The New Shape of Work: What Changed Between 2025 and 2026

Every few years the way people work shifts enough that the old vocabulary stops describing it. The year between the middle of 2025 and the middle of 2026 felt like one of those periods, and it left me with a practical problem when I sat down to write about it. Almost everything said on the subject falls into one of two camps. In the first, artificial intelligence is about to remove large parts of the white collar workforce, and every layoff announcement is evidence. In the second, the technology is overhyped, the data show no disruption, and nothing important has changed. Having read the evidence available up to the middle of June 2026, I think both camps are wrong in instructive ways. Work has changed a great deal, but not mostly in the place where the headlines look, which is the total number of jobs. It has changed in the composition of tasks inside jobs, in who is hired and at what level, in what different kinds of workers are paid, in what managers actually do, and in where the benefits of higher output end up. This study works through those layers one at a time. I have drawn on primary publications where I could, such as the Stanford AI Index, the PwC Global AI Jobs Barometer, Microsoft's Work Trend Index, the Bureau of Labor Statistics and the monthly Challenger layoff reports, and I have flagged claims that come from vendor surveys or from secondary summaries that repeat one another. Where I offer an opinion or a prediction, I say so.

Start with what the large studies say about adoption, because it sets the scale. The Stanford AI Index for 2026 reports that 88 percent of surveyed organisations now use artificial intelligence in at least one business function, up from 71 percent a year earlier, and that generative AI reached roughly 53 percent adoption among the population within three years of its mass market debut, faster than the personal computer or the internet. Four in five university students now report using it. At the same time, the Index finds that deployment of AI agents, the software that carries out multi step tasks on its own, remains in the single digits across nearly all business functions. McKinsey's survey from the previous November tells a similar story: almost every organisation uses AI somewhere, roughly a third have begun to scale it across the enterprise, and only about six percent qualify as high performers who attribute more than five percent of their earnings before interest and tax to it. Adoption, in other words, is nearly universal and shallow, and the interesting variation is in depth.

Where depth exists, the measured gains are real but concentrated. The Stanford Index summarises studies finding productivity improvements of roughly 14 to 15 percent in customer support and 26 percent in software development, with larger gains reported in marketing output, and it emphasises that gains are biggest in structured work where the output is easy to monitor and smaller, sometimes negative, in tasks that demand deeper reasoning or judgement. It also notes emerging evidence that heavy reliance on AI may carry learning penalties that slow skill development over time, a point I will return to when I discuss beginners. This pattern, large gains in measurable tasks and unclear gains elsewhere, explains much of what follows. It means the technology reshapes work unevenly, task by task, and that the effect on any particular job depends on the share of its tasks that are structured, repeatable and checkable.

The largest single piece of evidence on how this is playing out in the labour market is PwC's 2026 Global AI Jobs Barometer, which analysed more than a billion job advertisements across 27 countries and territories, combined with company financial data and occupational task data. Its headline findings are worth stating carefully. The average wage premium for jobs requiring AI skills rose to 62 percent, from 57 percent in the previous edition. Jobs requiring specific AI skills grew about 69 percent since 2019, against 9 percent for the jobs market as a whole. Companies in the most AI exposed sectors recorded productivity growth of 34 percent in 2025 relative to 2018, against 24 percent for the least exposed, and the top fifth of the most exposed companies recorded labour productivity growth of 163 percent. Perhaps most surprising to those who expect automation to shrink payrolls, headcount at the most AI exposed companies grew 52 percent over the same period, against 36 percent at the least exposed, with wage growth of 24 percent against 17 percent.

I would offer three cautions about these figures before drawing conclusions. The first is that they describe correlation. Companies in AI exposed sectors may be growing faster for reasons unrelated to AI, such as being concentrated in technology, finance and professional services, which have had strong years. The second is that exposure is an index constructed by the researchers from task data, not a measure of what a given company actually does with AI, so the findings tell us about sectors with more automatable tasks more than about firms that adopted tools. The third concerns the wage premium, which depends heavily on method. PwC compares postings that require AI skills with otherwise similar postings that do not, within the same sector, and finds a premium above sixty percent. Lightcast, which compares advertised salaries for postings that list AI skills, has reported a premium of about 28 percent. Both can be true, because they measure different things, and the gap is a reminder that a single headline number for the AI wage premium conveys false precision. What is robust is the direction: employers are paying more for people who can combine domain skill with the ability to use these tools well.

The Barometer's most useful conceptual contribution is a distinction between two kinds of AI exposed jobs. In what PwC calls professionalised jobs, AI automates routine tasks and makes human expertise more valuable, so the work becomes better paid and more in demand. Radiologists and recruiters are its examples. In democratised jobs, AI makes the task easier for non experts, so the role becomes less scarce and its wage advantage erodes. IT service managers and medical secretaries are its examples. PwC reports that professionalised roles are growing about twice as fast and seeing salaries rise about 42 percent faster than democratised ones. The framework is attractive because it explains why the same technology can raise pay in one occupation and compress it in another. I think it is right in outline and I want to add a caution about its edges. Whether a job is professionalised or democratised is not fixed by its title. It depends on how the employer designs the work, how much judgement it leaves with the person and how quickly the tools improve. A role that is professionalised today can be democratised in two years if the tool takes over the part that required expertise. Workers and employers should therefore treat the classification as a direction of travel to be monitored, not a permanent label.

Now consider the people at the bottom rung of the ladder, because this is where the evidence is most worrying and also most contested. Stanford researchers using payroll records have reported employment declines of around twenty percent since 2024 for software developers aged 22 to 25, with smaller or no declines for older workers in the same occupations. Job posting data from several providers show entry level openings weakening while postings for experienced staff hold up better. PwC's Barometer includes a targeted analysis of early career roles and finds that entry level jobs in highly AI exposed occupations are increasingly demanding skills traditionally associated with senior staff, such as judgement and leadership, which I read as a sign that the tasks which once taught beginners are being absorbed by software while the expectation of what a beginner can already do rises. Surveys of new graduates in the spring show growing anxiety, and some forecasters predict falling enrolment in computer science as a result.

I should be honest about how much is known. The aggregate evidence does not show a collapse in employment for young workers in general, and the causes of the weakness in entry level hiring are disputed. Interest rates, the correction after the pandemic hiring boom, remote work and offshoring all contributed, and researchers have not yet separated them cleanly from automation. What can be said with more confidence is the mechanism by which AI could matter. Tasks done by beginners, such as drafting routine documents, writing boilerplate code, preparing first pass analysis and handling standard enquiries, are precisely the structured, checkable tasks where the technology performs best. An employer that can have a senior person direct a tool to produce the same output faces a real choice about whether to hire a beginner. If many employers make that choice at once, the pipeline of future experienced workers narrows, and the cost arrives years later. I regard this as the most important unresolved question in the study of work today, and I will say more about what to do about it below.

Turn now to the layoffs, which dominate the news and which require careful reading. The outplacement firm Challenger, Gray and Christmas tracks announced job cuts and the reasons employers give for them. In its report for May, published on 4 June, it recorded 97,006 announced cuts, the highest May total since 2020, and attributed 38,579 of them to artificial intelligence, the highest monthly figure for that reason since it began tracking it in 2023. AI was the leading stated reason for the third month running, rising from about seven percent of cuts in January to ten in February, 25 in March, 26 in April and close to forty in May. For the year to May, cuts attributed to AI totalled 87,714, about 22 percent of the total and already well above the roughly 55 thousand attributed to it in all of 2025. Technology was by far the leading sector, with 38,242 cuts in May alone, and financial technology companies reported a large number of cuts citing AI.

How should we read these numbers? Challenger's own commentary is revealing: regardless of whether individual jobs are being replaced by AI, the money for those roles is. That sentence captures what I believe is the real mechanism behind many announcements. Companies are redirecting budgets from headcount to AI investment, whether or not the software actually performs the departed workers' tasks. The announced reason is a statement by the employer, not a finding of cause, and there are incentives to prefer an explanation that sounds forward looking over one that sounds like financial pressure. Several observers have noted that AI is a flattering explanation for ordinary cost cutting. Another limit is that announced cuts are not realised unemployment, and earlier studies this year, including one by Anthropic in March, found no clear rise in unemployment among workers in the most exposed occupations. The sensible reading, in my view, is that AI is genuinely changing how companies budget for labour, especially in technology, and that the announcements overstate the direct replacement of workers by machines while accurately signalling a shift in where companies are willing to spend.

The productivity statistics add a further layer of complication. The Bureau of Labor Statistics reported in early May that nonfarm business labour productivity rose 0.8 percent at an annualised rate in the first quarter of 2026 and 2.9 percent from a year earlier, the thirteenth consecutive quarter of positive year on year growth, with the revised estimate in early June cutting the quarterly figure to about 0.3 percent. The same release reported that labour's share of output fell to 54.1 percent, the lowest since the series began in 1947. Economists at Indeed's Hiring Lab noted that the productivity rise has coincided with heavy capital spending, including on AI, and argued that AI is currently contributing to productivity growth through that spending, not yet through efficiency gains in how work is done. A research group in the United Kingdom reached a similar view, observing that technology firms appear to be making an important contribution to the productivity revival while other firms learn from their own experiments.

Two things strike me about this picture. The first is that the aggregate productivity data are still too noisy and too influenced by other factors to attribute to AI with confidence, which is the same conclusion economists reached after earlier technological waves. The second is that a falling labour share is not a statistic to dismiss. If output per hour rises while the share of income going to workers falls, then the gains are accruing to owners of capital faster than to wages. That is not a statement about AI in particular, since the share has trended down for decades, but it frames the distributional question that the technology raises. Whether workers capture a fair share of any productivity gains will depend on bargaining power, skills and institutions, and the PwC finding that wages are growing faster in the most exposed companies suggests that at least some workers are sharing in the gains, especially those with the skills to use the tools well.

That covers what the aggregate data show. The next question is what has changed inside the work itself.

Microsoft's 2026 Work Trend Index is the best single source I know on that question, because it combines survey responses from 20,000 workers who use AI in ten countries, collected between February and April, with anonymised telemetry from Microsoft 365. It reports that 49 percent of a sample of Copilot conversations support cognitive work, meaning analysis, problem solving, evaluation and creative thinking, not the production of routine text. About 58 percent of AI users say they are producing work they could not have produced a year ago, and 66 percent say AI lets them spend more time on high value work. Active agents on the Microsoft 365 platform grew about fifteen times year on year, although the company did not disclose the starting number, which makes the multiple hard to interpret. I would treat all of this as a company describing its own product, and I would note that survey respondents were selected because they use AI, so the figures describe users, not the workforce.

The more interesting finding is about organisations. The Index sorts workers into zones according to whether their own capability and their organisation's readiness line up. Roughly one in five sit in what it calls the frontier zone, where capable workers work in organisations that support them. About one in ten are blocked, meaning skilled workers inside organisations that have not caught up, and about half sit in an emerging zone in between. It describes a transformation paradox: 65 percent of workers fear falling behind if they do not adopt AI quickly, yet 45 percent say they prefer to stay with current goals over redesigning how they work, and only 13 percent feel rewarded for reinventing their work with AI. A separate Microsoft study of 1,800 workers found that employees were about 1.4 times more likely to be high frequency users of agentic tools when their managers created psychological safety around experimentation. The message I draw is that the main bottleneck to benefiting from these tools is not the technology or the individual's skill. It is the organisation's willingness to let people change how work is done, and to reward them for it.

That conclusion matches McKinsey's findings from the previous autumn, which I regard as the most consistent result in the whole literature. The organisations that attribute meaningful profit impact to AI are, by a wide margin, the ones that fundamentally redesigned their workflows, 55 percent of high performers against about 20 percent of others, and the ones that defined clear human review processes, 65 percent against 23 percent. The tools are available to everyone, but a firm that places them on top of an unchanged process gets modest results, while one that rethinks the process around them can get large ones. This is an old lesson from earlier technological waves. Electric motors did not transform factories until managers redesigned the layout of the factory floor, a change that took decades. The difference now is that the cycle is shorter and the penalty for slowness is higher.

The nature of managerial work is changing too, though I would be careful about confident claims here, because the evidence is mostly descriptive. As software takes on more of the execution of routine tasks, the human part of the job shifts toward setting direction, specifying what good looks like, reviewing results, handling exceptions and taking responsibility. Microsoft's report describes four patterns of collaboration between people and agents, in which the person authors, edits, directs or orchestrates, and argues that the choice depends on the task. I find this more useful than the usual talk of replacement, because it points to a skill that can be taught: the ability to specify work clearly, to judge output quickly and accurately, and to know when to trust a tool and when to check it. Workers who develop that skill appear to be the ones whose pay is rising, and it is a skill that depends on knowing the underlying domain well enough to recognise a wrong answer.

This connects to a point that I think is under appreciated. Review is harder than production in one important respect: it requires knowledge. A beginner who produces a draft learns from the process of producing it. A person asked to review a draft produced by a machine must already know enough to spot a subtle error, and the more fluent the machine, the more subtle the errors tend to be. That is why the entry level problem and the productivity story are linked. If beginners no longer do the production work through which they learned the craft, they must acquire the judgement needed for review in some other way, and employers have not yet worked out how. The most thoughtful firms I have read about are experimenting with deliberate apprenticeship: having beginners review and correct machine output under supervision, rotate through the work that machines do, and explain their reasoning aloud so that their judgement can be developed. Those that do nothing will find that they have saved the cost of a beginner and lost the senior engineer or analyst they would have become.

Turn now to the place of work, where the evidence is surprisingly stable. The return to office debate generates enormous noise, but the underlying numbers have moved little. Gallup's tracking of remote capable American employees shows roughly half working hybrid, about a quarter to 28 percent fully remote and about a fifth fully on site, a distribution that has been remarkably steady since 2022. Stanford's work from home research puts remote work at about a quarter of all paid workdays, again stable for several years, and a joint survey by Stanford and the Federal Reserve Bank of Atlanta found that only about one in eight executives with hybrid or remote staff planned a return to office mandate in the coming year. What has changed is the structure of hybrid work. The Flex Index, which tracks policies across thousands of firms, reports that the share of structured hybrid firms requiring three days in the office rose to 66 percent from 53 percent in mid 2024, that large firms are the ones pushing for more office time, and that the average Fortune 100 company now requires close to four days a week. Smaller firms remain far more flexible, and they employ about half the workforce.

What do these figures imply for the relationship between AI and the workplace? My reading is that the two shifts are largely independent but interact in one important way. Remote and hybrid work reduced the opportunities for informal apprenticeship, since beginners learn by overhearing, observing and being corrected in passing, and those opportunities are scarcer when people are not in the same room. AI tools then removed a second source of learning, the routine tasks. A beginner starting in 2026 may therefore get less of both kinds of exposure than a beginner in 2019. Researchers have begun to examine whether remote work or AI better explains the timing of weakness in entry level hiring, and the question is unsettled. For employers, the practical implication is that if they want to develop junior staff, they must create structured, deliberate opportunities for learning, whether in person or remotely, that no longer happen by accident.

The demand side of the skills market shows the same shift from a different angle. Lightcast's data, summarised in the Stanford Index, shows AI skills mentioned in about 2.5 percent of all United States job postings, with the cluster of skills related to agents growing by about 280 percent in a single year from a very small base, and a dashboard maintained with the Bipartisan Policy Center reported that postings requiring AI skills were up 144 percent year on year in April against about 7 percent growth for postings overall. Earlier Lightcast research found that more than half of the postings mentioning AI were outside traditional technology sectors, which supports the view that AI skill is becoming a general workplace competence and not a specialist trade. Indeed has reported a rising share of postings mentioning AI as well. The numbers are not directly comparable, because each source defines AI skill differently, but all point the same way. The signal for a worker is that the combination of domain expertise and AI fluency is being priced into offers, while AI fluency alone, without domain knowledge, is a weaker credential than the headlines suggest.

What should an individual worker do with all this? I would offer a few suggestions, drawn from the evidence and from my own experience building software and teaching others. First, become competent with the tools in the part of your own work that is structured and checkable, since that is where the gains are largest and where you can verify the results. Second, invest in the domain knowledge that lets you judge output, because review is the skill that holds its value as production gets cheaper. Third, keep a record of what you have done with the tools and what you learned, since employers increasingly ask for evidence of judgement and not just a list of tool names. Fourth, practise explaining your reasoning, since the ability to explain why an output is right or wrong is the best proof that you understand it. Fifth, build relationships and reputation, which are not automated and which count for more as the volume of generated work rises. None of this is exotic. It amounts to becoming the person who understands the work well enough to be trusted with the machine.

For employers, the evidence suggests five practical moves. First, redesign workflows deliberately instead of adding tools to existing ones, and measure the result with a baseline taken before the change. Second, reward people who reinvent their work, since the Work Trend Index finding that only a small minority feel rewarded for doing so is a warning that incentives, not capability, are the constraint. Third, make managers responsible for creating safety to experiment, since the manager effect is among the largest in the data. Fourth, protect the apprenticeship pipeline by hiring beginners deliberately and giving them review and exception handling work with supervision, accepting that this has a cost today and a return in several years. Fifth, be honest in communication about layoffs and AI. Attributing cuts to AI when the cause is financial pressure erodes trust and encourages employees to resist tools they might otherwise embrace.

Now turn to Bangladesh, where these shifts arrive through particular channels. The country's large pool of young graduates, its growing freelance and outsourcing sector and its heavy dependence on export manufacturing mean that the labour market effects of AI will not look like those in the United States. The outsourcing and freelance sector is the most directly exposed. Work such as transcription, basic translation, routine content writing, data entry, simple design and first level customer support is highly structured and checkable, which is exactly where the productivity gains are largest and where buyers can most easily substitute tools. Freelancers whose work consisted of such tasks face real pressure on prices. The opportunity lies in moving up the value chain, toward work that combines tools with judgement, such as editing and quality assurance, localisation that requires cultural knowledge, specialised technical work and consulting that involves understanding a client's business. Bengali language work deserves a particular mention: the quality of automated Bengali output is still uneven, which makes skilled human review of Bengali text, speech and translation a service with real value.

In the garment sector, the office side of the business, covering merchandising, compliance documentation, planning, procurement and customer communication, is full of structured, repeatable tasks that software can speed up. I would expect gradual change as firms adopt tools for document handling, scheduling and reporting, with the main benefits going to firms that already keep clean records. The factory floor is a different story, as machine automation of sewing is slow and depends on the economics of specific products. In education and healthcare, the more immediate effect is on the work of teachers, clerks and administrators, where tools can reduce routine paperwork. In every case, the lesson from the global evidence applies: the benefits go to organisations that redesign work and to workers who combine domain knowledge with the ability to direct and check the tools, while those who simply wait are likely to find the value of their routine tasks eroding.

I also want to say something about education and training for the next generation of workers in Bangladesh, since the entry level problem will matter here too. Universities and training programmes should assume that students will use AI for their work and should design assessment and projects that develop judgement: oral explanations, live problem solving, real projects with real users, and reflection on errors. Employers who are willing to hire and train graduates, even when tools could do part of the work, will be building the senior workforce of the 2030s. Government and industry bodies could help by supporting structured internships, apprenticeship schemes and shared training resources, particularly for small firms that cannot fund them alone. A country that treats the junior pipeline as infrastructure will be better placed than one that leaves it to chance.

Now for my own predictions, which are judgements and not forecasts from any institution, written from the middle of June without knowledge of what the coming months will show. Over the next twelve months I expect the aggregate unemployment rate for exposed occupations to remain close to that of other occupations, while entry level hiring in the most exposed fields stays weak. I expect the share of layoffs attributed to AI to remain high in announcements, with growing scepticism from economists about how much is genuine replacement and how much is budget reallocation or a convenient explanation. I expect the wage premium for AI skills to persist but to narrow as skills become common, and for premiums to shift toward people who combine AI fluency with scarce domain expertise. I expect companies to experiment with formal apprenticeship and early career programmes designed around review and exception handling, as the pipeline problem becomes harder to ignore. I expect hybrid work to remain the dominant arrangement for knowledge workers, with office requirements creeping up in large firms and holding steady in smaller ones. And I expect measured productivity growth to remain positive but noisy, with debate about how much to attribute to AI continuing for years.

It is worth adding a word about small businesses and self employed people, since most coverage concerns large companies and the evidence about them is thin. The same logic applies at a smaller scale, with one difference: a very small firm has no bureaucracy to slow its adoption, and it also has no spare capacity to absorb mistakes. The owner of a small shop, clinic, school or agency who learns to use the tools for the structured parts of the business, such as drafting customer messages, preparing quotations, organising records and summarising documents, can reclaim hours each week. The risk is the same as at larger firms, that speed is gained and judgement is lost. A message written by a tool that promises something the business cannot deliver, or a financial summary that contains a subtle error, can cost more than the time saved. The simple discipline of having the owner or a trusted person read everything that reaches a customer or touches money remains the best safeguard, and it is cheap.

It may also help to describe how a manager can measure the effect of these tools inside a team, since so much of the public debate rests on claims that nobody inside the organisation can check. The method I would suggest is modest. Choose one repeated task that the team performs, such as preparing a weekly report, answering a category of customer enquiry or reviewing a type of document. Record how long it takes and how many errors it contains today, over a few weeks, before any change. Introduce the tool for that task only, with a clear instruction on how it should be used and who reviews the result. Then measure again over the same number of weeks, paying attention not only to time saved but to errors caught late, to how often a human had to redo the work and to how the team feels about it. If the numbers improve and the team can explain why, extend the approach to a second task. If they do not, find out whether the problem is the tool, the process or the training. This is slow and unglamorous compared with announcing a company wide rollout, but it produces something most rollouts lack: evidence from your own work.

Finally, consider three short sketches of how the same technology can leave different workers in different positions. A claims processor in an insurance office whose job consists mainly of reading documents and entering standard fields finds that software now does most of this. If the office responds by cutting the team, she loses her job. If it moves her to handling the unusual claims, speaking to customers who are distressed and checking the software's decisions, her role becomes more skilled, and her pay may rise. A junior software developer who used to be assigned small bug fixes finds that a tool does them in minutes. If the firm stops hiring juniors, he never gets the job. If it makes him responsible for reviewing and testing what the tool produces, he learns faster than his predecessors did and becomes a senior engineer sooner. A freelance translator who worked on routine documents sees her rates fall as clients try automated output. If she competes on speed, she loses. If she repositions as a reviewer and cultural adviser for businesses that need Bengali text to read naturally and accurately, she can charge for judgement the tools cannot supply. In each case, the technology is the same, and the outcome depends on a decision made by the employer or the worker about where the human contribution lies.

There are limits to what this study can claim. Most of the evidence comes from the United States and a few other advanced economies, and much of it comes from surveys, corporate reports and job posting data, which are imperfect measures of what actually happens in workplaces. Several of the most cited figures come from companies with an interest in the story, including Microsoft, whose survey covers only workers who already use its tools, and PwC, whose exposure index is a modelling construct. Layoff data reflect what employers say, not necessarily why they act. The period covered is short, and conclusions drawn from one year of data may look naive in hindsight. I have tried to say where each source is strong and where it is weak, and I encourage readers to go back to the originals and check the definitions and the dates before relying on any figure.

To bring the threads together: the shape of work changed between 2025 and 2026 less in the number of jobs than in what jobs contain, who is hired into them and how they are paid. The technology is widely used and shallowly integrated, with the biggest measured gains in structured, checkable tasks. A two track labour market is emerging in which roles that combine expertise with the tools gain value and roles that the tools make easy for anyone lose it, though the line between them is not fixed. Entry level work is narrowing, and the cost of that will arrive in the form of a thinner pipeline of experienced people unless employers deliberately rebuild apprenticeship. Layoffs attributed to AI reflect a real shift in budgets more than a simple replacement of workers by machines. Productivity is rising and labour's share of income is at a record low, which frames the question of who benefits. Hybrid work has settled into a stable pattern that makes informal learning harder. The organisations that gain most are those that redesign work, reward people who reinvent it and protect the path by which beginners become experts. For a worker, the best response is to become the person who understands the work well enough to direct the tools and judge their output. For a country like Bangladesh, the opportunity is to move up from routine tasks to judgement, language and local knowledge, and to treat the training of young people as the foundation of everything else.

Sources: Stanford HAI, AI Index Report 2026, economy chapter (https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_4_economy.pdf); PwC, 2026 Global AI Jobs Barometer (https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html) and global findings (https://www.pwc.com/gx/en/issues/artificial-intelligence/job-barometer/2026/2026-global-ai-jobs-barometer-global-findings.pdf); Microsoft, 2026 Work Trend Index (https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization); Challenger, Gray and Christmas, Job Cut Announcement Report for May 2026 (https://www.challengergray.com/wp-content/uploads/2026/06/Challenger-Report-May-2026.pdf) and for April 2026 (https://www.challengergray.com/wp-content/uploads/2026/05/Challenger-Report-Apr2026001249.pdf); CBS News on AI and April layoffs (https://www.cbsnews.com/news/ai-layoffs-job-cuts-challenger-report-april-2026/); Bureau of Labor Statistics, Productivity and Costs, first quarter 2026, preliminary (https://www.dol.gov/newsroom/economicdata/prod2_05072026.pdf) and revised (https://blsmon1.bls.gov/news.release/prod2.nr0.htm); Indeed Hiring Lab on the first quarter productivity release (https://www.hiringlab.org/2026/05/07/q1-2026-productivity-and-costs-release/); Productivity Institute on US productivity growth (https://lab.productivity.ac.uk/insights/q1-2026-us-productivity-growth-revival/); McKinsey, The state of AI in 2025 (https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai); Lightcast on the Stanford AI Index 2026 (https://lightcast.io/resources/blog/stanford-ai-2026) and Bipartisan Policy Center skills dashboard, April 2026 (https://bipartisanpolicy.org/article/navigating-skills-trends-data-dashboard-analysis-april-2026/); Flex Index and workplace policy data summarised by OfficeRnD (https://www.officernd.com/blog/hybrid-work-statistics/) and Rewordin (https://www.rewordin.com/blog/hybrid-remote-work-statistics-2026); FlexJobs, remote work statistics report, April 2026 (https://www.flexjobs.com/blog/post/flexjobs-remote-work-statistics-report).