (2019, November 6). Microsoft Japan introduced a four-day workweek and the benefits went far beyond a long weekend. World Economic Forum.
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 Time Dividend Trap: Why AI's Time Savings Are Intensifying Work, and How Leaders Can Reinvest Them Wisely. Human Capital Leadership Review, 40(1). doi.org/10.70175/hclreview.2020.40.1.6
