Retraining works when it targets one industry with local demand: in randomised trials of US sectoral programmes, earnings rose 11 to 40% after training. Generic training has a weaker record. In Germany, a two-year retraining raises women's employment by up to 28 percentage points within four years. Romania offers free courses, and its adults train less than most Europeans.
This part of my report on the future of work rests on the strongest kind of evidence in it, because it is the one place where the evidence is experimental. Most of what follows comes from randomised trials, which compare people offered a programme with similar people who were not. The American and German evidence is labelled as such; they are different systems and should not be read as one.
Does retraining actually work?
Some of it does, and the best evidence comes from the US. A review of randomised trials of several sectoral programmes, which train people for jobs in one industry where local employers are hiring, found earnings gains of 11 to 40% after training ended (NBER). The effects last. Project QUEST in San Antonio was followed for eleven years, and its participants earned $31,395 more than the control group over the whole period (Economic Mobility Corporation). At Per Scholas, an IT training provider in the WorkAdvance trial, earnings were still 14% higher in the seventh year (MDRC).
The broad programmes look different. In a national trial of the US federal workforce system, job seekers offered one-to-one help from staff earned about $7,100, or 20%, more over 30 months than those without it. Evidence that the training the programme paid for raised employment or earnings was inconclusive (Mathematica). For workers displaced by trade, the federal Trade Adjustment Assistance programme left participants earning either $3,300 less than a comparison group in the fourth year, or about the same (US Department of Labor). The programme has since expired (North Carolina Department of Commerce).
Why does sectoral training work when generic courses often do not?
Put the two results side by side and a pattern appears. The federal trial measured two things separately: a person who helps you search, and a course. The person showed up in the earnings data. The course did not. The sectoral programmes, which pay off best, start from the other end: an industry that is hiring locally, and a course built for its jobs.
We tend to think of retraining as acquiring a skill. The trials suggest it works more like a route: a skill plus a specific employer who needs it, plus someone who helps you walk from one to the other. Take away the employer and the course becomes a certificate. Take away the help and many people never reach the employer.
How long does it take for retraining to pay off?
Longer than people budget for, and the first year often goes backwards. Year Up participants earned $5,338 less than the control group in the year after they enrolled, then $5,181 more in the second year and $7,011 more in the third (Policy Impacts Library, summarising the trial). Time in a course is time not spent earning, and the trials count that.
Germany shows the same shape over a longer horizon. A study of publicly funded training before the 2003 Hartz reform found a negative effect that disappeared after about 12 months, with positive effects of about 5% after about 22 months (IZA). Today the Federal Employment Agency runs two kinds of programme. Short further-training courses (FbW): people who finished them were 12 percentage points more likely to be in unsubsidised work 180 days later than comparable people, of whom 45% were working (Bundesagentur für Arbeit). And the full retraining into a new trade, the Umschulung, which usually takes two years (Bundesagentur für Arbeit). Within four years of starting one, its effect on employment reaches up to 28 percentage points for women and almost 18 for men (IAB-Forum).
| Programme | Where | Measured effect |
|---|---|---|
| Sectoral training (several programmes) | US, randomised trials | Earnings +11 to 40% after training |
| Project QUEST | US, randomised trial | +$31,395 over eleven years |
| Per Scholas (WorkAdvance) | US, randomised trial | Earnings +14% in year 7 |
| One-to-one staff help | US, randomised trial | +$7,100 (20%) over 30 months |
| Short further training (FbW) | Germany, matched comparison | +12 points in work at 180 days |
| Umschulung (about two years) | Germany | Up to +28 points (women), +18 (men) within four years |
What retraining is available in Romania?
The national employment agency, ANOFM, organises free qualification and requalification courses for the unemployed (ANOFM). At its regional training centre in Dolj, a level-3 qualification course runs 720 hours, and welding is among them (ANOFM).
In 2024, 18,859 people went through ANOFM training or skills assessment, 16,276 of them unemployed. Of the unemployed it trained, 44.0% were over 45 and only 12.1% under 25. By the time of the report, 3,315 unemployed graduates of the 2024 courses had been placed in jobs (ANOFM annual report). That last figure is not a success rate: the window for 2024 graduates had not closed, and there is no comparison group. Romania has no trial evidence of the kind the US and Germany have.
The wider picture is low participation, with a twist. In the 2022 Adult Education Survey, 25.6% of Romanians aged 25 to 64 had taken part in education or training in the previous twelve months, against 46.6% across the EU and 60.4% in Germany (Eurostat). On the narrower four-week measure in 2024, Romania was at 8.9%, close to Germany's 9.3% and below the EU's 13.5% (Eurostat). The two surveys measure different things, so read the gap as large, not as an exact size.
Does AI change what is worth retraining for?
None of the trials above was designed around AI, and some are a decade old. What they test is the mechanism, and the mechanism does not depend on the technology: a course pays when it leads to a specific employer who is hiring. AI changes which employers those are. It does not change the rule.
Where AI does enter is the choice of trade. Desk work exposed to AI is where hiring of beginners has slowed first, as I described in AI and entry-level jobs, and the jobs projected to grow fastest are mostly physical and in person, from care to the trades. The pay and the length of each route are in which jobs grow in Europe to 2035.
What should an employer take from the evidence?
Two things, both drawn from the trials rather than from intuition. Training tied to a role you are actually hiring for behaves like the sectoral programmes; training offered in general behaves more like the inconclusive federal courses. And the one-to-one help mattered on its own, which inside a company means a person who checks how the new skill is being used on real work, not only a course catalogue.
For AI tools specifically, that means training a team on its own processes and its own documents, with someone looking at the output afterwards. It is slower to organise than a generic team training session. The trials are a fair argument that it is also the format that shows up in the results.
We usually ask whether people can learn new skills. The trials answer that they can, at almost any age. The harder question is whether anyone is waiting for them at the end of the course.
If you are planning how your team learns to work with AI: vladtudor.com/workshops.