The previous piece argued that AI got pointed downward — at the writers, the analysts, the customer service reps — because deployment follows authority, not value. The CEO should have been the first role replaced. The fact that it wasn’t is a confession about who held the lever.
That diagnosis only matters if there’s a prescription. So this is the prescription, and the reason it works is that we already know it works. We have the data. We ran the experiment for a quarter century and we measured the result.
The lockstep
From 1948 to 1973, productivity and median worker compensation grew together. Not approximately. Together. Plot them on the same chart and the two lines are almost indistinguishable, both rising at roughly 95% over the period, almost in unison.
After 1973 they separate.
Productivity keeps climbing. Compensation flattens, then drifts. By 2024, the cumulative gap is roughly 60–70 percentage points… productivity has grown that much more than typical worker compensation since the divergence began. That gap is the surplus that used to flow to workers and now flows somewhere else. It is not a rhetorical claim. It is a measurement, published annually by the Economic Policy Institute, and the chart has been getting more damning every year for fifty years.
If you have never seen this chart, you should look at it before you read another word about AI and labor.1 It is the single most important piece of evidence in American economic life, and most public discussion of automation, productivity, and worker outcomes proceeds as if it doesn’t exist.
The reason it matters here is that the lockstep era is our control group. We know what an economy looks like when productivity gains reach workers. We have twenty-five years of data on it. The question is not theoretical. The question is what changed.
What produced the lockstep
A bundle of constraints, operating as a system. Six are worth naming.
CEO-to-worker pay ratios sat in the 20–30x range, depending on the source and year. Not by accident… by a combination of social norm, tax structure, and the fact that boards hadn’t yet evolved into the comp-committee revolving door they would become.
Defined-benefit pensions were the norm for full-time workers at large firms. The firm bore the longevity risk and the investment risk. Workers accumulated a senior claim on the firm’s future cash flows that vested with tenure and could not be unilaterally rescinded.
Marginal income tax rates on top earners ran above 70% for most of the period… 91% under Eisenhower, 70% under Kennedy and Johnson, dropping to 50% only under Reagan. Whatever the executive class extracted at the top, the federal government recaptured a majority of it and recycled it through public investment.
Union density ran above 30% of the private workforce through most of the period, peaking at roughly 35% in the mid-1950s. Collective bargaining was the mechanism that translated firm-level productivity gains into wage gains, and the threat of organizing kept non-union firms paying competitively.
Glass-Steagall separated commercial and investment banking. The financial sector was a service to the productive economy, not a competitor with it for talent and capital. Finance as a share of GDP sat in the 3–4% range for most of the lockstep era. It is roughly double that now.
The corporate purpose itself was understood differently. The Business Roundtable’s pre-1997 framing still acknowledged the corporation as having obligations to employees, customers, and communities alongside shareholders. The shareholder-primacy doctrine that came to dominate after 1980 was a deliberate ideological project, not a natural state.
Six constraints. Operating together. Reinforcing each other.
The system framing matters because it explains why dismantling the constraints individually didn’t restore equilibrium, and why proposals to restore them individually won’t either.
The constraints worked because they were redundant.
Tax rates capped extraction at the top. Union density forced distribution at the bottom. Defined-benefit pensions locked in deferred ownership. Glass-Steagall kept the financial sector from outbidding the productive sector for human and financial capital. Each one closed a path the surplus might have taken to escape redistribution. Together they formed a system in which productivity gains had nowhere to go except back to the workers who produced them.
Remove enough pipes from a system designed to route water downhill, and the water finds another way. That’s what happened from roughly 1978 to 1999. Marginal rates collapsed under Reagan. Union density was actively crushed. Glass-Steagall was repealed in 1999. Defined-benefit pensions were shed in favor of 401(k)s, transferring longevity and investment risk from firms to workers. Shareholder primacy hardened from one theory among many into the only theory taught in business schools. The 20–30x ratio expanded into the 200–400x range it occupies today.
The lockstep didn’t break. It was disassembled. Piece by piece, with each removal justified on its own terms, until the system that produced the lockstep no longer existed.
Why this matters now
We are about to add a new extraction mechanism, the automation of coordination work, to a system whose redistributive constraints have already been removed.
This is the part of the argument that’s actually unanswerable, and it’s worth stating slowly.
The 1960s configuration could absorb labor-saving technology without producing runaway inequality, because the surplus from the technology had nowhere extractive to flow. Cap the executive comp ratio, tax extraction at 70%, force collective bargaining, and any productivity gain a firm captures has to be split with workers, taxed, or reinvested. The constraints route the surplus.
The 2026 configuration has the opposite property. Every redistributive pipe has been narrowed or closed. Executive comp is uncapped. Top marginal rates are at historic lows. Union density is below 7% in the private sector. Pensions are gone. Shareholder primacy is the default theory of the firm. So when AI lands on the coordination layer — which is the highest-value layer in the firm to automate — the surplus has only one direction to move.
This is not a forecast. It’s already happening.
Look at where the productivity gains from the first three years of generative AI deployment have shown up. Margins. Buybacks. Executive comp. Not wages. Not pensions. Not hours reduced. Not jobs preserved at higher pay. The pattern of the post-1973 divergence is being reproduced in compressed time, on a smaller dataset, with a sharper slope. We are watching the same movie at 4x speed.
The prescription
The constraints have to come back. Not all six at once, that’s not politically achievable in the next decade, but the load-bearing ones, in a sequence designed to re-tension the system before AI completes the extraction.
Two are the entry points.
Cap CEO-to-worker compensation at 20–30x. This is the historical range. It was the norm during the most prosperous and most equally-distributed period in American economic history. It is not a radical proposal. It is a return to a configuration the economy demonstrably operated under, profitably, for a quarter century. The objection that “we’ll lose our best CEOs to other jurisdictions” is the same objection raised against the OECD minimum tax and the same objection that’s been raised against every constraint on extraction for forty years. The empirical case for executive irreplaceability is weak, the cap doesn’t bind below the threshold, and the firms that adopt it first will see internal alignment and morale benefits that no spreadsheet captures until later.
Restore pension obligation through sectoral, multi-employer plans backed by sovereign reinsurance for longevity risk. The American DB model died because it loaded all the risk onto individual firms, which is what made the 401(k) shift competitively rational for any single employer. The Northern European model, the Dutch and Danish sectoral pension funds, solves this by pooling the risk across an entire industry. Workers get the deferred-ownership claim that a DB plan provides. Firms get a competitive level playing field because every firm in the sector contributes on the same terms. The longevity tail risk is socialized through public reinsurance, which is the only entity with a balance sheet long enough to absorb it.
These two together do what the original 1948–1973 system did… they cap extraction at the top and re-establish the worker as a residual claimant on firm productivity at the bottom. The surplus from automating coordination has somewhere to land that isn’t a buyback.
The control group
We already ran the experiment.
We know what the constrained configuration produces: twenty-five years of broadly shared prosperity, the lowest inequality in the modern industrial era, productivity and wages rising together, a middle class that could buy a house on a single income.
We know what the deconstrained configuration produces: fifty years of decoupled productivity and wage stagnation, the highest concentration of wealth since the Gilded Age, a middle class that has been told the problem is its own choices.
The question is not whether the constraints work. The question is whether we’re willing to reimpose them before AI completes the extraction the deregulatory cycle started.
We are running out of time to make that decision under conditions that allow it to be made democratically. Once the coordination layer is fully automated and the gains have flowed to a smaller and smaller set of capital holders, the political coalition required to reimpose the constraints becomes structurally weaker every year. The window in which this is winnable is the window in which workers still have enough collective income share to constitute a political force.
That window is closing. It will not reopen on its own.
The first piece called the misread a confession. This is what the confession demands: that we look at the period when the lockstep held, name the constraints that produced it, and reimpose the load-bearing ones before the machine that’s coming finishes the work the last fifty years began.
We have the data. We have the precedent. We have the working models in Northern Europe right now.
What we don’t yet have is the will.
That’s the only thing left to argue about.
This is the second piece in a three-part sequence.
Piece one, "The CEO Should Have Been First", diagnosed why AI was deployed downward instead of upward.
Piece three, "The Mechanism Has a Name", documents the specific signals visible in the public record right now that suggest the constraints described above will be actively prevented from being reimposed, and names the literature that predicts how.