The AI productivity paradox

The illusion of value, or why AI automation fails

Everyone claiming that AI automation saves time is only telling half the story. The other half, the time that evaporates checking, correcting, and rewriting what the machine produces, remains invisible in the spreadsheets of enthusiastic managers. Research by Workday among 3,200 employees shows that nearly 40 percent of time saved through AI is lost to rework [1]. Only 14 percent of knowledge workers consistently achieve positive net outcomes when verification time is factored in.

This isn’t an implementation problem that better training can solve. It’s a structural property of how generative AI functions.

The assumption behind every AI business case is elegantly simple: machines take over work, people do less, productivity rises. But work isn’t a bucket of water you can move around. Work is a web of judgments, context, and meaning. When you automate a task, you shift where human input is needed. From doing to checking. From producing to validating. automation

The paradox nobody wanted to see

In 1987, Robert Solow wrote a sentence that has since become a monument to technological unease: computers are visible everywhere, except in the productivity statistics. Economists call this the Solow paradox, and nearly four decades later, the pattern repeats itself with remarkable precision.

Research from the National Bureau of Economic Research describes four explanations for this gap between expectation and reality: false hope, measurement error, redistribution, and implementation lags [2]. The authors conclude that delays likely contribute the most. AI’s most impressive capabilities, particularly those based on machine learning, have not yet spread widely. And more importantly, just as with other general purpose technologies, the full effects will not be realized until complementary investments in processes, skills, and organizational structures have been made.

But there’s a fifth explanation that receives less attention: the possibility that AI doesn’t create less work, but different work. That automation doesn’t eliminate labor but transforms it. And that this transformed labor, the checking and correcting, is heavier than the original work, because it demands constant alertness without the satisfaction of one’s own initiative.

The data support this reading. The Faros AI research report analyzed telemetry from more than 10,000 developers and found that teams with high AI adoption complete 21 percent more tasks and merge 98 percent more pull requests. But the time for code review increased by 91 percent [3]. The bottleneck shifted from production to validation. The machine produces faster than people can check.

When plausible gets mistaken for true

In 1988, Hans Moravec described what is now known as the Moravec paradox: the things people do effortlessly, walking, recognizing faces, understanding context, are exceptionally difficult for machines. While what humans find difficult, chess, mathematics, logical deduction, is relatively easy for computers.

Steven Pinker summarized this with characteristic sharpness: the main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard. Evolution has had billions of years to optimize perception and motor skills. Abstract thinking is a recent addition, perhaps a hundred thousand years old, and therefore less deeply anchored in our cognitive architecture.

This explains why generative AI can write, code, and analyze, but fails at what a four-year-old does effortlessly: distinguishing what’s relevant, understanding what’s missing, sensing when something is off. AI doesn’t produce correct output, it produces plausible output. The difference is fundamental.

Plausibility is statistical probability: words that frequently occur together, patterns that have worked before, answers that look like answers. Truth is something else. Truth requires verification, judgment, the capacity to recognize when statistics mislead you. And precisely those capacities, the domain of ethics and human discernment, are what machines lack and humans must compensate for.

A meta-analysis in Nature Human Behaviour of 106 experiments showed that human-AI combinations on average perform worse than the better of the two working in isolation [4]. Improvements occur only in open-ended tasks like brainstorming. In decision-making and judgment tasks, AI assistance leads to overconfidence or confusion about responsibility. The cyborg fantasy, human and machine united in superior performance, turns out in practice to be a source of new problems.

The invisible labor of checking

Workday’s research reveals an unequal distribution of the verification burden. Employees between 25 and 34 years old make up 46 percent of those who spend the most time checking and correcting AI output. HR professionals represent the largest functional share of heavy rework users. IT functions, by contrast, more often convert AI use into net productivity gains.

This pattern has a logic. Younger employees are assumed to be digitally proficient and receive AI tools first. But they lack the experience to quickly distinguish when output is reliable and when it isn’t. HR work is by definition judgment-heavy and context-sensitive, precisely the domain where AI output most often fails.

The organizational response lags behind. Nearly nine out of ten companies have adapted fewer than half of their roles to AI capabilities. Only 37 percent of the heaviest AI users received additional skills training. Employees are using tools from 2025 within job structures from 2015.

The result is a hidden tax on talent. Cognitive load rises while autonomy and support remain unchanged. The machine accelerates; the human absorbs. impact

The measurement error that distorts everything

Measuring productivity through hours saved is like judging a diet by skipped meals without looking at compensatory snacking. The number looks good; the result disappoints.

When a consultant writes a report in two hours instead of four, the system registers two hours gained. When that same consultant then spends an hour and a half correcting factual errors, the system registers nothing. That time disappears into the general category of professional work. The illusion of value completes itself in the spreadsheet.

This measurement error has strategic consequences. Organizations that define success as hours saved optimize for speed while quality erodes. The California Management Review warns that managers should not treat AI as a general productivity enhancer but as a targeted accelerator that requires skill-diagnostic deployment strategies [4]. That demands dual measurement systems: usage analytics combined with output quality metrics.

Why the promise keeps seducing

Despite the data, the promise of AI automation remains attractive. People are poor at estimating the costs of cognitive labor. Physical labor is visible and measurable. Cognitive labor is diffuse, hidden, exhausting in ways we can barely articulate.

When a machine produces a first draft of a text, that feels like real help. The blank page is filled. That the hours that follow, devoted to refinement and correction, aren’t experienced as AI costs but as normal work is how our minds operate. We overestimate what we see; we underestimate what we feel.

This explains why individual employees are often more positive about AI than organization-wide results justify. 85 percent say they save time, but only 14 percent consistently achieve net-positive outcomes.

The transformation we didn’t choose

Joanna Maciejewska’s viral tweet from March 2024 captured the frustration: the biggest problem with AI is that it’s going in the wrong direction. People want AI to do the laundry so they can write, not the other way around.

The paradox is technically explainable. Moravec’s evolutionary argument still holds: motor skills and perception have been optimized for billions of years, abstract thinking only millennia. AI can write before it can wash because writing, however complex, is cognitively younger than object manipulation.

But the consequences reach further. If AI is best at what we want to do and worst at what we have to do, then automation isn’t liberation but replacement of meaning with work. The creative tasks that provide satisfaction get automated. The verification tasks that are exhausting remain, particularly as autonomous agents amplify this burden through independent decision-making. governance

What working actually is

Here we touch the core that the productivity discussion avoids. The assumption that work is automatable presupposes that work consists of discrete tasks: identifiable, delimitable, transferable. But working isn’t a collection of tasks. Working is a form of attention.

When a lawyer reads a contract, she isn’t just recognizing patterns. She brings decades of experience, intuition about human motives, understanding of what isn’t said. When a doctor makes a diagnosis, that isn’t just symptom matching. It’s seeing the patient, weighing the uncertainty.

AI can recognize patterns. What AI cannot do is give meaning. And meaning isn’t an extra layer on top of task execution, it’s the core of what professionals do.

The return of the human

The conclusion isn’t that AI is worthless. The conclusion is that the value lies elsewhere than where we’re looking. AI doesn’t accelerate productivity; AI transforms the nature of human contribution. From execution to judgment formation. From production to validation.

This isn’t degradation if we design it consciously. Experts who use AI as an amplifier of their judgment can do more with better quality. But that requires recognition of what AI is: a generator of plausibility, not a producer of truth. A tool for first drafts, not for final judgments.

Organizations that understand this invest differently. Not in more licenses but in better training. Not in broader adoption but in deeper strategy integration. They redesign roles around human judgment, instead of expecting employees to compensate for systems that weren’t built to support them.

The question that remains

The illusion of value forces a more fundamental question than “does AI work.” The question is: what are we assuming when we believe that work is automatable?

We assume that efficiency is the same as value. That speed is the same as productivity. That output is the same as outcome. Each of these assumptions is questionable.

The productivity paradox isn’t a technical problem waiting for a technical solution. It’s a symptom of deeper confusion about what work is, what value is, what human contribution means.

The organizations that thrive aren’t those that automate fastest. They’re those that understand most clearly what is irreplaceable. That invest in judgment, not just in tools. That measure what matters, not just what’s measurable.

Seeing through the illusion of value is the first step. The next is recognizing that real value emerges where machines fail: in meaning, in judgment, in the human capacity to distinguish what is plausible from what is true.

And perhaps that’s precisely the insight we needed: not that AI can do too little, but that humans are more than we thought.

Related signals

• Lazy chatbots make people dumb - Shows how the same hidden cost mechanism operates at the individual level: AI assistance erodes professional competence, mirroring the organizational productivity loss you see in verification work.
• Why Your AI Budget Will Cost 89% More Than Your CFO Realizes - Directly addresses the measurement error problem: 95% of AI pilots fail because hidden costs (implementation, rework, retraining) aren’t factored into ROI calculations, proving the productivity paradox isn’t theoretical.
• The Silent AI Takeover Inside Your Business - Explores how AI shifts decision-making authority without explicit handoffs, creating the governance vacuum where verification burdens accumulate and productivity gains evaporate.
• Managing Autonomous AI Agents in the Enterprise - Reveals why autonomous systems amplify the paradox: as AI agents make more decisions independently, oversight and validation become the new bottleneck, transforming work from execution to constant monitoring.

References

[1] Workday. Beyond productivity: measuring the real value of AI. Pleasanton: Workday; 2026. Available at: https://erp.today/workday-research-finds-ai-productivity-gains-are-lost-to-rework

[2] Brynjolfsson E, Rock D, Syverson C. Artificial intelligence and the modern productivity paradox: a clash of expectations and statistics. Cambridge: National Bureau of Economic Research; 2017. NBER Working Paper No. 24001. Available at: https://www.nber.org/papers/w24001

[3] Faros AI. The AI productivity paradox research report 2025. San Francisco: Faros AI; 2025. Available at: https://www.faros.ai/blog/ai-software-engineering

[4] Vaccaro M, et al. Human-AI collaboration: a meta-analysis. Nature Human Behaviour. 2024;8:1892-1905. Available at: https://cmr.berkeley.edu/2025/10/seven-myths-about-ai-and-productivity-what-the-evidence-really-says/