Does AI Actually Make Workers More Productive

The current scientific consensus demonstrates that generative artificial intelligence acts as a profound, albeit uneven, accelerator of knowledge worker productivity, yielding average efficiency gains between 15% and 40% across various cognitive tasks. However, these gains are highly heterogeneous, consistently acting as a skill-leveler that disproportionately benefits novice and low-performing workers while offering marginal, or occasionally negative, returns to highly skilled experts operating at the boundary of current technological capabilities. Ultimately, while microeconomic field experiments demonstrate staggering task-level efficiencies, macroeconomic diffusion models urge caution, noting that realizing aggregate global economic growth will require significant structural adaptation, infrastructure investment, and human-in-the-loop oversight.

For the modern professional, the advent of generative artificial intelligence represents the most significant shift in the valuation of human labor since the industrial revolution. Whether a professional is drafting a routine executive summary, analyzing complex financial derivatives, representing a client in appellate proceedings, or architecting a million-line software repository, understanding precisely where artificial intelligence excels and where it catastrophically fails is no longer a purely academic exercise. It has become the determining factor in whether professional human capital will be exponentially amplified or rendered structurally obsolete over the coming decade. As this technology permeates everyday enterprise software suites, the organizations and individuals that thrive will not necessarily be those possessing the most advanced technical engineering acumen. Rather, the economic winners will be those who master the nuanced art of delegating appropriate cognitive load to machines while fiercely protecting and cultivating human judgment in the opaque areas where algorithms falter.

What is the current scientific consensus on AI's impact on worker productivity?

Historically, successive waves of technological automation have predominantly impacted routine, codifiable tasks governed by explicit sequences of steps, such as manufacturing assembly, data entry, or bookkeeping ¹. Generative artificial intelligence, however, fundamentally alters this historical paradigm by demonstrating immense, scalable proficiency in cognitive, creative, and non-routine knowledge work - domains that have long relied exclusively on specialized and highly educated human capital ¹. Over the past several years, a robust body of empirical evidence, spanning randomized controlled trials, difference-in-differences econometric models, and extensive corporate field experiments, has established that large language models significantly boost human output across diverse intellectual disciplines.

The baseline productivity gains observed in controlled professional environments are striking and consistent. In an extensive 2023 study by researchers at the Massachusetts Institute of Technology, college-educated professionals assigned midlevel writing tasks, such as drafting delicate corporate communications and conducting cost-benefit analyses, experienced a 40% reduction in task completion time when granted access to ChatGPT ¹²³. Crucially, this rapid acceleration in output generation did not come at the expense of quality. Independent evaluators, grading the outputs in a blinded process, determined that the quality of the AI-assisted work actually rose by 18% compared to the control group ²³. The durability of this intervention was also notable; workers exposed to the chatbot during the experiment were twice as likely to report integrating it into their actual employment workflows two weeks later ²³.

These cognitive enhancements extend deep into highly specialized, technical domains where precision is paramount. In the realm of software engineering, a controlled experiment analyzing the deployment of GitHub Copilot, an AI pair programmer, found that software developers tasked with implementing an HTTP server in JavaScript completed the assignment 55.8% faster than a control group relying purely on traditional internet search and documentation ⁴⁵⁶. In the highly regulated legal sector, a 2025 randomized controlled trial involving upper-level law students demonstrated that next-generation legal AI tools dramatically improved efficiency ⁷. When utilizing Vincent AI, a legal assistant powered by retrieval-augmented generation (RAG), alongside OpenAI's reasoning models, participants achieved statistically significant productivity gains of up to 140% across litigation-related tasks, such as drafting complex memos and client letters ⁷. Unlike earlier iterations of generative models that suffered from severe hallucination rates, the RAG-powered tools grounded in verified legal source material yielded factual error rates no higher than those of unassisted human students, while substantially boosting analytical depth, clarity, and organizational flow ⁷.

The integration of generative AI into high-volume business operations also transforms customer-facing roles, altering both the speed and the qualitative experience of the labor. In a landmark National Bureau of Economic Research working paper, subsequently published in The Quarterly Journal of Economics in 2025, researchers examined the staggered rollout of a generative AI conversational assistant among 5,172 customer support agents at a Fortune 500 software firm ⁴⁹⁵¹¹. By utilizing a difference-in-differences methodology to track the differential timing of the software deployment, the researchers isolated the causal impact of the AI on human labor. The introduction of the AI tool increased aggregate productivity, measured by issues successfully resolved per hour, by an average of 14% to 15% ⁴⁹¹². Beyond raw operational speed, the technology improved the subjective psychological experience of the work itself. The real-time conversational guidance led to demonstrably higher employee retention, more polite customer interactions as measured by sentiment analysis, and a 25% reduction in hostile escalations to managerial staff ⁴⁹.

However, beneath these impressive aggregate figures lies a highly complex web of variable impacts. The scientific consensus definitively rejects the notion of artificial intelligence as a simple, linear multiplier of human effort. Instead, the empirical data reveals that the technology alters the very nature of human capital accumulation, completely reshaping the distribution of performance and capability within organizations.

Does AI make everyone equally faster, or does it replace human skill?

A pervasive myth surrounding the enterprise deployment of artificial intelligence is that it acts as a uniform multiplier, making every worker proportionally faster and more effective across the board. If an organization adopts a large language model, the prevailing assumption among many executives is that the top performers will become even more untouchable, while the bottom performers will experience a similar, parallel boost in output, maintaining the existing hierarchy of talent. The empirical evidence vehemently contradicts this narrative. Instead, generative AI consistently functions as a profound "skill-leveler," fundamentally compressing the distribution of performance by elevating the bottom tier while offering diminishing returns to the top tier.

This skill-leveling mechanism operates primarily by effectively capturing the tacit, unwritten knowledge of top performers and instantly distributing it to the rest of the workforce. In the aforementioned study of 5,172 customer support agents, the 14% average productivity gain masked extreme underlying heterogeneity ⁴¹⁴. Novice and low-skilled workers experienced a staggering 34% improvement in their issue resolution rates ⁴¹⁴. In stark contrast, the most experienced and highly skilled workers saw minimal gains in speed and, notably, experienced a slight decline in the qualitative accuracy of their output when relying on the tool ⁹⁵¹⁵. The AI system essentially served as an automated, real-time training mechanism. Because the model was trained on the historical chat logs of the company's best agents, it disseminated those optimized practices to veteran and novice employees alike, helping newer workers move down the steep experience curve at an artificially accelerated pace ⁴.

This phenomenon of performance compression is corroborated across multiple intellectual disciplines. In the MIT professional writing study, performance inequality between workers decreased significantly; participants who initially scored poorly on baseline tasks saw the largest leaps in quality and speed upon utilizing the chatbot ¹³¹⁶. Similarly, in the evaluation of GitHub Copilot, less experienced developers, older programmers adapting to unfamiliar languages, and those struggling with syntax recall derived the most substantial benefits ⁴⁶. This suggests that AI pair programming holds massive promise for transitioning individuals into software development careers by drastically lowering the barrier to entry for coding ⁴⁶¹⁷. In the 2025 legal sector study, the average law student improved their work quality and speed drastically, while the most excellent students improved their speed at a similar rate but saw only marginal, if any, bumps in the quality of their legal reasoning ¹⁸.

Perhaps the clearest visualization of this dynamic comes from a 2023 field experiment analyzing elite management consultants. The data indicates that consultants who initially performed below the average baseline threshold experienced a massive 43% increase in performance when given access to AI ¹⁹. By comparison, those who were already performing above average saw a much smaller 17% increase in their performance ¹⁹. This dynamic effectively levels the playing field, drastically closing the historical gap in output quality between novice practitioners and seasoned experts.

The implications of this skill compression are profound for macroeconomic labor dynamics and organizational talent management. If entry-level employees can instantly produce output matching that of seasoned veterans through algorithmic augmentation, the traditional apprenticeship model of knowledge work is severely disrupted. Researchers have identified a consequential irony in this dynamic: generative AI is increasingly tasked with taking over the routine, foundational work through which entry-level professionals have traditionally built the deep, intuitive expertise required to advance into senior roles ¹⁵. Without a solid foundation of experiential learning gained through tedious, repetitive tasks, future workers may lack the critical judgment required to evaluate complex AI outputs strategically ¹⁵. Over the long term, this could trigger a localized "Matthew Effect" within highly specialized fields, where only those who accumulated deep domain expertise prior to the AI revolution possess the foundational knowledge necessary to govern the models, thereby increasing the premium on verified human expertise even as the floor for average output is permanently raised ¹⁵.

What is the "Jagged Technological Frontier," and why does it matter?

To understand why a state-of-the-art language model can sometimes produce brilliant, synthesis-level insights and other times fail at rudimentary logic, one must understand the concept of the "Jagged Technological Frontier." Introduced in a landmark 2023 field experiment conducted collaboratively by researchers from Harvard Business School, the Massachusetts Institute of Technology, the Wharton School, and Boston Consulting Group (BCG), this framework explains the highly uneven and unintuitive landscape of modern artificial intelligence competence ²⁰⁶²².

In the BCG study, 758 elite management consultants, representing approximately 7% of the firm's individual contributor workforce, were monitored in a pre-registered experiment to determine how AI transforms high-end, complex knowledge work ¹⁹⁶²³. The researchers designed a suite of 18 realistic consulting tasks representing product innovation, go-to-market strategy, creative ideation, and analytical persuasion ¹⁹⁷. Through rigorous testing, they discovered that current AI capabilities do not represent a uniform, monotonic boundary of intelligence. Instead, the boundary separating tasks that an AI can effortlessly handle from those that exceed its abilities is highly irregular, or "jagged" ⁶²⁵. Tasks that appear to human knowledge workers to be of identical cognitive difficulty may fall on completely opposite sides of this frontier ⁶⁸.

For tasks falling inside the jagged frontier, where the AI is highly capable, the empirical results were historic. Consultants randomly assigned to use GPT-4 completed 12.2% more tasks on average, finished them 25.1% faster, and produced results graded at a staggering 40% higher quality than the control group utilizing standard human workflows ¹⁹²³²⁵. The AI acted as a massive cognitive force multiplier, particularly for creative brainstorming, drafting marketing materials, and synthesizing qualitative market data into structured formats ²³²⁵.

However, the researchers purposefully designed a specific task that fell outside the frontier. This was a complex brand strategy case requiring consultants to reconcile subtle, qualitative interview nuances with quantitative spreadsheet data - a task intentionally designed to mimic the blind spots and reasoning cliffs of modern language models ⁶²²⁷. On this specific task, consultants using the AI were 19 percentage points less likely to produce correct solutions compared to the control group of humans who did not use AI ¹⁴¹⁹²⁷.

This sharp degradation in performance outside the frontier introduces a critical psychological and operational risk into the modern workplace: the risk of unengaged interaction and blind adoption. Because large language models are remarkably fluent, confident, and capable of executing real work with virtually no technical skill required of the user, their failure points remain highly opaque ⁸. Knowledge workers, lulled into a false sense of security by the AI's sheer brilliance on inside-the-frontier tasks, often fail to see the technology's limitations. They become overly reliant, accepting incorrect, hallucinatory outputs without sufficient interrogation ²⁰⁶⁸. When recruiters, analysts, or consultants trust high-quality AI too deeply, they can become intellectually lazy, allowing the machine's plausible hallucinations to override their own expert judgment, leading to catastrophic decision-making errors in high-stakes environments ²⁷.

To successfully navigate this jagged, unpredictable landscape, the researchers observed the organic emergence of two distinct, highly effective human-AI integration strategies among the top-performing consultants: 1. Centaurs: Named after the mythical half-human, half-horse creature, these workers practice a strategy of rigid, strategic delegation. They maintain a clear, cognitive boundary between human work and machine work. Centaurs hand off specific, well-defined tasks, such as generating initial feature lists or formatting data, entirely to the AI, while retaining full, unassisted human control over tasks requiring deep strategic refinement, ethical reasoning, or emotional intelligence ¹⁹²⁰²⁷. 2. Cyborgs: In contrast to the rigid division of labor employed by Centaurs, Cyborgs intertwine their efforts intimately with the AI, moving back and forth in a continuous, deeply integrated workflow. A Cyborg might prompt the AI for a single sentence, rewrite the next paragraph themselves, ask the AI to evaluate their underlying logic against a specific framework, and continually co-create at the micro-task level, blurring the lines of authorship ¹⁹²⁰²⁵.

Both Centaur and Cyborg methodologies succeed because they actively manage the jagged frontier. These workers continuously deploy cognitive effort and expert judgment to validate, steer, and interrogate AI outputs, rather than slipping into the dangerous trap of passive reliance ⁸²⁷.

To contextualize the highly variable nature of these impacts, the table below synthesizes the core productivity, quality, and skill-leveling metrics observed across disparate fields of professional knowledge work based on recent empirical literature.

Are there limits to AI productivity gains, and can AI actually slow us down?

While the overwhelming majority of early academic studies and corporate press releases tout massive, frictionless productivity gains, highly rigorous recent field research indicates that generative AI is not a universal panacea. Under specific conditions involving dense institutional complexity, artificial intelligence can actively impede productivity and generate negative returns on human time. A compelling set of studies conducted between 2024 and 2025 by METR, a nonprofit research organization dedicated to evaluating AI capabilities, rigorously challenged the prevailing optimistic narrative regarding software developer efficiency ²⁸.

Unlike earlier, highly publicized experiments that tested software engineers on small, isolated, greenfield tasks, such as generating a single script or an HTTP server from scratch, the METR study placed highly experienced open-source developers into massive, real-world, million-line legacy codebases ²⁸²⁹³⁰. The participating developers were compensated at a premium rate of $150 per hour, and their screens were continuously recorded to ensure rigorous adherence to the methodology and to accurately track idle time ³⁰⁹. They were tasked with tackling complex bug reports, deep architectural refactoring, and nuanced feature requests using state-of-the-art agentic coding tools ³⁰⁹¹⁰.

The empirical results were a stark, data-driven counterpoint to industry hype. Developers who were randomly assigned to use AI tools took, on average, 19% longer to complete their assigned issues compared to the control group of developers working entirely without AI assistance ²⁸²⁹⁹.

Researchers identified several mechanical reasons for this significant negative productivity shock. Primarily, AI capabilities degrade precipitously in environments that demand deep, highly specific, and often undocumented contextual knowledge ⁹. When a human developer is already an expert in a specific repository and requires no external documentation to comprehend the architecture, forcing an AI assistant into the workflow introduces massive friction ²⁸. The AI introduces a new category of cognitive overhead. Instead of simply typing the solution, the human must spend time writing elaborate, context-rich prompts, waiting for generation latency, reading the often-verbose output, verifying its logic against a massive web of undocumented dependencies, and remediating inevitable hallucinations ¹⁴. In complex, interdependent systems, identifying and fixing an AI's subtle architectural mistake often takes significantly longer than a human writing the correct code from scratch ¹⁴.

Equally fascinating to the time loss was the study's revelation of a profound psychological phenomenon: the illusion of speed. Prior to the commencement of the experiment, the developers forecasted that integrating AI would accelerate their work by 24% ²⁸²⁹. After utilizing the tools, and empirically losing 19% of their actual time to completion, the developers were surveyed again. Astoundingly, they still fervently believed that the AI had sped them up by 20% ²⁸⁹.

This extreme cognitive dissonance likely stems from the fact that generative AI outsources the physical act of typing and the immediate cognitive strain of syntax recall, replacing it with the more passive acts of reading, scrolling, and verifying. Because the worker feels less immediate mental strain in the moment of creation, the work subjectively feels faster, even if the aggregate time-to-completion stretches significantly due to debugging, reprompting, and verification overhead ¹⁴²⁹. These findings introduce a crucial, sobering caveat to the jagged frontier framework: generative AI is immensely powerful for greenfield creation, ideation, and isolated tasks, but its efficacy drops into the negative when forced to navigate the dense, pre-existing structural complexities of legacy systems and deeply specialized, uncodified institutional knowledge.

How will AI affect macroeconomic growth and global inequality?

Extrapolating microeconomic, task-level productivity gains into macroeconomic growth requires navigating a complex maze of capital constraints, adoption lags, regulatory hurdles, and deeply entrenched global economic structures. While individual consultants or administrative workers may become 25% or 40% faster at specific discrete tasks, this localized efficiency does not instantly translate to a parallel increase in global Gross Domestic Product (GDP).

Leading financial institutions and macroeconomic researchers hold widely divergent views on the ultimate size of AI's aggregate economic footprint. Early, highly optimistic models produced by Goldman Sachs in 2023 projected that the widespread adoption of generative AI could drive a staggering 7% increase in global GDP over a decade, translating to an annual labor productivity growth boost of around 1.5 percentage points, mirroring the historical impact of the electric motor or the personal computer ¹¹³⁴.

However, rigorous academic macroeconomic modeling, most notably by Massachusetts Institute of Technology economist Daron Acemoglu, offers a much more tempered and conservative outlook. Acemoglu estimates that the aggregate total factor productivity (TFP) gains over the next ten years are unlikely to exceed 0.66% to 1.1% cumulatively for advanced economies ¹¹¹²¹³. This caution stems from the mathematical reality that while AI can automate a wage-bill-weighted 20% of tasks in the US economy, the actual corporate cost savings, deployment timelines, and the immense capital required to train and run these models severely limit the immediate macroeconomic surge ¹³. Furthermore, a massive portion of the real economy, including physical trades, construction, hospitality, and hands-on healthcare provision, remains largely immune to generative text and image automation, diluting the aggregate productivity impact ¹³³⁷.

Institutions like the European Central Bank (ECB) and the Organisation for Economic Co-operation and Development (OECD) emphasize that any aggregate gains are entirely dependent on the pace of firm-level adoption, which is currently lagging behind technological capability ¹⁴¹⁵. In 2025, while an impressive 57.3% of Information and Communication Technology firms across the OECD utilized AI, overall firm adoption across all sectors stood at only 20.2% ¹⁶. The OECD estimates that AI could add between 0.4 and 1.3 percentage points to annual aggregate labor productivity growth over the next decade in countries with high AI exposure and widespread adoption, such as the United States and the United Kingdom ¹⁵. In contrast, countries with less favorable digital infrastructures, stricter regulatory environments, or different sectoral compositions, such as Italy or Japan, may see their productivity gains capped between a mere 0.2 and 0.8 percentage points ³⁴.

When viewed through a global macroeconomic lens, artificial intelligence threatens to drastically widen the economic divide between the Global North and the Global South. Comprehensive reports from the International Monetary Fund (IMF), the United Nations Conference on Trade and Development (UNCTAD), and the World Bank paint a highly precarious picture for developing nations in the age of AI.

The IMF estimates that nearly 40% of global employment is currently exposed to AI disruption ³⁷⁴¹. However, this exposure is heavily skewed by the structural realities of economic development. In Advanced Economies (AEs), approximately 60% of jobs are exposed due to the high density of cognitive, administrative, and knowledge-intensive roles ¹⁷⁴³. Emerging Market Economies (EMs) face a 40% exposure rate, and Low-Income Countries (LICs) face only a 26% exposure rate, as their labor forces remain highly concentrated in physical and manual labor ¹⁷.

While lower exposure means less immediate job displacement and labor market shock in the Global South, it concurrently implies a terrifying lack of access to the productivity dividends required for long-term economic convergence ³⁷¹⁷. Automation inherently favors capital over labor. For decades, developing economies have relied on abundant, low-cost labor as their primary competitive advantage to attract global supply chains. Generative AI threatens to completely erode this structural advantage by allowing capital-rich nations to onshore administrative, coding, and knowledge processes via highly efficient software ⁴⁴⁴⁵.

The World Bank identifies four critical pillars, termed the "4Cs," that are absolutely required for a nation to harness AI for economic growth: Connectivity (reliable broadband and electricity), Compute (access to data centers and advanced AI chips), Context (locally relevant, native-language training data), and Competency (digital workforce skills) ¹⁸⁴⁷¹⁹. Currently, there is a severe, escalating global monopoly on these critical resources. Just 100 technology companies, primarily based in the United States and China, account for 40% of all global corporate R&D spending on artificial intelligence ⁴⁴⁴⁹. The market capitalization of a single AI hardware provider, Nvidia, currently rivals the entire GDP of the African continent ⁴⁴⁴⁵. Furthermore, over 73% of data centers are located in high-income countries, and 45% of global internet traffic is conducted in English. This means foundational models often completely lack the cultural, linguistic, and socioeconomic "context" to serve local populations in the developing world effectively ⁴⁷. Politically, 118 nations remain completely absent from global AI governance discussions, leaving them subject to technological trade rules and ethical frameworks they have no hand in writing ⁴⁴⁴⁵.

Despite these massive structural deficits, international development organizations point to a highly promising, grassroots trend: the rise of "Small AI" ¹⁸¹⁹. Rather than relying on massive, trillion-parameter foundational models that require hyperscale data centers and massive energy grids to run, innovators in the Global South are increasingly deploying localized, affordable, and highly efficient AI applications designed to run edge inference on ordinary mobile devices ¹⁸⁵⁰. These localized, highly tuned models are currently helping rural farmers predict localized rainfall and optimize crop yields, assisting under-resourced teachers in tracking student progress in crowded classrooms, and enabling community micro-enterprises to manage inventory efficiently ⁵⁰. By bypassing the absolute need for massive, centralized computing infrastructure, "Small AI" represents the most viable, scalable pathway for developing nations to leverage machine learning for sustainable development and poverty reduction over the next decade ¹⁸⁴⁷.

What this means for you: Actionable advice for the everyday worker and manager

The rapid, unyielding integration of artificial intelligence into the modern workforce requires an immediate paradigm shift in how individuals manage their professional development and how organizational leaders evaluate technology investments. The popular narrative that "AI will replace humans" is fundamentally inaccurate; the precise economic reality is that humans who are highly adept at collaborating with AI will inevitably replace humans who are not. To secure a competitive edge and ensure long-term relevance in this evolving landscape, everyday workers and managers must internalize several actionable, empirically backed principles:

1. Cultivate Centaur and Cyborg Workflows Professionals must cease interacting with AI as if it were an infallible oracle or a traditional search engine. Instead, it should be treated as an exceptionally fast, highly knowledgeable, but occasionally clumsy intern. Workers should actively adopt the "Centaur" approach by explicitly delegating routine, bounded tasks, such as summarizing long meeting transcripts, drafting initial boilerplate email responses, or generating basic code frameworks, entirely to the AI. This preserves finite human cognitive energy for high-level strategic alignment, complex ethical judgment, and empathetic communication ²⁰²⁷. Alternatively, professionals can practice the "Cyborg" method by integrating AI intimately into the creative process, utilizing it to continually challenge assumptions, suggest alternative analytical frameworks, or debug logic in real-time, thereby utilizing the model as a relentless intellectual sparring partner ²⁰²⁵.

2. Guard Against the Illusion of Speed and Blind Adoption Every worker must remain intensely aware of the jagged technological frontier. Before blindly trusting an AI output, one must critically evaluate whether the task is highly contextual, relies on undocumented internal corporate knowledge, or requires absolute, perfect factual accuracy. If the task falls outside the AI's current capability frontier, relying on it will likely result in subtle, dangerous errors that take significantly longer to audit and fix than generating the work manually from scratch ⁶²⁸. Organizations must establish rigorous verification protocols. An individual's ultimate value to an enterprise no longer lies in their ability to generate content quickly, but in their human capability to audit, fact-check, synthesize, and take absolute accountability for the final output ⁸²⁷.

3. Shift from Outputs to Outcomes For managers, executives, and organizational leaders, the primary metric for AI success can no longer simply be "time saved" or "tasks completed per hour." The rigorous METR developer study conclusively proved that perceived time savings are often entirely illusory, masked by unmeasured verification overhead ²⁸²⁹. Furthermore, if an organization utilizes AI to generate marketing copy 50% faster, but that copy fails to convert customers because it lacks genuine human empathy and distinct brand voice, the AI investment has fundamentally failed to deliver value ⁵¹. Leaders must aggressively shift their measurement frameworks from inputs and efficiency to ultimate business outcomes, such as revenue growth, profit margins, and customer retention ⁵¹. Generative AI delivers real, scalable value only when it is deeply embedded into redesigned, holistic business workflows, rather than being slapped onto existing, legacy processes as an isolated chat interface ⁵¹.

4. Evaluate Organizational AI Intentions When companies implement new AI tools, employees should closely observe the primary motivations behind the deployment. Do executives view AI as a collaborative tool designed to augment capabilities, reduce cognitive load, and elevate the workforce to higher-level strategic work? Or do they deploy granular, AI-driven productivity-monitoring algorithms to squeeze maximum output from every keystroke and optimize labor costs at the expense of culture? Workers should heavily prioritize developing highly transferable human-AI collaboration skills that make them valuable across the broader, global job market, rather than becoming overly dependent on proprietary, company-specific monitoring systems ⁵². To thrive, professionals must lean heavily into uniquely human skills - such as nuanced creativity, ethical reasoning, cross-functional empathetic collaboration, and complex stakeholder management - as these are the exact attributes that current generative models cannot replicate ⁵³⁵⁴.

Bottom line

Generative artificial intelligence is definitively proving to be a massive catalyst for human productivity, but its economic benefits are far from uniformly distributed. At the microeconomic level, the technology acts as a profound skill-leveler, drastically accelerating the operational capabilities of novices and below-average performers while offering limited, and occasionally negative, utility to true domain experts. Because AI operates on a highly unpredictable "jagged technological frontier," it can simultaneously execute wildly complex creative synthesis flawlessly while failing at basic logical contextualization. Consequently, those who succeed in the future of work will not be passive consumers of automation, but strategic, discerning operators who know precisely when to delegate to a machine and when to rely exclusively on human judgment. Zooming out, the macroeconomic reality suggests a bifurcated future: while advanced economies stand to gain modest but meaningful productivity growth, developing nations face a perilous digital divide. Without massive, coordinated global investments in localized digital infrastructure, data sovereignty, and human capital, the AI revolution risks severely exacerbating global economic inequality, prioritizing capital accumulation over labor, and leaving much of the Global South behind.