The Plans Are Already Made

The headline finding of the Atlanta Federal Reserve's new working paper on AI and the labor market is the one that will be widely cited: AI is not causing aggregate unemployment. "We find little evidence of near-term aggregate employment declines due to AI," the paper concludes.

That finding is real and worth stating. It is also the least interesting thing in the paper.

The working paper — "Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives," released March 25, 2026 — surveyed approximately 750 corporate executives across sectors. It documents substantial heterogeneity in AI adoption, significant productivity gains concentrated in high-skill services and finance, and a productivity paradox: executives perceive productivity gains from AI that are systematically larger than what measured output data shows. The paper attributes the gap to revenue delay — the capability has been deployed but the economic output hasn't yet materialized in the numbers.

These are all useful findings. The one that matters most for the labor market is buried in the distributional detail.

The aggregate finding — little near-term employment decline — is produced by averaging across firms of different sizes. When the Atlanta Fed breaks it out, the picture changes.

Smaller firms expect modest employment gains from AI adoption. This is consistent with the Jevons mechanism: cheaper AI tools expand the productive capacity of small operations, generating demand for more workers to manage the expanded output. Larger companies anticipate AI-driven workforce reductions. This is the Leontief mechanism at scale: large firms with established processes are using AI to substitute for specific human functions, and they are planning the workforce implications accordingly.

The paper documents this as "compositional reallocation of labor both within and across firms" — routine clerical roles declining, relative demand for skilled technical roles increasing. In plain language: the jobs being eliminated are the function-type occupations, and the jobs being sustained or created are the resource-type ones. The bundle is coming apart.

Why the timing matters

The Atlanta Fed finding that larger companies are planning AI-driven reductions while aggregate employment remains stable is not a contradiction. It is a description of a lag structure.

Corporate workforce decisions precede labor market data. A firm that decides in March 2026 to reduce its clerical headcount through attrition and hiring freeze will not produce a visible unemployment spike. The positions won't be backfilled. When the current occupants retire or leave — over a period of years — the occupation's total employment count will gradually decline. That decline will appear in the BLS data eventually. It will appear as a trend. The decision that produced it was made now.

The Atlanta Fed paper's most important finding is therefore this: the plans are in place. The executives have made their workforce decisions based on AI's current and expected productivity. The lag between the decision and the data is not evidence that the displacement won't happen; it is a description of how displacement happens when it happens through attrition and hiring suppression rather than mass layoffs.

This is the compression mechanism documented in a primary-sourced survey of the executives who are doing the compressing.

The productivity paradox and what it implies

The paper's documentation of a productivity paradox — perceived gains larger than measured gains, attributed to revenue delay — has a mirror image in the labor market. If productivity gains are being realized before they appear in output data, labor displacement is also being planned before it appears in employment data. The executives know both sides of this equation. The public data shows neither.

The policy apparatus responds to the public data. Workforce development programs, reskilling initiatives, and the legislative recommendations in the White House's March 20 framework are calibrated to the current employment picture, which shows near-term stability. They are not calibrated to the plans that the Atlanta Fed's executives have already made.

The Tufts Fletcher AI Jobs Risk Index projects that 4.9 million workers in tipping-point occupations could see displacement risk jump from under 10% to over 40% within two to five years. The Atlanta Fed paper provides a mechanism for why that projection might be right even though the current data doesn't show it: the decisions that will produce that outcome are being made now, inside firms, by executives who told a Federal Reserve researcher about them but haven't told their workers yet.

*Sources: Atlanta Federal Reserve Working Paper 2026-4, "Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives," March 25, 2026; Tufts Fletcher School / Digital Planet, American AI Jobs Risk Index, March 2026; Bureau of Labor Statistics Employment Projections 2024–2034.*