NextFin News - After three years of trillion-dollar bets on artificial intelligence in the workplace, the first controlled real-world trials have produced a verdict, and it is more awkward than either the utopians or the doomsayers predicted: AI delivers large productivity gains at the task level, but most firms are failing to convert those gains into measurable business value, and the workers who should benefit most are the ones most likely to lose their foothold on the career ladder.
The question has moved from speculation to measurement. Since ChatGPT's launch in late 2022, companies have raced to put AI tools in employees' hands. By late 2024, nearly 40% of U.S. adults ages 18-64 reported using AI tools — an adoption pace that exceeded the comparable early stages of the personal computer and the internet. The spending has been staggering: hyperscalers and large technology companies have committed hundreds of billions of dollars to AI infrastructure, and enterprise software budgets have been reoriented around "copilots" and "agents."
Now the bill is coming due, and the experiments are producing hard numbers. The evidence comes from three distinct kinds of trials, each with a different answer.
First, controlled field experiments inside firms — the gold standard — show large effects. A study of 5,179 customer support agents, published in the Quarterly Journal of Economics, found that access to an AI assistant raised productivity, measured by issues resolved per hour, by 15% on average, with a 34% improvement for novice and low-skilled workers and minimal impact on experienced staff. Three randomized field experiments involving 4,867 software developers at Microsoft, Accenture, and an anonymous Fortune 100 company found developers with access to an AI coding assistant completed 26.08% more tasks per week. A field experiment with Boston Consulting Group consultants found substantial performance gains on tasks within the system's capability range — and performance declines on tasks just beyond that frontier, the phenomenon researchers called the "jagged technological frontier."
Second, real-world deployment tells a humbler story. When the New York Times deployed AI agents to act as office workers with full laptop access in July 2026, the agents completed some tasks but failed in revealing ways: they wrote code instead of using human-designed interfaces, and they made a consequential error when asked to recommend staff cuts, adding employees on leave to the cut list without considering when those employees would return. Scale AI, which tested agents on real freelance projects, found the best-scoring model produced client-ready work only about 16% of the time. METR, which ran a randomized trial with experienced open-source developers, found AI tools caused tasks to take 19% longer in its early-2025 study, with a confidence interval between 2% and 39% — a result it attributed to the overhead of managing unreliable AI output.
Third, the macroeconomic and survey data show almost no disruption. A Federal Reserve Bank of Atlanta survey of nearly 750 corporate executives, published in March 2026, found that nearly 60% of firms invested in AI in 2025, yet there was little evidence of near-term aggregate employment decline. The European Central Bank's March 2026 survey of about 5,000 euro-area firms found no overall employment gap between AI-using and non-using firms; intensive AI users were about 4% more likely to take on additional staff. The Yale Budget Lab found the U.S. occupational mix is shifting no faster than it did when the personal computer or the internet arrived, with no relationship between an occupation's AI exposure and its employment or the duration of unemployment.
The combination is the story. Task-level capability has changed permanently. Employment has not. Something sits between the two — and that something is the organization.
The Productivity Paradox Is an Organizational Problem, Not a Technological One
The central puzzle of the AI era is why large task-level gains have not translated into measured productivity at the firm or economy level. The Atlanta Fed study put a number on the gap: labor productivity attributable to AI investment was 1.8% in 2025 and is expected to strengthen in 2026, but the same study documented a "productivity paradox" in which perceived productivity gains were larger than measured gains. The authors attributed the gap to a delay in revenue realizations — the gains are real but have not yet shown up in the top line.
Microsoft's 2026 Work Trend Index, which analyzed trillions of anonymized Microsoft 365 productivity signals and surveyed 20,000 workers using AI across 10 countries, pointed to a different explanation: the constraint is organizational. Only 19% of workers are in what Microsoft calls the "Frontier" zone — high AI capability combined with organizational readiness. About half are in the "emergent" zone, experimenting at the edges. The report's most consequential finding: organizational factors — culture, manager support, talent practices — account for 67% of reported AI impact, compared with 32% for individual mindset and behavior.
That ratio should unsettle any executive who believes buying licenses is the same as adopting technology. Frontier Professionals — the 19% pulling ahead — are significantly more likely to say their manager openly uses AI (85% versus 64%), sets quality standards for AI work (83% versus 57%), creates space for experimentation (84% versus 61%), and encourages more ambitious work redesign (87% versus 61%). They are also twice as likely to say they are rewarded for the reinvention of work with AI regardless of outcome (26% versus 11%).
ADP Research reached a similar conclusion from payroll and survey data representing more than 25 million U.S. workers. In its 2025 Global Workforce Survey, half of workers globally said they use AI at least multiple times a week, and one in five uses it nearly every day. Daily users were more engaged (30% fully engaged versus 14% for non-users) and less stressed (11% overloaded versus 23% of non-adopters). But they did not feel more productive: daily users were four times as likely as non-users to say they were less productive than they could be.
The most plausible reading is not that AI fails to work, but that firms have measured the wrong thing. Companies are measuring productivity when they should be measuring transformation: decision quality, cycle times in the parts of the business that matter, learning velocity, agent reliability, governance maturity. A tool that lets a junior analyst produce a first draft in ten minutes looks like a productivity gain only if the analyst's job was to produce first drafts. If the job is to make a recommendation, the relevant metric is whether the recommendation is better — and that requires redesigning the workflow, not installing software.
The Jagged Frontier: Why AI Helps the Inexperienced and Trips Up the Expert
The field experiments reveal a consistent pattern that upended the early consensus. The initial wisdom, built on the customer-support study, was that AI is a "leveler": it helps the least experienced and lowest performers the most, compressing the skill distribution. That part held. The 34% gain for novice support agents versus minimal gains for veterans is the cleanest example. In taxi fleets, an AI demand-prediction system narrowed the earnings gap between the best and worst drivers by 14%, with gains accruing almost entirely to low-skilled drivers.
But the BCG consultant experiment added the crucial caveat. Consultants using GPT-4 performed substantially better on tasks that fell within the model's capability range — and worse on tasks just beyond it, because they over-relied on AI-generated suggestions for problems the system could not reliably solve. The "jagged technological frontier" is not a smooth curve; it is a cliff edge, and the most dangerous place to stand is close to it, confident that you are on the safe side.
This explains the METR result that looked like a paradox: experienced open-source developers given AI tools were 19% slower at completing tasks in the early-2025 study. These were not novices. They were experts working near the frontier, where the tasks AI cannot do reliably are exactly the tasks experts are hired for. The time saved on the easy parts was consumed by verifying, correcting, and sometimes discarding AI output on the hard parts. METR revised its experiment design in February 2026 and noted that developers are likely more sped up by early-2026 AI than by early-2025 AI, but flagged that its new data remained too weak to estimate the size of the increase reliably.
The practical implication is counter-intuitive: the workers most exposed to displacement are not the experts, but the competent intermediates — the people whose jobs consist of tasks that sit safely inside the frontier. The experts are needed to catch the errors; the novices are being trained faster. The middle is hollowed out.
The Entry-Level Problem: A Leadership Pipeline Nobody Is Measuring
The most under-discussed consequence of skill compression is not unemployment today but a leadership shortage tomorrow. Junior jobs have always served a dual purpose: they produce output, and they train the seniors who will run the organization. If AI absorbs the tasks that juniors used to do to learn — the first drafts, the code reviews, the data pulls, the client memos — then the training ground disappears even if headcount does not.
The data already show the pressure point. The World Economic Forum, in collaboration with PwC, reported in June 2026 that more than one in three young workers (37%) are employed in occupations with medium to high exposure to AI-driven task change, including three in four young workers in Eastern Asia (75%) and two in three in Northern America (69%) and Europe (63%). Stanford Digital Economy Lab research, using ADP payroll microdata covering millions of U.S. workers, has linked AI adoption with slower job creation among young workers in some sectors. ADP's own survey found that young workers, including frequent AI users, are less optimistic about their job security than older workers. The Atlanta Fed study found compositional reallocation, with routine clerical roles declining and relative demand for skilled technical roles increasing; firms expect the share of routine clerical workers to fall by 0.76% in 2026 and 2.19% in 2028.
This is not mass displacement. It is something slower and harder to see: the rung at the bottom of the ladder is being removed, and the organization will discover the gap only when it needs to promote someone who never learned to climb.
The Second-Order Question the Market Has Not Priced
The market has largely priced a binary narrative: AI either replaces jobs or it does not. The trials suggest a third outcome that is more consequential for how capital should be allocated. AI does not replace jobs; it replaces tasks, and it changes which firms win. The companies that redesign work around human-AI collaboration pull ahead; the companies that hand out licenses and wait for the productivity number to appear fall behind.
The second-order effect runs through the cost of coordination. When a novice can produce output that previously required a senior, the bottleneck shifts from production to judgment. The scarce resource is no longer the ability to generate a draft; it is the ability to decide which draft is right, to navigate a user interface that was designed for human eyes, to know how long an employee's leave will last. The New York Times agent experiment found AI struggled with exactly these things: understanding the nuances of human language, navigating the Chrome browser, knowing when not to cut a role.
A.I. may be less capable of replacing tacit knowledge, the idiosyncratic tips and tricks that accumulate with experience but which are never digitized.
That is why the real-world trial results are not a repudiation of AI. They are a map of where human judgment remains the scarce input — and therefore where the economic value will accrue.
The Strongest Case Against This Reading
The counter-thesis is serious and deserves its due weight. The productivity gains documented in controlled experiments are task-level measurements in bounded settings; the real-world success rates are far lower (Scale AI's 16% client-ready figure; METR's 19% slowdown). The aggregate data show no disruption yet — but that is exactly what a Solow paradox would predict. When electricity was introduced, factories initially saw no productivity gain because they had to be rewired; the gain came a generation later, when the architecture changed. If AI follows the same path, today's "no displacement" data are not evidence that displacement will not happen — they are evidence that it has not happened yet. The Atlanta Fed's own finding that larger companies anticipate AI-driven workforce reductions, while smaller firms expect modest gains, is consistent with a delayed but real shock.
This argument is strongest on timing, weakest on mechanism. The Solow analogy assumes the technology works and the organization lags. But the trials show something more fundamental: the technology itself is unreliable at the frontier, and the tasks it cannot do are the ones that define many jobs. A factory rewired for electricity could run all day. An organization rewired for AI still needs a human to catch the leave-date error. That is not a lag; it is a permanent feature of the technology as it exists today.
The falsifying signal is concrete: if measured labor productivity attributable to AI stays below 1% annually through 2027 despite adoption above 50% of firms, the structural-shift thesis is wrong and this is a cyclical hype cycle whose costs will be written off. A second signal: if routine clerical employment does not decline by the expected 2.19% by 2028 while AI adoption continues to rise, the compositional-shift thesis fails and the "AI changes work" narrative has been oversold.
Conclusion: Who Wins, Who Loses, and What to Watch
The mechanism, cashed out, points to a specific asymmetry. The beneficiaries are not the AI vendors alone and not the firms that automate headcount. They are the organizations that treat AI as a redesign problem: managers who model use, set quality standards, and create space for experimentation; firms that measure decision quality and learning velocity rather than draft volume; workers who keep some tasks deliberately human to preserve the skills that let them judge AI output. Microsoft's data show that when managers actively modeled AI use, employees reported a 17-point lift in AI value, a 22-point lift in critical thinking about their AI use, and a 30-point lift in trust in agentic AI.
The exposed are the competent intermediates whose tasks sit safely inside the frontier, and the young workers whose entry-level rungs are disappearing. They are not being laid off in mass today. They are being denied the repetition that used to build expertise.
Split by horizon, the picture diverges. In the short term — the next six to eighteen months — the dominant forces are sentiment and liquidity: experimentation costs, ROI pressure, and the budget strain that has already led some companies to rein in AI usage. The Atlanta Fed's productivity paradox and ADP's finding that heavy users feel less productive will feed skepticism. In the medium term — two to five years — the fundamentals dominate: task reallocation, skill compression, compositional shifts away from routine clerical work, and a widening gap between firms that redesign work and firms that do not. In the long term — five years and beyond — the shift is structural: the definition of "work" changes, the leadership pipeline problem surfaces, and new occupational categories emerge around judging, coordinating, and governing AI output.
Three scenarios frame the path. The base case is muddling through: adoption continues, measured productivity creeps toward the Atlanta Fed's expectation of strengthening gains in 2026, and employment shifts compositionally without aggregate disruption. The upside case is the J-curve payoff: organizational redesign catches up, the productivity paradox resolves, and the gains documented in controlled experiments finally appear in firm-level data. The downside case is the Solow trap without the payoff: the technology's frontier reliability does not improve fast enough, real-world success rates stay near the 16% range, and capital begins to write down AI investments as the costs strain budgets faster than the returns arrive.
The trial is not over. But the first results are clear enough to act on: the question is no longer whether AI can do the task. It is whether the organization can be rebuilt around the answer.
The market priced a jobs apocalypse and a productivity miracle; the trials say both are wrong. What is actually happening is quieter and more expensive — a rewrite of the organization that most companies have not yet started.
Explore more exclusive insights at nextfin.ai.
