The AI Productivity Paradox Is Back

Why a $700 Billion AI Industry Still Can’t Replace You In the late 1800s, when electric motors first arrived in factories, many owners simply bolted them onto their old steam‑powered layouts. Little about the work itself changed. It took decades before they redesigned the factory floor around electricity. When they finally did, productivity rose sharply. In 1987, Nobel Prize–winning economist Robert Solow remarked that “you can see the computer age everywhere but in the productivity statistics.” Companies had been pouring money into information technology for years, yet productivity in advanced economies remained stubbornly weak. Robert Solow remarked stubbornly weak The “productivity paradox” lingered for roughly a decade, until a clear productivity boom showed up in the United States in the mid‑1990s. productivity paradox In February 2026, Torsten Slok, Chief Economist at Apollo Global Management, put it bluntly: “AI is everywhere except in the incoming macroeconomic data.” AI is everywhere New buzzword, same observation, nearly forty years later. MIT researchers describe adoption of major technologies, including AI, as a productivity J‑curve: many firms see performance and productivity dip at first before any gains appear. adoption of major technologies, including AI a productivity J‑curve productivity dip Why? Because AI does not just slot into existing processes. It demands new infrastructure, new training, and new ways of organizing work. Without those changes, even the best technology adds drag. It creates new problems instead of solving old ones. Put simply, things get worse before they get better. Stanford economist Erik Brynjolfsson believes the United States may already be turning the corner. He has predicted that U.S. productivity growth will be about 2.7% in 2025—nearly double the average of the previous decade—and sees that as a sign AI‑related gains are beginning to show up. But a recent National Bureau of Economic Research study of about 6,000 executives across several countries tells a different story inside companies. More than two‑thirds of those executives say they use AI at work, but on average for only about 1.5 hours per week, and over 80% of firms report no impact on either productivity or employment so far. Research study 6,000 executives across several countries 80% of firms report no impact productivity or employment so far The pattern is clear. Technology alone changes little. The reorganization of work around the technology changes everything, and that reorganization usually takes a decade or more. History suggests that major technologies often take a decade or two before they show up as broad, economy‑wide transformation. With the current wave of generative AI, we are only a few years into serious deployment. The hype talks as if decades have passed. The Vibe Economy Meets Reality You may have seen the essay “Something Big Is Happening.” It has racked up tens of millions of views across social platforms, and its message is simple: AI will take over a huge share of today’s jobs, faster than you think, so you should prepare now or be left behind. Something Big Is Happening It was written by Matt Shumer, who runs an AI company and helps manage an AI‑focused venture fund. AI researcher Gary Marcus pushed back hard, arguing that the essay had “not a shred of actual data” behind its sweeping claims. Gary Marcus pushed back hard The people selling AI need you to believe it will replace everything. That belief helps justify the hundreds of billions of dollars in capital expenditure that companies like Alphabet, Microsoft, Meta, and Amazon plan to pour into AI infrastructure around 2026. need you to believe If AI only makes your spreadsheets a little faster, nobody gets that money back. But if AI replaces lawyers, doctors, accountants, marketers, and managers, suddenly the investment makes sense. The stock prices make sense. The valuations make sense. So they sell you a future where you are the cost to be cut. The problem is that this story runs on vibes and opinions, not on the numbers coming out of places like McKinsey, MIT, the NBER, Gartner, Deloitte, Challenger, and Wyndham. McKinsey’s 2025 State of AI report found that 88% of organizations now use AI in at least one business function. Yet only about 6% qualify as true high performers, where AI contributes more than 5% of enterprise‑level EBIT. That NBER‑linked study found that more than 80% of firms report no measurable impact from AI on productivity or employment so far. MIT researchers have estimated that today’s AI tools could technically take over tasks worth about 11.7% of the U.S. labor market, or roughly $1.2 trillion in wages. But industry surveys suggest that most generative AI pilot projects never make it out of the experimental phase and into scaled production. A Deloitte‑cited survey found that 47% of business executives have already made at least one major decision based on faulty or fabricated AI output. Gartner predicts that more than 40% of ambitious “agentic AI” projects will be scrapped by 2027 because costs keep rising, business value stays fuzzy, and risk controls lag behind. In hospitality, Wyndham reported in January 2026 that 98% of hotels have begun using AI somewhere in their operations, but only 32% have it embedded across most of their systems. Roughly three‑quarters of hoteliers say they feel unsure or overwhelmed about how to scale it effectively. Challenger, Gray & Christmas, an outplacement firm, reported that total U.S. job cuts reached about 1.2 million in 2025, up 58% from the year before, with tens of thousands of those layoffs explicitly attributed to AI. McKinsey’s 2025 State of AI report found that 88% of organizations now use AI in at least one business function. Yet only about 6% qualify as true high performers, where AI contributes more than 5% of enterprise‑level EBIT. McKinsey’s 2025 State of AI report That NBER‑linked study found that more than 80% of firms report no measurable impact from AI on productivity or employment so far. NBER‑linked study MIT researchers have estimated that today’s AI tools could technically take over tasks worth about 11.7% of the U.S. labor market, or roughly $1.2 trillion in wages. But industry surveys suggest that most generative AI pilot projects never make it out of the experimental phase and into scaled production. MIT researchers have estimated industry surveys suggest A Deloitte‑cited survey found that 47% of business executives have already made at least one major decision based on faulty or fabricated AI output. Deloitte‑cited survey Gartner predicts that more than 40% of ambitious “agentic AI” projects will be scrapped by 2027 because costs keep rising, business value stays fuzzy, and risk controls lag behind. Gartner predicts In hospitality, Wyndham reported in January 2026 that 98% of hotels have begun using AI somewhere in their operations, but only 32% have it embedded across most of their systems. Roughly three‑quarters of hoteliers say they feel unsure or overwhelmed about how to scale it effectively. Wyndham reported Challenger, Gray & Christmas, an outplacement firm, reported that total U.S. job cuts reached about 1.2 million in 2025, up 58% from the year before, with tens of thousands of those layoffs explicitly attributed to AI. reported that total U.S. job cuts 58% from the year before explicitly attributed to AI The vibe says something big is happening. The numbers say something big is missing. The Part They Never Explain There is one piece of this puzzle that almost never gets discussed honestly. Not in viral essays. Not in investor decks. Not on keynote stages. It is what happens after you buy the AI. Take a simple case. A hotel wants to reach customers online. There is nothing exotic about that. It is a basic business challenge. There are three kinds of AI tools this hotel could use. AI tools Level 1: AI as a Faster Pen Tools like ChatGPT can write listings, ad copy, and marketing content at speeds no human can match. A hotel manager can use ChatGPT to draft ten different listings for ten types of travelers in the time it once took to write one. That is real value. But someone still has to prompt the AI, read what comes out, choose what fits, copy it into the booking platforms, and check that it displays correctly. The tool does the writing. It does not do the work. Level 2: AI as a Semi‑Autonomous Assistant Agentic AI tools can string steps together. Write a listing, post it to multiple sites, track performance, maybe even tweak copy based on results. Agentic AI tools On paper, it sounds like the leap everyone is waiting for. In practice, it demands that the hotel manager know which platforms matter, what the AI is actually doing, and how to verify the results. Their expertise is running a hotel, not orchestrating software agents. Deloitte reports that 62% of companies are experimenting with this kind of AI. Only 23% have scaled it successfully. Gartner predicts over 40% of these projects will be abandoned by 2027. The tools are not plug‑and‑play. They are work. Level 3: AI as a Specialized Platform Then there are industry‑specific platforms. A hotel management suite that promises to handle “everything” automatically once you feed in your information. This is the promised endgame. Yet Wyndham’s 2026 numbers show that while 98% of hotels have started using AI somewhere, only 32% have it operating across their systems. Seventy‑three percent are overwhelmed. Even with a powerful platform, somebody has to manage integrations, handle exceptions, keep information accurate, and watch for when the system quietly fails. And this is just one channel: web presence. The same hotel also has social media, email marketing, booking engines, review platforms, and more. Each tool adds complexity. Each one demands attention. The hotel manager still has a hotel to run. Guests to serve. Staff to schedule. Maintenance to supervise. Bills to pay. The problem was never just skill. It is time, attention, and care. A 2025 report from digital marketing firm WSI found that time remains the number one barrier for 35% of businesses trying to adopt AI. A 2025 report In February 2026, IBM Chief Human Resources Officer Nickle LaMoreaux announced the company is tripling its entry‑level hiring. Her logic was blunt. The companies that will be most successful three to five years from now are the ones investing heavily in human talent now, not cutting it. announced the company is tripling its entry‑level hiring Automating tasks does not erase the need for people. It changes what those people do. What the Machines Cannot Fake Researchers at MIT Sloan went to U.S. Bureau of Labor Statistics data and asked a simple question: where do humans have staying power? They grouped the answer into five capabilities and called the framework EPOCH. the framework EPOCH Empathy: the ability to understand what another person feels. Presence: the ability to show up, collaborate, and build trust in real time. Opinion: the ability to make ethical and moral judgments. Creativity: genuine imagination and intuition. Hope: the capacity for leadership, purpose, and vision. Empathy: the ability to understand what another person feels. Presence: the ability to show up, collaborate, and build trust in real time. Opinion: the ability to make ethical and moral judgments. Creativity: genuine imagination and intuition. Hope: the capacity for leadership, purpose, and vision. These are not nice‑to‑have add‑ons. They are the glue holding real business relationships together. A client who insists on a specific color for their brand because it ties to a family story. A hotel guest who returns year after year because the receptionist remembered their name and their kid’s allergy. A manager who handles crises better because they lived through them, not because an AI summarized leadership quotes. Consumers already know this. In one 2025 survey of U.S. consumers, 93.4% said they prefer interacting with a human over AI in customer service, and 84% said humans are more accurate. Other recent research finds that most people still suspect companies deploy AI mainly to cut costs, not to improve service. 2025 survey of U.S. consumers AI in customer service Other recent research A study in the Journal of Interactive Marketing analyzed 887 real chat transcripts and ran experiments with 989 participants. A study in the Journal of Interactive Marketing The verdict was clear. AI chatbots perform well in simple, transactional situations. When the job is answering a straightforward question or processing a form, AI can handle it. The moment the interaction demands trust, warmth, or the sense that someone actually cares, human agents win. Customer satisfaction goes up. Repeat business goes up. Loyalty goes up. The replacement story pretends most business interactions are simple transactions. They are not. They are relationships. As a February 2026 Forbes article put it, AI frustration is the new customer service crisis. February 2026 Forbes article Service interactions do not succeed because they are automated or fast. They succeed when they are clear, accurate, and centered on the human experience. Companies deploying AI to shave costs are solving for their own margins. The people they serve are asking for something else. People go back to the same hotel because a person made them feel welcome. They trust a brand because a person behind it communicated with honesty and personality. People remember people. Nobody remembers how an algorithm treated them. Nobody feels loyalty to a software product. AI’s Achilles Heel Anyone who has used AI tools deeply already knows the problem the industry tries to bury under jargon. AI lies. The industry calls it hallucination to make it sound mysterious. The truth is simpler. AI systems fabricate things all the time, with full confidence and perfect grammar. Recent enterprise risk analyses drawing on Deloitte’s survey work estimate that about 47% of business executives have already made major decisions based on faulty AI content. Recent enterprise risk analyses Employees in firms using AI now spend roughly 4.3 hours per week checking AI output—about $14,200 per employee per year in verification overhead. The average cost of a single serious hallucination ranges from around $18,000 in customer‑service cases to roughly $2.4 million in healthcare. The average cost This is not a bug waiting for a patch. It is baked into how current systems work. AI language models generate text by predicting which word is most likely to come next. They are not checking against a ground truth database. They are optimizing for what sounds plausible, not what is accurate. The European Union’s AI Act, whose obligations for high‑risk systems start to bite in 2026, is an explicit acknowledgment of this. It mandates documented risk management, human oversight, and formal conformity assessments for high‑risk AI. The European Union’s AI Act Lawmakers concluded the error problem will not fix itself. The U.S. Air Force reached the same conclusion in a very different domain. In testing its Maven Smart System to speed up targeting decisions, the Air Force’s explicit design principle was to have the machine support, not replace, the human in the loop. Even in a domain built on speed and lethality, the military insists on human oversight. The spectrum runs from a throwaway social post at one end to a military strike at the other. Most business decisions sit somewhere between: accounting, customer interactions, brand representation, strategy. In all of these, the acceptable error rate is much lower than the hype suggests. The industry has no serious answer to basic questions. How do you know when the AI has made a mistake? Who is responsible for checking its work? How far back do you go to unwind errors once you find them? Who carries the blame when the machine’s mistake costs real money or lives? These are not theoretical puzzles. They are daily headaches in every organization that has tried to hand the keys to a machine. Anyone who has spent time trying to “train” ChatGPT to obey simple instructions—stop using a certain format, avoid certain phrases, stick to a specific structure—only to watch it drift minutes later knows this at a gut level. AI does not learn like a person. There is no shame, no fear of being fired, no deep memory that says “never again.” No consequence. No lasting change. You Were Meant To Struggle A family sits down to pick a movie. Nobody can agree. People throw out titles the others have never heard of. The conversation wanders into old stories, running jokes, half‑remembered moments. It takes twenty‑five minutes to settle on something. Now imagine a perfectly efficient recommendation app. It quietly reads everyone’s data, scans the streaming catalog, and serves the mathematically optimal movie in seconds. Faster. Smarter. More optimized. And the family sits down in silence, watches the movie, and misses the part that actually mattered. The friction was the point. The search created the conversation. The conversation created the connection. UCLA psychologist Robert Bjork calls these desirable difficulties. A desirable difficulty is a challenge that feels harder in the moment but produces better learning, deeper understanding, and longer‑lasting growth. desirable difficulty Spacing your studying instead of cramming, mixing different problem types instead of repeating one, or getting delayed feedback instead of instant correction all feel less efficient in the short term yet lead to stronger performance in the long term. AI does the opposite. It delivers polished output immediately. It skips the struggle. And in skipping the struggle, it skips the learning. Harvard‑linked researchers have documented the IKEA Effect: people reliably value things they helped build far more than identical things they did not assemble themselves. IKEA Effect Effort enhances meaning. Achievements feel more personal and significant when you have had to work for them. People value outcomes more when they invest real effort. Psychologists Edward Deci and Richard Ryan, through Self‑Determination Theory, identified three core human needs. Self‑Determination Theory three core human needs Autonomy: the need to direct your own life. Competence: the need to grow your abilities through effort. Relatedness: the need to connect deeply with others through shared experience. Autonomy: the need to direct your own life. Competence: the need to grow your abilities through effort. Relatedness: the need to connect deeply with others through shared experience. AI that tries to do everything for you cuts against all three. It erases your choices. It blocks your path to mastery. It steals the shared process that bonds people. There is a reason climbing Mount Everest and taking a helicopter to the summit do not feel the same, even if the view is identical. For millions of business owners and managers, the work is not just a means to a paycheck. The work is the point. Solving problems. Building something with your own judgment and hands. Growing through difficulty. Forming relationships along the way. Wiping that away in the name of “efficiency” does not liberate people. It hollows them out. The World Is Bigger Than the Valley More than twenty years ago, Marc Andreessen said “Software will eat the world.” Recently the line has been updated to “AI will eat the world.” What people in Silicon Valley often forget is that the world is bigger than their friend group. As of late 2025, roughly 2.5 billion people are still offline or lack reliable internet access, while well over 5 billion are online—but their access varies massively. roughly 2.5 billion people About 94% of people in high‑income countries use the internet, compared with just 23% in low‑income countries. In many cities around 85% of people are online, versus well under two‑thirds in rural areas. The internet is slowly becoming a basic utility for work, education, and finance. But the reality on the ground is uneven. Around three‑quarters of small businesses now have a website, but many of those sites are bare‑bones. Only a minority—less than 20%—have what researchers would call an “advanced digital profile,” meaning a genuinely effective, high‑performing online presence, and only about a third report meaningful online sales. Around three‑quarters of small businesses Roughly four in ten small firms say they plan to invest in improving their websites over the next year. By 2025–2026, subscription models dominate software and content. Roughly a third of internet users say they pay for recurring software‑based services. Surveys also show that a majority of people have at least one paid subscription they rarely or never use, and a substantial share—around a third—have canceled at least one subscription in the past year because of economic pressure. paid subscription they rarely or never use Even if AI wiped out the entire current software‑as‑a‑service landscape—and that is a huge if—it would not “upend the world” on the timeline the industry keeps promising. Technologies that have been around for decades are still slowly spreading and taking shape. The world does not bend to startup pitch decks and investor dreams. The Age of AI, or the Age of Expensive Disappointment? If businesses keep adopting AI at the current pace without reorganizing around it, the most likely future is a long, costly period of frustration. High adoption. Low impact. Growing resentment. Many companies are using AI to cut costs rather than solve problems. Customers see it, feel it, and resent it. The gap between what AI promises and what it delivers keeps widening. In this scenario, AI adoption stays broad but real impact stays confined to a small minority—the 6% of companies McKinsey already classifies as true high performers today. McKinsey already classifies Jobs shift more than they vanish. Net displacement lands in the 6% to 7% range that Goldman Sachs has estimated for the global workforce. Small businesses keep stumbling through integration headaches. Consumer anger at AI customer service keeps climbing as Europe’s AI Act reins in the sharpest edges. Goldman Sachs has estimated While hundreds of billions of dollars a year flow into AI, more than four out of five companies still report that they cannot see any clear productivity or employment impact at all. still report And in the middle of this, real people lose real jobs, not because AI actually replaced their work, but because executives believed it would. Harvard Business Review ran an article in January 2026 titled “Companies Are Laying Off Workers Because of AI’s Potential, Not Its Performance,” reporting that roughly 60% of surveyed organizations had already reduced headcount based on what they expect AI to do someday, not what it is doing today. article in January 2026 There are other possibilities. A smart minority of companies might figure out how to truly restructure around AI. Governments might invest in serious retraining. Firms might follow IBM’s lead and invest in humans and AI together. Productivity might finally climb its J‑curve, the way factories did in the 1920s. Or a catastrophic AI failure in healthcare, finance, or infrastructure could trigger a regulatory freeze and a consumer backlash. Alternatively, a true breakthrough could make domain‑specific AI genuinely autonomous in entire sectors. Either shock would create rapid change. But the real danger lurks in the background. Middle management is where future leaders are forged. It is where people learn to make real decisions without clear right answers, to manage conflict, to earn trust across departments. Yet Korn Ferry reports that more than four in ten companies plan to replace roles with AI, including 37% that are specifically targeting entry‑level positions—the very pipeline where tomorrow’s leaders are supposed to start. Korn Ferry reports In the short term it looks cheap, fast, efficient. In the long term, it is a disaster. You produce a generation of senior leaders who have never been tested. People who perform well when everything is organized and clear. People who can operate a dashboard but not navigate a crisis. Nickle LaMoreaux of IBM said it plainly: the companies that will win are the ones investing in human talent now, not cutting it. “We are tripling our entry‑level hiring—and yes, that is for software developers and all these jobs we’re being told AI can do.” said it plainly The Future Belongs To You At every level of capability, AI needs humans. It needs supervision, judgment, and correction. The productivity gains lag years or decades behind the investment. Removing struggle from work removes the learning that strengthens people and organizations. Customers consistently prefer human warmth over automated efficiency. And the leadership pipeline depends on the messy, difficult, human experiences that AI is built to bypass. Most AI analysis treats business like a machine. Inputs go in, outputs come out, and the only things that matter are speed and cost. But businesses are not machines. They are living networks of human relationships running on trust, care, judgment, and shared experience. AI is very good at the base of the work. Logistics. Templates. Data processing. That layer will keep improving. The top layers are different. The layers that make a customer come back, convince an employee to stay, make a brand actually mean something, and make the work worth doing. Those layers are produced by people who care. people who care The AI profiteers trying to sell us a robotic, hollow future where we cast aside the most meaningful parts of our humanity will fail. The next wave of businesses will be built by people who understand that people remember people, and AI, like many technologies before it, will soon be relegated to the junkyard of obsolescence. Sources: Frictionless AI comes at a human cost to learning, growth and connection AI-Powered Customer Service Fails at Four Times the Rate of Other Tasks CEOs aren’t seeing any AI productivity gains, yet some tech industry leaders are still convinced AI will destroy white collar work within two years Caregiving in Philosophy, Biology & Political Economy The US Air Force let AI help operators find targets to speed up kill chain decisions US Air Force Tests AI to Shrink the Kill Chain Frictionless AI comes at a human cost to learning, growth and connection Frictionless AI comes at a human cost to learning, growth and connection AI-Powered Customer Service Fails at Four Times the Rate of Other Tasks AI-Powered Customer Service Fails at Four Times the Rate of Other Tasks CEOs aren’t seeing any AI productivity gains, yet some tech industry leaders are still convinced AI will destroy white collar work within two years CEOs aren’t seeing any AI productivity gains, yet some tech industry leaders are still convinced AI will destroy white collar work within two years Caregiving in Philosophy, Biology & Political Economy Caregiving in Philosophy, Biology & Political Economy The US Air Force let AI help operators find targets to speed up kill chain decisions The US Air Force let AI help operators find targets to speed up kill chain decisions US Air Force Tests AI to Shrink the Kill Chain US Air Force Tests AI to Shrink the Kill Chain