Arman @MetzYurman

Most debate about AI and the economy focuses on software and jobs. The physical constraint gets less attention: electricity. The International Energy Agency's Energy and AI report projects: 1. Global electricity demand from data centres is set to more than double by 2030. 2. That takes it to around 945 terawatt-hours, slightly more than Japan's entire electricity consumption today. 3. In the United States, data centres are on course to account for almost half of the growth in electricity demand. That turns AI into an infrastructure question: generation, transmission, permitting, and the skilled trades needed to build all of it. The regions that can add power fastest will capture much of the investment and the jobs that come with it. My bet: the binding constraint on AI's economic impact this decade is energy infrastructure, and that makes it a jobs and permitting story.

The usual worry is that AI rewards people who are already skilled. One of the best-identified workplace studies found the opposite pattern. Brynjolfsson, Li and Raymond studied the staggered rollout of a generative AI assistant to 5,179 customer support agents (Quarterly Journal of Economics, 2025): 1. Productivity, measured as issues resolved per hour, rose 14% on average. 2. Novice and low-skilled workers improved 34%, with minimal impact on experienced and highly skilled workers. 3. The tool appeared to spread the practices of the most able workers, and it improved customer sentiment and employee retention. If AI compresses the time it takes to become competent, the value of tenure changes. Training programs, pay ladders and promotion criteria built around years of experience will need rethinking. My bet: AI's biggest workplace effect may be compressing the experience curve, which changes how firms train and promote.

One of the most careful early studies of AI and wages comes from Denmark, and its headline result is a null. Anders Humlum and Emilie Vestergaard linked large-scale adoption surveys to Danish administrative labor records (NBER Working Paper 33777): 1. Most employers in exposed occupations have adopted chatbot initiatives, and workers report productivity benefits. 2. Yet the authors estimate precise null effects on earnings and recorded hours, ruling out effects larger than 2% two years after ChatGPT's launch. 3. What moved was the structure of work: new tasks in content generation, AI oversight and AI integration, and adopters moving into higher-paying occupations where chatbots are more relevant. Their summary is worth repeating: technological change reshapes work well before it surfaces in earnings or hours. If we only watch wage and employment data, we will see the shift late. My bet: the first AI shock shows up in job descriptions, not wage statistics, and policy should track tasks, not just pay.

The AI adoption story looks very different depending on the size of the business. The Census Bureau's Business Trends and Outlook Survey, covering December 2025 to May 2026: 1. 37% of firms with at least 250 employees reported using AI in their business operations. 2. 32% of firms with 100 to 249 employees did, as of the period ending May 3, 2026. 3. The national rate was 19.8%, and fewer than 20% of firms with four or fewer employees reported using AI. Use rose among firms with 20 or more employees but didn't change significantly among smaller ones. If smaller firms keep adopting AI more slowly, productivity gains will concentrate in larger employers, along with the wage and hiring advantages that follow. That is a question for small-business policy, procurement and training programs, not just for technologists. My bet: the next productivity divide in the U.S. runs between large and small employers, not between industries.

Generative AI spread through the U.S. workforce faster than the PC or the internet did. Whether it is lifting productivity is a harder question. The St. Louis Fed's Real-Time Population Survey (November 2025 update) finds: 1. 54.6% of adults ages 18 to 64 used generative AI as of August 2025. Three years after launch, the internet was at 30.1% and the PC at 19.7%. 2. Counting nonusers, workers report time savings equal to 1.6% of all work hours. 3. Fed into a standard production model, that implies generative AI may have raised labor productivity by up to 1.3% since ChatGPT's release. The authors add the caveats that matter: saved time can go to lower-value tasks or on-the-job leisure, and their estimates don't capture firms reorganizing work, which could add more. Individual adoption is fast. Institutional adoption is slower. My bet: the gains are real but small until firms, not just workers, reorganize around the tools.

The headline number on AI and jobs is often the net. The more consequential number is the churn. The World Economic Forum's Future of Jobs Report 2025, based on a survey of more than 1,000 employers across 55 economies, projects for 2025 to 2030: 1. 170 million jobs created, about 14% of today's employment. 2. 92 million jobs displaced, about 8%, for a net gain of 78 million. 3. 39% of workers' existing skill sets transformed or outdated. Employers also say that if the world's workforce were 100 people, 59 would need training by 2030, and 11 of them would be unlikely to get it. A positive net figure can coexist with a great deal of individual disruption. My bet: the policy fight of this decade is over transitions, not totals: who pays to move 92 million people into new work.

If AI were already reshaping the U.S. labor market at scale, it should be visible in the aggregate data. So far, it mostly isn't. The Budget Lab at Yale compared how fast the U.S. occupational mix has changed since ChatGPT's release with earlier technology waves. Its findings, as of October 2025: 1. 33 months in, the broader labor market showed no discernible disruption. 2. During the internet era, the occupational mix shifted by at most around 7 percentage points between 1996 and 2002. 3. The post-ChatGPT path runs only about 1 percentage point above that internet-era path, and the shift began before generative AI arrived. The authors are explicit that this is a snapshot, not a forecast, and that better usage data from AI companies is needed. Historically, technology reshapes offices over decades, not months. My bet: the first clear AI signal will show up in hiring of new entrants, long before it shows up in unemployment.

For most of AI's history, the barrier to using it at work was cost. That barrier is falling faster than almost anything else in the economy. Stanford's 2025 AI Index puts numbers on it: 1. The inference cost of a system performing at GPT-3.5's level dropped more than 280-fold between November 2022 and October 2024. 2. 78% of organizations reported using AI in 2024, up from 55% the year before. 3. U.S. private AI investment reached $109.1 billion in 2024. When a capability gets that much cheaper that quickly, the hard questions move from the IT budget to the org chart. Which tasks get redesigned, which roles get rebuilt, and who is trained to work alongside the tools? Those are management and policy decisions, not technology ones. My bet: the constraint on AI at work is no longer cost. It's how fast organizations can redesign jobs around it.

Most automation waves hit routine work first. AI is different: it reaches high-skilled jobs, which means the richest economies are the most exposed. IMF staff estimate that almost 40% of global employment is exposed to AI. The split by income level is the part worth sitting with: 1. Advanced economies: about 60% of jobs. Roughly half may benefit from AI; for the other half, AI may take over key tasks, lowering labor demand and wages. 2. Emerging markets: about 40%. 3. Low-income countries: about 26%. That means fewer immediate disruptions, but less infrastructure and fewer skilled workers to capture the gains, which raises the risk of falling further behind. The IMF's prescription is unglamorous: social safety nets, retraining, digital infrastructure and regulation that can keep up. Its AI Preparedness Index across 125 countries ranks Singapore, the U.S. and Denmark highest. My bet: AI readiness, meaning safety nets, retraining and digital infrastructure, becomes a competitiveness policy, not a welfare one.

The loudest AI-and-jobs question is "how many layoffs?" The more useful one may be "how many people never got hired?" Stanford Digital Economy Lab researchers, using ADP payroll data on millions of U.S. workers through June 2026, find no evidence of widespread, economy-wide job displacement. But employment of workers aged 22–25 in AI-exposed occupations now stands 19% below where it would be had it kept pace with less-exposed peers. Experienced workers in the same occupations show no comparable gap. Three details matter for employers and policymakers: 1. The gap comes mainly from reduced hiring of young workers, not more separations. 2. It is concentrated where AI substitutes for tasks; where AI complements workers, employment is flat or rising. 3. It has widened steadily since the authors first documented it in August 2025. The authors are careful to call these early, descriptive indicators, not causal estimates. Still, if entry-level roles are where people learn a profession, a quiet hiring freeze at the bottom becomes a skills shortage at the middle a decade from now. My bet: the entry-level pipeline, not the layoff count, is the AI metric that matters.

An AI project needs research beyond the first architecture decision. Before building, understand the provider’s constraints, the dependencies between stages, and the assumptions that could change the design. Then implement in small, useful increments. Let what actually happens inform the next decision. A plan can cover the whole system without committing every part of it to code at once. Each research step should answer something concrete: Which approach fits? What could invalidate it? What evidence would justify moving forward? That’s how I’d pursue speed without lowering the quality bar: reduce avoidable waiting and rework, while keeping the checks that tell us whether we’re making progress.

If AI replaces jobs faster than people and institutions can adapt, the damage won’t stop at unemployment. For a century, we’ve built much of our social system around work: income, identity, status and stability. Rapid displacement could tear that structure apart before a replacement exists. “Learn to use AI” is not a complete answer. Workers must adapt, but businesses and governments must too. Otherwise, we risk enormous productivity gains alongside a profound social and psychological crisis.

My view is that current voice AI makes large-scale replacement of call-center work theoretically feasible. The remaining challenge is enterprise execution: integrating records, transactions, exception handling and reliable oversight. Businesses that get this right could operate at costs that human-heavy competitors struggle to match. That competitive pressure could accelerate adoption. What happens if every call-center operator replaces its human agents within two years? The U.S. telephone-call-center industry employs roughly 303,000 people. The Philippines’ broader customer-experience industry employs 1.68 million. Both figures include roles beyond phone agents, but show the scale involved. The risk is earnings disappearing faster than workers find comparable jobs. That would pressure household spending, local businesses, tax revenues and outsourcing export income. Cheaper services and new jobs could create gains elsewhere. They wouldn’t necessarily reach the same people or places quickly enough. The speed of that transition could determine whether this becomes a manageable productivity shift or a severe economic shock.

My bet: enterprises that rebuild around agents will outscale enterprises that simply give employees AI tools. Some enterprises will run a bad pilot and retreat. Others will add LLMs to existing workflows and get real productivity gains, while keeping the same handoffs, queues and coordination overhead. A third group will redesign how the work gets done: agents executing connected tasks, with humans setting objectives, governing decisions and handling exceptions. I think current models already have enough capability for much repetitive cognitive work. For those workflows, the harder problem is increasingly orchestration: connecting systems, verifying outputs, recovering from failures and deploying reliably at enterprise scale. That changes the hiring equation. More demand for people who build and govern systems. Less dependence on staffing every step of a workflow separately. Better models will help. I wouldn’t wait for another generation before redesigning the work.