The number that should haunt every general contractor in America is not the 439,000-worker shortage that Associated Builders and Contractors projected for 2025, the kind of figure that belongs to economists and lobbyists who trade in workforce projections without ever stepping onto a job site. The number that should keep you up is the one sitting in the Department of Labor’s own apprenticeship database, confirmed by a University of Washington analysis of 335,212 construction apprentice records spanning 2012 to 2023: 40.1% of construction apprentices cancel their programs before completing them.
Not 40% of marginal candidates who wandered in off the street and decided construction was harder than they imagined. Forty percent of people who enrolled in a registered apprenticeship program, passed the screening, found a sponsor, and showed up on day one. They did everything right. Then they left.
Half of them were gone within six months.
The Program That Lost Its Only Apprentice
A RAND Corporation study published in April 2025 mapped the structural problem with a precision that the industry has not yet absorbed. There are roughly 7,000 registered construction apprenticeship programs in the United States, and nearly half of them have one or two apprentices in training at any given time.
Program coordinators told RAND researchers that the problem was not a lack of applicants. Some programs only bring on new apprentices once every two years and have to turn away most who apply, because the bottleneck is capacity: mentor availability, equipment, and state laws that limit how many apprentices a single journeyman can supervise at once. So when a program with a single apprentice loses that apprentice in month three, it does not lose 40% of its output. It loses everything, sometimes for years, because there is no one waiting in the queue to take that slot until the next intake cycle.
“We’re talking about an opportunity that’s paid, that can lead to meaningful employment, and that can offer a pathway to the middle class,” said Marwa AlFakhri, the RAND labor economist who led the research. “If we could close that completion rate gap, that could meaningfully change the supply of construction workers in the near future.”
RAND estimated that an apprenticeship credential is worth more than $240,000 in additional lifetime earnings over a 36-year career, which means that the roughly 116,000 who walk away each year represent something close to $28 billion in aggregate forfeited earning power. Not construction industry revenue lost, but lifetime economic potential that individual workers chose not to claim, or could not hold onto long enough to claim.
Where the Money Is Going
On July 4, BlackRock announced Future Builders, a $100 million initiative to train and place 50,000 skilled trades workers over five years, targeting electricians, welders, plumbers, and HVAC technicians with a framing that is equal parts patriotic and pragmatic. “What better way to celebrate America’s 250th than to honor the men and women who actually built the country,” said John Kelly, BlackRock’s Global Head of Corporate Affairs. Larry Fink, BlackRock’s CEO, has estimated that America should plan to spend $10 trillion on data centers and energy infrastructure for AI, and Future Builders is a down payment on the labor force required to pour the concrete and pull the wire.
In April, North America’s Building Trades Unions and Microsoft expanded a nationwide partnership to deliver AI literacy training through union apprenticeship programs and LinkedIn Learning, building on a collaboration that had already trained 1,500 instructors in hands-on training centers across the country. It adds free AI courses, industry-recognized credentials, and integration with TradesFutures, a nonprofit that connects people to union apprenticeship programs in 34 states.
Federal investment hit $244 million for expanding apprenticeship opportunities in fiscal year 2024, the largest such investment in history. All of this money goes to the same place: the top of the funnel, recruiting new apprentices and teaching AI tools to workers who already made it through. None of it addresses why 116,000 people who already started a program decided to stop.
Who Quits, and When
The University of Washington study broke down the cancellation data by trade, demographics, and program type, and the patterns are not random.
Roofing has the worst retention, with 58.5% of apprentices canceling, followed by plastering at 55.6% and drywall at 53.2%. These are the trades that wreck bodies fastest: overhead taping on stilts, hauling shingle bundles up ladders in August heat, mixing and applying plaster with your arms above your shoulders for eight hours straight. None of that physical toll is a surprise to anyone who has spent a week on a roofing crew, but what is surprising is that nobody systematically screens for it or prepares candidates for the reality before their first day.
Nonunion programs lose more apprentices than union programs, 44.6% versus 38.4%, a statistically significant gap that likely reflects the structural advantages of union apprenticeships: more formalized mentorship, clearer wage progression schedules, and benefits packages that create staying power through the hardest early months. Nonunion programs have grown faster, with a 43% increase in participants since 2019 compared to 11% for union programs, which means the industry is channeling most of its growth through the pathway with the weaker retention.
Black apprentices face 41% higher odds of cancellation than white apprentices after controlling for other factors, while American Indian and Alaska Native apprentices face 13% higher odds and women face 11% higher odds. Srikanth and colleagues noted that these disparities likely reflect workplace climate, mentorship access, and systemic barriers that the pipeline itself does not address, which means every intervention that focuses solely on recruitment volume will reproduce the same unequal outcomes at greater scale.
Geography matters too, and the spread is wide. The Middle Atlantic division (New York, New Jersey, Pennsylvania) has the lowest cancellation rate at 28.7%, while the West South Central (Texas, Oklahoma, Arkansas, Louisiana) sits at 47.1%, a gap that encompasses differences in climate, labor culture, regulatory environments, union density, and a dozen other variables that no single intervention will flatten.
AI at the Bottom of the Funnel
A handful of programs are pointing technology at the actual leak, rather than the recruiting brochure.
In 2022, the National Center for Construction Education and Research ran a controlled pilot with ABC Illinois comparing VR-trained apprentices against a traditional-lab control group on two standardized tasks: circular saw operation and electric drill training. Results were lopsided, with the VR group averaging 17.67 on expert evaluation versus 13.46 for the traditional group, and on the standardized NCCER assessment, VR scored 75 against 53.46. But the number that matters most for the dropout crisis is the one at the end of the program: 100% of VR-group students completed their training and secured jobs, while the traditional group completed at 69%.
Small sample, one pilot, not proof of anything that could be generalized to a national policy. But the mechanism makes intuitive sense if you have spent any time on a job site watching a first-week apprentice try to keep up. A first-week apprentice faces a sensory assault: unfamiliar tools, unfamiliar danger, unfamiliar exhaustion, and a mentor who has his own deadlines and cannot slow down to explain how the miter saw sounds different when the blade is dull. VR lets a first-week apprentice make mistakes in a space where mistakes do not draw blood, do not waste material, and do not cost the mentor billable hours, and the confidence that builds in that safe space appears to translate into the kind of staying power that keeps someone on the job site through week six and month three and the first November rain.
ABC Institute, Florida’s largest registered apprenticeship provider, partnered with Uplift Training in late 2024 to build ABC Labs, a VR-powered training platform designed to expand program capacity without requiring more physical lab space or equipment. ABC Institute currently trains over 10% of all apprentices in Florida across seven trades. Their operating theory: if the bottleneck is physical space and mentor time, VR can multiply both without adding square footage or headcount.
In the UK, Aston University built a machine-learning platform called Smart Coach that analyzes patterns in apprentice behavior and flags at-risk learners within three months of enrollment. It does not just predict who will quit. It prescribes specific interventions for each learner. It works on a straightforward premise: if half of dropouts leave in the first six months, and you can identify the ones heading for the door by month three, you have a window to act. Whether the intervention is a different mentor assignment, a schedule adjustment, or a frank conversation about whether roofing is really what this person wants to do for forty years.
What None of This Fixes
VR headsets do not make drywall finishing easier on a 47-year-old shoulder, and predictive analytics do not change the fact that a first-year roofer in Dallas works in 104-degree heat for $18 an hour while an entry-level data center technician down the road works in air conditioning for $24. AI literacy courses from Microsoft, however well-intentioned, do not address why a Black apprentice electrician has 41% higher odds of quitting than a white one enrolled in the same program and working on the same crew.
RAND identified a blunter problem lurking underneath the technology question. State laws in many jurisdictions limit a journeyman to supervising one apprentice at a time, and changing that ratio to two-to-one would double capacity overnight, but the research on whether doubling the ratio can be done safely simply does not exist yet. In construction, an evidence gap and a body count are separated by a very thin margin. Apprenticeship mentors are also working full-time on active jobs, which means every hour of mentorship has an opportunity cost measured in concrete dollars that the mentor’s employer is absorbing in a labor market where every billable hour matters.
The deepest problem may be informational, a mismatch between what candidates expect and what the work actually demands. RAND found that the early timing of most dropouts suggests candidates did not fully understand what they were signing up for. The work is physically demanding in ways that are hard to convey in a recruitment presentation, the hours are irregular, the commute changes with every project, and the pay, while it leads to a $240,000 lifetime premium over non-credentialed work, starts low and stays low for the first two years in a way that is difficult to tolerate when rent is due on the first of the month. A realistic preview of what the first winter on a framing crew actually feels like might be the cheapest and most effective intervention nobody is deploying at scale.
VR could deliver that preview, not as a training tool but as an honest simulation of the job: the cold, the weight, the repetition, the 5:30 AM starts, the feeling of your hands going numb inside wet gloves while you try to hold a nail gun steady on a February morning. The technology to build that experience exists today. But the will to deploy something that might talk candidates out of starting does not exist, because every program needs bodies and nobody wants to build a tool that reduces enrollment, even if doing so would save money and heartbreak by reducing the dropout rate on the other end.
The Math for Builders
If you are a residential builder planning projects that break ground in 2027 or 2028, here is what the apprenticeship data means for your budget.
Registered apprenticeship programs currently produce about 35,000 to 40,000 new construction workers per year, and unregistered programs, community colleges, and immigration add tens of thousands more, but RAND’s estimate is unforgiving: even adding all pathways together, the nation falls roughly 250,000 workers short of estimated need.
That shortage is not evenly distributed, and you will not encounter it as a single number on an industry report. Critically, the trades with the highest dropout rates are the same ones with the most acute shortages, so a roofer shortage drives up roofing subcontractor bids, an HVAC shortage extends mechanical rough-in timelines, and what you feel on your project is a $3,000 surprise on the roofing line item and a two-week delay waiting for ductwork rather than a clean macro statistic.
If AI-powered tools could cut the dropout rate from 40% to 30%, based on FY 2024 participation numbers that would put an additional 29,000 workers into the market annually, equivalent to roughly three years of new program creation at current growth rates. Capital required to build and deploy those tools would be a fraction of what the industry is already spending to recruit replacements for the people it had and lost.
What We Did Not Prove
The Transfr VR pilot was a single controlled study at one ABC chapter. Sample sizes were small, and the tasks measured (circular saw, electric drill) are among the simpler skills in a four-year electrical or plumbing apprenticeship. Whether VR training sustains retention through year two of a pipefitting program is an open question with no published data.
Aston University’s Smart Coach was developed in the UK, where apprenticeship structures, employer incentives, and labor markets differ substantially from the U.S. system. Transferability is plausible but unproven.
That $28 billion figure in forfeited lifetime earnings is a rough aggregate: 116,000 estimated dropouts multiplied by the RAND-cited $240,000 career premium. It conflates different trades, regions, and program types, and it assumes every dropout would have completed and earned the full premium, which is not realistic. Some fraction of those 116,000 left for higher-paying work, military service, or other life events. It illustrates scale, not precision.
No published research has tested whether a “realistic job preview” delivered via VR reduces dropout rates when deployed at enrollment rather than after it. That study should exist. It does not.
Marcus Washington covers construction workforce and labor issues. His reporting on AI’s impact on construction workers draws on BLS data, academic research, and conversations with the people who build things for a living.