The United States now spends roughly $20 billion a year in federal workforce-development funding across the Workforce Innovation and Opportunity Act, Trade Adjustment Assistance, Pell-eligible short-term programs, Job Corps, and a long tail of sectoral grants — and almost nobody can tell you which dollars worked. That is not a complaint about accounting. It is a measurement crisis. The dominant evaluation logic still rewards "people trained" and "credentials issued," which are inputs, not outcomes. The dollars that demonstrably move wages — apprenticeships, sectoral partnerships, career pathways with employer co-investment — are a small share of the spend and an even smaller share of the political conversation. The ROI question, properly asked, is not whether workforce development "works." It is which two or three program designs work well enough to justify scaling and which deserve to be defunded.
This is the argument: stop evaluating workforce investments on enrollment metrics, and start evaluating them on three-year earnings differentials measured against matched non-participants, with employer retention as a secondary check. Everything else is theatre.
The evidence base is thinner than the spending suggests
The most rigorous evidence we have on U.S. workforce programs comes from a small set of randomized and quasi-experimental studies. Mathematica's 2017 evaluation of WIOA Adult and Dislocated Worker programs, building on the earlier ETA-commissioned impact study, found average earnings gains in the range of $500–$800 per quarter for adult-program participants relative to controls — statistically significant, but a modest return on a program that costs several thousand dollars per enrollee. The Year Up program, evaluated by MDRC through a randomized controlled trial published in 2019, produced earnings gains of roughly 30% (about $7,000 a year) three years after program entry — among the largest credible effects in the literature. Project Quest in San Antonio, a sectoral training program studied by Economic Mobility Corporation over nine years, generated annual earnings impacts of about $5,000 for participants well into year nine of follow-up.
Those are the success stories. They share three structural features: long training durations (often a year or more), tight employer alignment in a specific sector, and wraparound supports (childcare, transit, emergency cash). They are not what most of the $20 billion buys. The Department of Labor's own performance reports show median program completers earning under $20,000 in their first post-exit year — a number that almost certainly does not clear a benefit-cost threshold once opportunity cost is included.
Why ROI measurement keeps failing
There are four failure modes that recur in nearly every state and federal workforce evaluation I have seen, and they explain why the spending data and the outcomes data refuse to converge.
The denominator problem
Programs report on completers, not enrollees. A program with a 40% completion rate that posts strong wages for the 40% who finish is not the same thing as a program with a 40% completion rate that posts strong wages overall. The Census Bureau's Longitudinal Employer-Household Dynamics data, properly used, can correct this — but most state workforce agencies still publish completer-only outcomes.
The counterfactual problem
A trainee who enters a program in a tight labor market and earns more a year later may have earned more anyway. Without a matched comparison group — ideally constructed from UI wage records of non-participants with similar pre-program earnings trajectories — the wage gain is uninterpretable. Raj Chetty and co-authors have shown, in their work on Opportunity Insights, how badly intuitions about mobility break down when counterfactuals are missing; the same logic applies here.
The time-horizon problem
Most evaluations stop at two quarters or one year post-exit. That is the worst window. New hires from training programs often start at lower wages than their long-run trajectory, because employers are paying for general human capital and adjusting upward as productivity becomes legible. The Project Quest finding — that effects grew over nine years — is the rule, not the exception. If we evaluate at one year, we systematically understate the better programs and overstate the worse ones, because cheap programs tend to plateau early.
The attribution problem
When private employers co-invest in workforce programs — registered apprenticeships, sectoral partnerships, employer-led upskilling — the public ROI calculation has to net out what the employer would have spent anyway. The OECD's Getting Skills Right series has been pushing OECD members to use deadweight-adjusted ROI estimates for adult learning subsidies, and most U.S. evaluations still don't.
What a credible scorecard would actually measure
If I were designing the federal workforce dashboard from scratch — and the bones of one exist in the WIOA Performance Accountability System, which Congress could strengthen tomorrow — I would force every funded program to publish four numbers and only four numbers.
First, three-year median earnings of enrollees (not completers) drawn from linked UI wage records, with a matched non-participant comparison group constructed from administrative data. Second, employer-retention rate at 18 months post-placement, because a job that lasts three months and ends is not a workforce outcome. Third, cost per quality-adjusted job-year, where "quality-adjusted" means above the local self-sufficiency wage. Fourth, the deadweight share — the estimated fraction of placements that would have happened in the absence of the program, derived from the comparison group.
That is a hard ask. It is also exactly what the Treasury Department's Office of Economic Policy did, in narrower form, when it evaluated the Investing in America agenda's training provisions in 2024, and it is what the Brookings Workforce of the Future initiative has been pushing for since 2019. The data exists. The political will to publish it does not, because the rankings would be brutal: a small number of sectoral and apprenticeship programs would dominate, and a much larger number of generic short-term training programs would be revealed as roughly cost-neutral or worse.
The reallocation case
The bipartisan policy implication is uncomfortable but straightforward. Roughly 80% of new workforce-program dollars over the past three years — through the Infrastructure Investment and Jobs Act, the CHIPS and Science Act, and IRA-linked clean-energy programs — should be funneled into the program designs the evidence already supports: registered apprenticeship, sectoral partnerships modeled on Project Quest, and career-pathway programs with employer co-investment and Pell-eligible short-term credentials in genuinely demanded fields. Generic individual training accounts, "any-credential-is-fine" voucher models, and short-term certifications uncoupled from employer demand should be sunset or held flat. That is not austerity. It is concentration.
The U.S. spends a smaller share of GDP on active labor-market policies than nearly every other advanced economy — the OECD's most recent comparable figures put U.S. spending at roughly 0.1% of GDP versus an OECD median above 0.5%. The case for spending more is strong. But the case for spending more on the same diffuse mix is weak. ROI discipline is what makes a larger budget politically defensible.
For a longer treatment of which training models actually move wages in the AI era, see our flagship Reskilling for Real →.
The dominant evaluation logic still rewards "people trained," which is an input. Pay only for three-year earnings outcomes against a matched comparison group, and the workforce-development map redraws itself within a single funding cycle.
Two questions decide whether the next decade of workforce investment looks more like Project Quest or more like the lost decade of generic short-term training. Will Congress require linked UI-wage outcomes data as a condition of federal grants, the way WIOA already nominally does but rarely enforces? And will state workforce boards publish completer- and enrollee-level earnings against matched comparisons in public dashboards? The answers will determine whether $20 billion a year buys a workforce policy or a press-release machine. The data infrastructure is built. What is missing is the political appetite to read it.
Updated May 21, 2026. This piece was substantively rewritten as part of NWLB's 2026 editorial refresh.
