Hook
McKinsey says AI will create more American jobs than it cuts. Eleven million workers may need new careers. The number is a net figure, and that is the entire problem.
Thirteen years of watching markets absorb structural shocks has taught me one invariant: the aggregate headline prints first, the distribution is discovered second, and the financing gap is discovered last — by whoever is holding the liability when it clears. A net employment number is an accounting identity, not a forecast. It gives you the sign of a difference and nothing about the gross flows that generate the pain.
I watched this exact concealment run in DeFi for eighteen months. Net TVL climbed while gross redemptions exceeded gross deposits every single day. The dashboard said growth; the plumbing said a run. Netting is the accounting of people who never had to bear the gross.
So when a consulting firm leads with "creates more than it cuts," I do not read relief. I read a subtraction I have not been shown.
Context
The source material is a trade-press brief: five extracted points, four of them opinion, one of them a number. The number is eleven million. The underlying research is McKinsey Global Institute's July 2023 work on generative AI and the American labor market, whose headline figure is closer to "up to about 12 million US workers may need to change occupations by 2030." The brief rounds down and strips everything else — no methodology, no time span, no scenario weighting, no adoption curve.
That omission is not cosmetic, because the entire "net positive" conclusion rests on one methodological choice: task-level automation instead of job-level substitution. Decompose an occupation into activities, score each activity for automation potential against language-model capability, then re-aggregate. Under that frame a role survives as long as a single human task remains inside it. The result is not a measurement of displacement. It is a measurement of task overlap with a model class, converted back into headcount through an adoption rate that the brief never states.
Two more absences deserve naming. The research is silent on where the boundary of model capability actually sits — GPT-3.5, GPT-4, or something shipping in 2027 — and that boundary is the single most sensitive parameter in the whole exercise. It is also silent on who pays for the transition it describes. Both silences are load-bearing.
The brief's provenance compounds this. It is a summary of a summary, filtered through a crypto publication with no labor-economics desk, where the operative incentive is traffic, not precision. Every layer of paraphrase is a layer of fidelity loss. What "McKinsey says" in the headline is not what McKinsey wrote in the report.
I will treat the eleven million as a lower bound on gross churn, not an upper bound on net pain. Then I will price the gap.
Core
Start with the arithmetic the brief suppressed. Eleven million career changes between now and 2030 is roughly 1.6 to 1.8 million per year. Against a US labor market whose gross monthly flows — hires plus separations — run in the five-to-six-million range, that is not a catastrophe in aggregate.
The damage is never in the aggregate. It is in the concentration.
Every cross-institution study I have cross-referenced points the same direction: office support, customer service, food service, entry-level white-collar documentation. Language-dense, structured, low-physical. Meanwhile care work, skilled trades, STEM, and high-interpersonal roles stay flat or grow. The shock runs down the wage and education gradient. It does not spread evenly, and an average cannot describe a gradient.
That line is not idle. US job openings and labor turnover have run in the five-to-six-million monthly flow band for years, and the post-pandemic quit rate already pushed annual occupational switching above its pre-2020 baseline. The eleven million is therefore not a clean AI increment. It is a stack of AI-driven churn on top of a structural churn rate that was already rising. The brief merges the two and presents the sum as a single AI signal. That is the difference between a marginal shock and a total one, and it changes every policy response.
Now add geography. The replacement roles cluster in metros with AI capital, data infrastructure, and compute. The displaced roles are distributed across the country in smaller labor markets. Retraining a worker is a six-to-twenty-four-month project. Relocating one is a different problem entirely, and the report does not model it. Skills are portable across a training certificate. Lives are not.
Wage insurance and benefit portability are the tools that would actually address this. Both are absent from the brief, and both are politically harder than training grants, which is precisely why the recommendation defaults to training.
Then the financing gap, which is the part I actually care about. Eleven million transitions at a private cost of three to ten thousand dollars each implies thirty-three to one hundred ten billion dollars of reskilling expenditure. The United States currently funds workforce development through WIOA and adjacent programs at roughly ten billion a year. That is a one-order-of-magnitude gap between the liability the research implies and the apparatus that exists to service it. A forecast is a liability issued against your own uncertainty.
Who closes a gap like that? Three candidates, and the choice is the real story. Firms, if the replacement cost of a trained worker exceeds the retraining cost. Governments, if the political cost of unmanaged churn exceeds the fiscal cost of prevention. Or households, drawing down savings and absorbing wage cuts. The brief mentions "policy and training" without assigning the bill. That omission is not an oversight — it is the most contested question in the entire report, and it happens to be the one a corporate strategy consultancy is structurally least able to answer cleanly.
There is a second-order market here that the brief never names. Eleven million transitions imply a reskilling market of tens of billions annually, and capital will not wait for policy. Watch for tokenized credentialing, on-chain skill attestation, and talent-marketplace rails — not because the technology is elegant, but because the demand is inelastic and the incumbent apparatus is an order of magnitude too small. My caution is the same one I apply to every yield product: check the maturity of the liability against the maturity of the asset. A training-finance product that funds eighteen-month programs with thirty-day deposits is a maturity mismatch wearing a mission statement. It survives a bull market and fails in the first drawdown.
Write down the macro transmission, because it is where this becomes a crypto question. Labor displacement is disinflationary on wages. A weaker wage-negotiation position for the bottom two income quintiles compresses compensation, which compresses the inflation impulse, which loosens the Fed's reaction function. Liquidity expands. Bitcoin, which behaves like a liquidity sponge with a beta attached, reprices upward. That is the benign path.
The opposite path is equally legible. The AI buildout is energy-intensive and compute-constrained. Power, cooling, and advanced packaging are inflation-additive. If labor disinflation and compute inflation roughly cancel, the Fed has no clean signal and defaults to its reaction function on the labor side — because that is the mandate it can actually measure. Either way, the terminal variable is the same: the path of aggregate demand.
And here is what the brief never touches. If the transition is financed by households drawing down balance sheets, aggregate demand falls faster than AI productivity raises supply. That is the deflationary tail — and it is the tail that forces the most aggressive easing, which is the most bullish liquidity outcome for hard assets. If instead it is financed by fiscal expansion, duration grows and the term premium reprices. Both paths route through the same market. Neither is in the bullet points.
The dispersion across forecasters is the last signal, and I consider it the most honest one. Goldman placed roughly 300 million full-time-equivalent roles globally at automation risk. The IMF put global exposure near 40 percent of employment, and advanced economies near 60 percent, while noting that exposure is not replacement. The WEF projects net job growth alongside enormous skill reconfiguration. The OECD finds high exposure and low realized substitution. McKinsey lands mildest and most optimistic. Same technology, same decade, a three-hundred-million-to-eleven-million spread.
When a field's headline estimates span two orders of magnitude, the estimates are not describing the future. They are describing the modelers. I have seen this exact dispersion pattern in crypto price targets, and it always resolves the same way: the direction is knowable, the magnitude is not, and positioning should be built for the range rather than the point.
The honest uncertainty is the capability boundary. Every parameter in the McKinsey frame — the automation scores, the adoption curve, the eleven million — assumes a model class. If multimodal and agentic systems move the boundary faster than the labor market can retrain behind it, the numbers stop being a 2030 forecast and become a 2027 estimate. The direction holds. The magnitude does not, and the policy window compresses by three years. That is the scenario the brief cannot price and the one that matters most.
One structural bias must be named, not as an accusation but as a portfolio input. McKinsey's client base is the group that captures the upside of AI adoption. A report that concludes "net positive, manageable, needs training" lowers the perceived risk of adoption for exactly that client base. It is also a first-class demand-generation asset for AI transformation consulting. I am claiming frame selection — net instead of gross, opportunity instead of transition cost, adoption instead of capability. Volatility is the tax on unproven consensus, and this consensus is unproven in a specific, measurable way.
Contrarian
The consensus reading is now fixed: net positive, so macro is fine, so risk assets are fine, so the AI trade runs. I think that reading inverts the causality.
Eleven million people will not, in practice, be retrained by 2030. Not at scale, not with a ten-billion-dollar apparatus facing a hundred-billion-dollar liability. Systems rarely clear their liabilities through the intended channel. They route around them.
The route here is agentic substitution. A firm facing a retraining decision and a deployment decision will price both. Retraining takes eighteen months and produces a human with uncertain retention. Deployment takes a procurement cycle and produces an always-on process that does not negotiate. The firm does not need the worker to become an AI operator. It needs the process to become an agent.
Which moves the binding constraint away from model capability and onto settlement infrastructure. In March 2026 I analyzed a leading AI-crypto asset-management protocol and traced a twelve percent loss in simulated user funds to oracle unreliability — a feed latency problem, not a model problem. The agents were competent. The trust layer was not. That is where labor substitution actually stalls: not at whether a model can draft the contract, but at whether an autonomous process can hold, move, and settle value with verifiable execution. Trusted execution environments, not transformers, are the gating technology.
When that gate opens — and I expect it to open before 2030 — the substitution stops being task-level and becomes job-level. At that point the eleven million stops being a stock of displaced workers and becomes a flow of displaced tasks, priced in milliseconds. The marginal unit of labor in the affected roles becomes an agent wallet. Labor's share of that value chain gets repriced as compute, and the reallocation happens without a single retraining program being funded. The transition cost does not disappear. It moves onto household and public balance sheets, invisible, unmodeled, and politically unbudgeted.
Takeaway
So the useful question is not whether AI creates more jobs than it cuts. It is who holds the liability between the cut and the creation, and what that holder does when the retraining bill arrives. If the answer is the household, watch consumption and duration. If it is the state, watch the term premium and the fiscal impulse. If it is the firm, watch nothing — firms already price it, which is why they are deploying agents rather than funding academies. Position for the range the forecasters cannot narrow. But ask the one question McKinsey's frame cannot answer: when the retraining bill comes due and no one signed for it, which balance sheet clears it first?
