On-Base Percentage for an Uneven Labor Market Max Effgen, September 22, 2026September 16, 2026 The useful idea in Moneyball was never that scouts were stupid. It was that the market was paying for the wrong stat. Oakland could not outbid the Yankees for home runs, so Billy Beane paid for on-base percentage — a cheaper way to not make an out. Michael Lewis’s 2003 book is a story about a budget constraint meeting a measurement error. That is closer to the present labor market than the phrase “AI job apocalypse.” Artificial intelligence is changing the price of certain kinds of work. It is not flattening the entire workforce. The World Economic Forum’s Future of Jobs Report 2025, built from an employer survey, projects 170 million jobs created and 92 million displaced by 2030, a net of 78 million. Those are not census counts. They are a structured guess about how firms think the next five years will go. Treat them as a direction, not a headcount. The tighter evidence is narrower and more uncomfortable. Stanford’s Digital Economy Lab, using ADP payroll data, still finds no economy-wide collapse. What it does find is a growing gap for workers ages 22–25 in highly AI-exposed occupations. By June 2026 that group sat about 19 percent below where it would have been if it had kept pace with same-age peers in less-exposed jobs. The earlier public versions of the paper had already shown a hit to entry-level hiring rather than a wave of mid-career firings. Experienced workers in the same occupations look much more stable. The adjustment is happening at the door, not only in the layoff announcement. PwC’s 2025 Global AI Jobs Barometer points the other way for people already inside the building who can show AI-related skills: a 56 percent wage premium, and wages rising twice as fast in more-exposed industries. Job counts in those industries still grew. The skills employers list in those jobs are changing faster than in the rest of the market. That is a market that is repricing tasks, not deleting work as a category. Read those three findings together and the Moneyball analogy holds only if you keep it precise. The undervalued asset is not “a skill robots cannot steal.” Almost no interesting skill is permanently unstealable. The undervalued asset is work that is hard to specify, hard to check, and expensive to get wrong — plus the ability to use the new tools without pretending they replaced judgment. What the Apocalypse Story Gets Wrong Mass-displacement headlines collapse two different questions. Can a model draft an email? Yes. Can a firm fire the person who knew which email not to send? Sometimes, and then it discovers the cost in the next quarter. The Stanford pattern is the one to watch. If generative tools compress the first two years of a knowledge job — the years that used to be spent writing the first draft, cleaning the file, building the slide — firms hire fewer juniors and keep the people who already know what good looks like. That is not an apocalypse. It is a thinner apprenticeship. Software, analysis, customer-facing writing, and junior design are where that shows up first because the output is already digital. The WEF net-job figure can be true in the same world. Care, construction, logistics, energy, and a long list of frontline roles are not sitting inside a text box. Technology roles grow in percentage terms. Clerical and some customer-service titles shrink. The average disguises the fact that a 24-year-old analyst and a 24-year-old electrician are not in the same market. Carnegie is a weaker import than Beane. How to Win Friends and Influence People is a book about attention and status in rooms. Those things still matter. They are not a labor-market strategy by themselves. Empathy without a domain is hospitality. The hybrid that actually shows up in hiring is narrower: someone who can read a model’s output, argue with it, and then get other humans to act. Four Skills That Are Mispriced, Not Magic Judgment under a bad dashboard. Someone has to decide whether the model’s answer is usable. That is audit, QA, compliance, and editorial taste wearing different badges. Bias review and governance are real hiring categories. They are also easy to staff with theater — a policy PDF and no authority. The version that holds value is closer to Beane’s plate discipline than to a new job title: you measure the system against an outcome you care about, and you are willing to kill a convenient number. Translation across a room. EQ-as-leadership is the brochure version. The working version is getting a technical team, a risk team, and a customer to agree on what “done” means when the tool keeps moving. That is not smiling. It is specification. Firms that “adopt AI” without this person accumulate half-finished automations. Taste with a constraint. Creative work that is only vibe is exactly what image and text models flood. Creative work that is a constrained problem — this audience, this budget, this legal line, this brand that already exists — is still scarce. The model can produce options. It cannot be the client. Negotiation when the other side also has a model. Vendor contracts, internal resource fights, and customer exceptions are where the tool is now on both sides of the table. The human job is not charm. It is knowing which term is actually expensive. None of these are safe forever. They are expensive to fake, which is why they still clear a market. A résumé that invents “audited 15 models and saved $50K in fines” is the opposite of sabermetrics. Beane’s point was better measurement of real events. If you cannot point to a shipped workflow, a rejected output, a retained account, or a cycle time that moved, you do not have an on-base percentage. You have adjectives. The Other Side of the Same Market The entrepreneur story is the same pricing error from the other chair. For twenty years a small firm bought time in the form of people: a writer, a support inbox, an analyst, a designer, a person to build the deck. Generative tools do not eliminate those functions. They change the minimum viable staff. A founder can now draft, summarize, support a first-line FAQ, and produce a passable visual much faster than in 2019. That is leverage. It is also how junior hiring gets skipped. Use the tools that way and you are running a Moneyball roster: cheap outs avoided, not a 40-man payroll. The mistake in the second source draft is treating a long menu of vendors as strategy. Most of those names will churn. The durable move is smaller. Pick one workflow that already exists — inbound questions, first-draft marketing, a weekly numbers pull — and put a model on the first pass. Keep a human on the exception and the send button. Measure hours returned, error rate, and whether customers noticed. If you cannot measure those three, you did not adopt AI. You subscribed to software. Customer-service automation is the cleanest example and the easiest to overdo. A bot that answers order status is a good out. A bot that invents a refund policy is a loss. Data tools that forecast churn are useful if you already have a retention motion. Generated video is cheap until the seventh clone of the same avatar shows up in the same feed. The constraint is still taste and liability. Affordable tiers matter because they lower the cost of an experiment. They do not lower the cost of a bad process. A bootstrapped company that automates a messy inbox just produces mess at higher volume. What to Do With the Template Instinct People want a checklist. Fine. Make it boring. For a mid-career worker: keep a log of work the model drafted and work you changed. The delta is the job. Learn enough of the tools in your field to stop being the person who pastes raw output. Put one number on a résumé that a manager could verify. For a junior worker in an exposed field: the apprenticeship just got shorter and meaner. Seek rooms where you see the full cycle — problem, draft, correction, customer — not rooms that only want the draft. Adjacent skills that still require being on-site or on the hook (implementation, sales engineering, operations, care, trades tied to physical plant) are not a consolation prize. They are the part of the WEF forecast that does not depend on a chatbot. For a founder: do not hire a “head of AI.” Hire the next unit of work you were going to hire, and see whether a tool ate part of it. If it did, keep the savings. If it did not, you learned something cheaper than a transformation program. The Line That Survives Beane did not win because he loved spreadsheets. He won, for a while, because the rest of the league paid for theater. The AI labor market has the same shape. Firms pay a premium for people who can use the tools. They hire fewer people whose job was to be the first draft. They still need someone who knows when the draft is wrong. That is a smaller, sharper disruption than an apocalypse and a harder one than a TED Talk about empathy. The robots are not stealing careers, but they are changing which stats the market will still buy. Measure the ones that show up in the box score. Ignore the ones that only look good in the scouting report. Sources 1. Michael Lewis, Moneyball (2003): undervalued on-base skills as a response to a budget constraint and a measurement error in player evaluation. 2. World Economic Forum, Future of Jobs Report 2025: employer-survey extrapolation of 170 million jobs created and 92 million displaced by 2030 (net 78 million); 59 percent of workers projected to need reskilling. 3. Brynjolfsson, Chandar, Chen et al., Stanford Digital Economy Lab, “Canaries in the Coal Mine” update (August 2026): no economy-wide displacement in ADP payroll data; ages 22–25 in high AI-exposure occupations about 19 percent below a less-exposed peer path as of June 2026. 4. PwC, 2025 Global AI Jobs Barometer: 56 percent average wage premium for listed AI skills; wages growing twice as fast in more AI-exposed industries; faster skill churn in exposed occupations. 5. U.S. Bureau of Labor Statistics occupational projections remain the baseline for specific titles (including software development growth above the all-occupation average); treat title-level forecasts as separate from generative-AI exposure studies. 6. Practical constraint on “AI-resistant” lists: resistance is about current task mix and error cost, not a permanent exemption from automation. “Do it yourself; do not wait for the institution to price you correctly.” ~ Ian MacKaye Avanti. Measure what matters. ...