The $380 Billion Retraining Industry Has 90 Years of Evidence. Almost None of It Works.

In January, Salesforce CEO Marc Benioff told investors the company would hire no more software engineers in 2026. "Agentforce," he said — the company's AI agent platform — had made it unnecessary. Salesforce offered affected employees access to Trailhead, its free online learning platform. Learn AI skills. Adapt. Reskill.

Trailhead has existed since 2014. Salesforce's own internal data, reported at Dreamforce 2025, shows that fewer than 12% of employees who start a Trailhead learning path finish it. Of those who finish, fewer than half report using the new skill in their current role.

This is the retraining economy in miniature. The company fires you. The company offers you a course. Almost nobody finishes the course. The company counts the offer as workforce investment. Everyone moves on.

The 2.6-Point Miracle

The Inter-American Development Bank published the most comprehensive meta-analysis of worker retraining programs ever conducted in 2024. Forty programs. Rigorous causal identification — not before-and-after surveys, but randomized controlled trials and quasi-experimental designs that actually isolate the training effect from selection bias.

The result: program participants were 2.6 percentage points more likely to find a job after training. Wages rose by 0.08 standard deviations — roughly $1,200 a year for a median worker.

That's it. Ninety years of federal workforce programs — from the New Deal's Civilian Conservation Corps to CETA to JTPA to WIA to WIOA — and the best causal evidence says retraining moves the needle by less than three points.

Elisabeth Jacobs at Brookings put it more bluntly in her 2026 review: "The evidence base for large-scale workforce retraining as an effective response to technological displacement is, at best, weak." She traced the entire history — nine decades of programs, thousands of evaluations, billions of dollars — and found the same pattern repeating. Politicians announce retraining. Money flows to providers. Providers report enrollment numbers. Nobody tracks whether enrollees actually get jobs that pay what they lost.

Why Nobody Tracks Outcomes

Because tracking outcomes would kill the business.

The corporate training market hit $380 billion globally in 2024, according to The Business Research Company's market report. It's projected to reach $490 billion by 2028. LinkedIn Learning, Coursera, Udemy, Pluralsight, and their enterprise competitors have built a combined revenue base larger than the GDP of Denmark.

These companies report completion rates, not employment outcomes. They report "skills acquired," not "skills used." They report satisfaction scores — did you enjoy the course? — not income changes. And they have powerful reasons to keep it that way.

Harvard economists Ni, Oberman, and Zhang found something worse. In a 2024 NBER working paper, they tracked workers who specifically retrained for jobs in AI-exposed occupations — the exact move every LinkedIn thought leader recommends. Those workers faced a 29% earnings penalty compared to peers who retrained into non-AI-exposed fields. The people who followed the advice were punished for following it.

The mechanism is straightforward. When everyone retrains into "AI skills," the supply of AI-adjacent workers spikes. Employers, who were already automating the jobs these workers trained for, face a buyer's market. Wages compress. The training industry profits from the enrollment. The worker absorbs the loss.

The Nash Equilibrium Nobody Will Name

The retraining narrative persists because it serves every powerful actor simultaneously.

Corporations need it because the alternative is regulation. If displacement is a training problem, it's the worker's problem. If displacement is a structural problem, it's the corporation's problem. Every Fortune 500 AI announcement includes a line about "investing in our workforce" — typically a Coursera license that costs less than one engineer's monthly salary. Salesforce spent $4.2 billion on Agentforce development and offered Trailhead.

Politicians need it because it's bipartisan and actionable. Both parties can support "skills training." Neither party has to confront the possibility that millions of jobs might simply be gone. The Workforce Pell proposal — expanding Pell Grants to cover 8-15 week certificates — passed a House committee in March 2026 with unanimous support. The same Congress eliminated the Workforce Innovation and Opportunity Act entirely. They defunded the old retraining program and funded a new one in the same legislative session.

Training providers need it because it's $380 billion. Coursera's stock rose 23% in Q4 2025 on "AI upskilling demand." LinkedIn Learning added 4,200 AI-related courses in 2025 alone. The industry's growth directly correlates with displacement anxiety, not with evidence of effectiveness.

Workers need it because the alternative is despair. If "learn to code" or "learn AI prompt engineering" is available, there's agency. There's a plan. The plan has a completion certificate and a progress bar. The alternative — that the job isn't coming back regardless of what you learn — offers nothing to do on a Tuesday morning.

This is a Nash equilibrium. No single actor benefits from defecting. Corporations don't want regulation. Politicians don't want hard conversations. Training providers don't want outcome-based funding. Workers don't want hopelessness. The system is stable. The system doesn't work.

What 7% Looks Like

Brookings researchers surveyed HR leaders at companies that had deployed AI tools in 2025. They found that 7% were working on reskilling strategies for AI-impacted roles. Seven percent. Meanwhile, 62% were focused on "piloting AI in HR management" — using AI to automate the human resources department itself.

The Georgetown CSET report on AI and workforce training documented the gap from the other side: federal workforce development spending has fallen from 0.35% of GDP in the 1980s to under 0.1% today. In absolute terms, the U.S. spends roughly $5 billion per year on public workforce programs — less than what Americans spend annually on Halloween costumes.

The private sector isn't filling the gap. It's exploiting it. When Gartner says "skills-based hiring" is the future, what they're describing is a system where the cost of skill acquisition shifts from employer to worker, the risk of skill obsolescence shifts from employer to worker, and the credential that proves the skill is sold by a company that profits regardless of whether the skill retains value.

The Temporal Problem Nobody Solves

Even if retraining worked — even if the 2.6 percentage points were 26 — there's a timing problem that no program design can fix.

The Anthropic Economic Index, released in March 2026, showed enterprise AI usage crossing from majority-augmentation (57%) to majority-automation (77%) in under a year. The "copilot" era — AI helping workers do their jobs — lasted roughly 18 months before companies started using the same tools to replace workers entirely.

A typical retraining program takes 6-12 months. A coding bootcamp takes 12-16 weeks. A community college certificate takes one to two years. By the time a displaced customer service agent completes a data analytics certificate, the data analytics job she trained for may itself be automated. Gartner estimates AI-relevant skills have a half-life of 2.5 years and shrinking.

You cannot retrain your way out of a disruption that moves faster than training.

Who Benefits From Honesty

The retraining narrative isn't harmless. It actively blocks structural responses.

Every dollar spent on "AI upskilling" is a dollar not spent on income support during transition. Every political minute spent on Workforce Pell is a minute not spent on the AI Prosperity Dividend or expanding H.R. 5830 — the Guaranteed Income Pilot Program Act that would actually test direct income support for 20,000 Americans over three years. Every corporate announcement about "investing in our workforce" substitutes for transparency about how many jobs are actually being eliminated.

The most honest thing the retraining industry could do is publish outcome data: what percentage of completers found jobs paying at least 80% of their pre-displacement income within 12 months? The IDB meta-analysis suggests that number is in the single digits. But nobody publishes it, because the answer would vaporize $380 billion in annual revenue.

The most honest thing a politician could say is: "Some of these jobs aren't coming back, and training can help at the margins but isn't the primary answer." Senator Mark Warner came close with the AI-Related Job Impacts Clarity Act, which would require quarterly reporting on actual AI displacement numbers. It's been in committee since November.

The most honest thing a displaced worker can hear is: retraining might help — it shifts your odds by about 3 points — but the primary response to mass displacement is income support, community preservation, and honest acknowledgment that we don't yet know how to manage this transition. That's a harder sell than "learn Python." It has the disadvantage of being true.