Between 1980 and 2016, US firms threw real capital at automating the workers earning the highest rents — coordinators, reviewers, mid-level analysts whose pay sat above their marginal product — and lost 60 to 90 percent of the productivity gains that automation would otherwise have delivered. That is the central number in Daron Acemoglu and Pascual Restrepo's Quarterly Journal of Economics paper, Automation and Rent Dissipation: Implications for Wages, Inequality, and Productivity, published in the May 2026 issue (vol. 141, iss. 2, p. 1521) (Acemoglu & Restrepo, QJE, 2026). The same paper attributes 52 percent of the rise in US income inequality over the period to this misallocation, with about 10 percentage points coming specifically from wage-premium replacement (Acemoglu & Restrepo, NBER w32536, 2024).
The four-decade history is finally a number. The 2026 question for a Head of Operations at a 200-FTE company is whether the agentic-AI roadmap on the desk this quarter is the same mistake at higher resolution — and the evidence says, by default, it is.
The Acemoglu–Restrepo Mechanism: Rent Dissipation, Not Displacement
The headline finding most readers know from the earlier Acemoglu literature is that automation displaced routine workers and compressed wages at the bottom. The 2026 QJE paper is a sharper claim and worth reading on its own terms. The authors extend their 2022 task-displacement model to incorporate worker rents — the gap between what a worker is paid and the marginal product of their labor — and show that when firms automate, they preferentially target the tasks where rents are highest, not the tasks where the productivity ceiling is highest (Acemoglu & Restrepo, NBER w32536, 2024).
The mechanism: a coordinator who earns 25 percent above their marginal product looks like a 25 percent saving on the spreadsheet the moment the automation pitch lands. A frontline worker performing high-leverage work with a tighter pay-to-output ratio looks like a smaller saving. Capital flows to the larger saving. The productivity gain — the difference between what the automation can do and what was being done — is smaller in the first case, because rent-extracting roles are, by definition, the ones where pay overstates output. The net result is a deployment pattern that maximizes visible payroll relief while minimizing real productivity uplift.
Acemoglu and Restrepo formalize this as rent dissipation: capital is spent to retire pay that did not need to be reduced for productivity to rise, while leaving on the table the much larger gains from automating the work where output actually moves. Quantifying across 49 industries and 500 demographic groups using BEA, ONET, and Census data from 1980 to 2016, they find that two-thirds to nine-tenths of the productivity dividend was lost to this dynamic (Washington State University working paper, 2024). On the inequality side, the same misallocation explains the bulk of the wage-structure shift: "wage decreases for workers specializing in routine tasks with high exposure to automation account for 50 to 70 percent of the changes in the US wage structure between 1980 and 2016" (WorkRise, 2021) — a figure the QJE paper now decomposes into rent vs. productivity components.
The conclusion the authors put in the paper, not in a press release: when you target automation at people instead of at output, you can capture inequality without capturing productivity. That is the empirical pattern of US automation from 1980 to 2016.
Why the 2026 Agentic-AI Rollout Repeats the Pattern
The natural reaction in an ops review is that 1980–2016 industrial automation is not 2026 agentic AI — different technology, different unit economics, different timeline. The unit economics are different. The targeting logic is not.
Walk into any mid-market agent-deployment review this quarter and the ROI math is almost always presented the same way: role X costs $Y per year; the agent can do 60 percent of role X's work; therefore the agent saves 0.6 × Y. The roles named in those slides are not chosen by where AI's marginal productivity uplift is largest. They are chosen by where the payroll line is largest and where the work is structured enough for an agent to look credible — which biases the targeting toward coordinators, reviewers, senior analysts, and customer-success leads. Those are precisely the wage-premium roles in Acemoglu and Restrepo's framework: roles where pay exceeds marginal product because of within-firm rents (information asymmetry, hard-to-measure judgment, internal bargaining power).
The agent is then evaluated against the saving on that payroll line, not against the productivity ceiling the deployment could have hit if aimed elsewhere. The two questions — what does this agent save? and where would this agent produce the most output? — are not the same question, and almost no mid-market deployment review separates them.
The MIT Initiative on the Digital Economy has been making a version of this argument for two years: that the productivity dividend from AI is concentrated in tasks where the human's current output is bottlenecked by cognitive throughput, not in tasks where the human's current pay is high (MIT IDE, 2024). The two distributions overlap, but they are not the same distribution. The Acemoglu–Restrepo paper is the first piece of historical evidence with the magnitude of the gap measured — and the gap is large.
The 60–90% Number, Read Two Ways
The 60–90 percent productivity tax has two operational readings, and a 50–500-FTE ops function should hold both.
The conservative reading is that the wage-premium-targeting penalty applied to industrial automation may not transfer one-for-one to agentic AI, because the marginal cost of deploying an agent on a different task — once the agent is built — is much lower than the marginal cost of redeploying industrial machinery. In principle, an agent that is targeted poorly can be re-targeted in a sprint, where a stamping press that is misallocated is a multi-year capital write-off. This is the steel-man for current deployment practice: the cost of getting the target wrong is recoverable.
The aggressive reading is that the wage-premium-targeting penalty is worse in the agentic case, not better, because the organizational politics around removing the targeting bias are harder. When the C-suite has been pitched the deployment as a payroll saving and the saving has been booked into next year's plan, redirecting the agent to a different (lower-payroll, higher-productivity) function is no longer a sprint decision — it requires reversing a financial commitment, defending the original framing, and explaining why a previously named role is no longer the target. The political cost of re-targeting is what makes the 1980–2016 pattern persist for forty years instead of being corrected in year two.
Both readings converge on the same operational implication: the targeting decision in quarter one is far more load-bearing than the technical capability of the agent. A correctly targeted weak agent outperforms a mis-targeted strong agent, because the strong agent's gains are dissipated against pay that did not need to be reduced.
The Counter-Argument: Wage-Premium Roles Are Where Judgment Lives
The strongest pushback from a Head of Operations is that the wage-premium roles are exactly where the most leveraged work happens — that the reason coordinators and senior analysts are paid above their marginal product is that they hold the institutional context that lets a 200-FTE company actually function. Pointing an agent at those roles is not rent dissipation; it is the highest-leverage automation target by definition.
The counter-argument is partly right and entirely consistent with the Acemoglu–Restrepo finding. The reason the wage premium exists in those roles is precisely the judgment workload — and judgment workload is also where most current agentic systems still fail in ways the deployment pitch does not flag. Recent randomized evidence on AI handling judgment-heavy tasks shows that confidence in agent output is uncorrelated with accuracy, particularly when the human user is no longer the domain expert (Bojinov et al., HBS working paper, 2024). So the targeting that looks highest-leverage in the deployment slide is also the targeting most likely to produce silent quality regressions that do not show up in the productivity dashboard until two quarters later.
What the QJE paper adds to this debate is the four-decade base rate: when firms target the wage premium, the productivity gain shrinks. The presence of judgment workload in those roles is why the targeting is tempting — but the same judgment workload is why the marginal productivity uplift is smaller than the payroll saving suggests. The right framing is not "wage-premium roles are bad targets" but "the saving on a wage-premium role is not a productivity number, and treating it as one is the four-decade mistake."
The Mid-Market Specifics: What Changes for a 200-FTE Ops Function This Quarter
For a Head of Operations finalizing 2026 agent targets, the QJE paper converts to three concrete deployment-review changes. None require a different vendor or a different agent.
One: separate the payroll-saving line from the productivity-uplift line. Every agent deployment proposal should score two distinct numbers: the gross payroll relief the deployment makes possible, and the measured productivity uplift (output per unit of judgment-time) the deployment is expected to produce. The two numbers are not interchangeable. When they diverge by more than 2x, the deployment is in rent-dissipation territory — the agent is being justified by the saving, not by the work. That is the moment to ask whether a different target produces the same productivity uplift without the wage-premium dependency.
Two: score targets against where output is bottlenecked, not where payroll is concentrated. A weekly close that takes four days because reconciliation queries take six hours per cycle is a productivity bottleneck. A senior analyst who earns $180K is a payroll concentration. The first is a high-uplift agent target; the second is not, even though the second produces a larger headline saving. Mid-market ops functions almost never run this scoring exercise explicitly; the deployment review defaults to the payroll view because the payroll view is the one the CFO can compute in a meeting.
Three: pre-commit to the re-targeting protocol. Acemoglu and Restrepo's finding is, in part, about persistence — the misallocation lasted four decades because no one corrected it. The agentic-AI equivalent persists because the original deployment pitch hardens into a financial commitment within a quarter. The hedge is to write the re-targeting trigger into the deployment proposal itself: at month three and month six, the agent's productivity uplift is measured against the original target; if the measured uplift is below 40 percent of the projected uplift, the agent is re-pointed to a different task before the political cost of re-targeting becomes prohibitive. This is the only structural defense against the persistence Acemoglu and Restrepo measure.
These three moves are not technical; they are review-process moves. They do not require buying a different agent or hiring a different team. They require running the deployment review in a different shape — one that does not collapse productivity and payroll into the same column.
This Quarter's Specific Move
The Acemoglu–Restrepo QJE paper is the first piece of empirical work in forty years to put a number on the cost of automating against pay rather than against output. The number — 60 to 90 percent of the productivity dividend lost, 52 percent of the inequality rise attributable to the same dynamic — is large enough to flip the ROI math on most current mid-market agent deployments if the math is done correctly.
The decision in front of a Head of Operations this quarter is narrow. Before signing off on the next agent target, run the deployment proposal through one filter: is this target chosen because the work is where productivity is bottlenecked, or because the payroll is where the saving is most visible? If the honest answer is the second one, the historical base rate says the deployment will dissipate two-thirds to nine-tenths of the productivity gain it could have captured.
Rescore the target. Separate the columns. Write the re-targeting trigger. The cost of doing this in quarter one is a meeting and a revised template. The cost of not doing it is the one Acemoglu and Restrepo have now put a hard number on — and the one your 2027 productivity review will be written about.
