- AI makes codified knowledge abundant and practical expertise scarcer. Policies must ensure learning opportunities for workers are preserved.
Key ideas
- As artificial intelligence makes codified knowledge abundant, tacit knowledge—the expertise gained only through doing—becomes scarcer and more valuable.
- Latin America and the Caribbean face a specific risk: the entry-level jobs where young workers acquire practical skills are precisely those most exposed to AI automation.
- Governments should invest now in apprenticeships, dual-training, and portable employment records to ensure AI adoption does not hollow out the next generation’s learning opportunities, an approach championed by the IDB in the region.
By Eduardo Levy Yeyati
For a while, the conversation about artificial intelligence and work revolved around a straightforward question: which tasks can machines do, and which will remain human? That question still matters. But the debate has already moved on. Recent work points to a broader concern: AI may shape not only what workers do, but also how human capabilities are formed and transmitted over time— a question with urgency for Latin America and the Caribbean (LAC).
What this shift underscores is that experience and practical judgment now matter more than ever. To ensure AI strengthens human capital in our region, policymakers should expand learning by doing through apprenticeships and dual training, make employment records portable to signal practical skills, and support early career experience through targeted subsidies. This has been an agenda long supported by the Inter-American Development Bank (IDB) to ensure AI adoption does not reduce learning opportunities.
A useful place to start is with a simple yet often overlooked distinction: the difference between explicit and tacit knowledge. Explicit knowledge can be codified, written down, taught in a classroom, stored in a database, or transmitted through manuals and documents. Tacit knowledge is different, it is practical expertise gained through experience, trial and error, and context. It includes the intuition of an experienced nurse, the field sense of a caseworker, the practical judgment of a teacher, or the implementation know-how of a public manager. Much of what matters in skilled work is hard to document and even harder to transfer through instruction alone.
This distinction matters because generative AI seems especially good at scaling explicit knowledge. It can summarize, classify, retrieve, compare, draft, and recombine information at very low cost. But that does not mean it can replicate the practical judgment that comes from experience equally well. If anything, the opposite may be true: the more abundant codified knowledge becomes, the more valuable may become the knowledge that is acquired only through doing.
When doing beats knowing
That possibility runs through much of the new literature, although from different angles. Acemoglu, Kong, and Ozdaglar study how generative—especially agentic—AI may weaken incentives to acquire context-specific knowledge, gradually eroding the knowledge base on which future decisions depend. Ide and Talamàs analyze how AI reshapes the organization of knowledge work, changing the role of workers who solve routine problems and those of experts who handle exceptions. Ide’s more recent paper focuses on a related risk: how automation of entry-level tasks may reduce the opportunities through which tacit expertise is transmitted across generations of workers.
The question I would add is closer to the classic economics of human capital: if AI scales codified knowledge more easily than tacit knowledge, what changes should we expect in how workers are paid and valued?
This brings back an older question in labor economics, associated with Ben-Porath and Mincer: how human capital is built over the life cycle, through schooling and experience, and how wages reflect the return to both. In the age of AI, that framework may acquire a new significance. AI may do more than shift workers across tasks. It may alter the composition of valuable human capital, lowering the relative scarcity of codified knowledge while raising the relative scarcity of judgment acquired through practice.
This is not just a loose conjecture. Economic models of human capital accumulation point to several observable implications. In occupations more exposed to AI, the return to formal education should weaken relative to the return to experience. The point in a worker’s career at which experience becomes more valuable than credentials should arrive earlier for more recent cohorts— a shift that researchers have tracked in earnings data across generations, and that recent work on how schooling and early work experience interact over the life cycle helps explain.
And firms may begin to rely less on titles and degrees alone, and more on documented experience: work histories, portfolios, and demonstrated performance when screening candidates—a scenario in which portable employment records may become increasingly relevant as a new kind of credential of practical knowledge.
Why on-the-job learning matters for Latin America and the Caribbean
In our region, a large share of formal employment growth over the past two decades has come from services and administrative occupations (back-office support, customer service, routine analysis, administrative intermediation), precisely because they offered an accessible ladder into the labor market. Many of those jobs are also among the most exposed to generative AI. If those entry-level tasks are compressed too quickly, the risk, besides layoffs, reskilling congestion, discouragement, and early exit, lies in fewer opportunities to accumulate the practical judgment that workers typically acquire at the beginning of their careers.
The concern is that AI offloads precisely the trial-and-error tasks through which junior workers become experienced workers in the first place. In a region where weak schooling is often compensated, at least partially, through on-the-job learning, that loss could be especially costly.
Beyond its implications for workers, the erosion of learning‑by‑doing channels also matters for productivity and long‑term growth. Much of the productivity gain from adopting AI depends not only on the technology itself, but on the firm’s ability to integrate it into complex production processes, a capability that relies heavily on tacit knowledge accumulated through experience. In sectors where Latin America and the Caribbean could build comparative advantages, such as global services and other knowledge‑intensive activities, the loss of early‑career learning opportunities risks weakening firms’ absorptive capacity and, ultimately, the sustainability of those productive specialisations.
The IDB has identified this dynamic as a central challenge of AI readiness in the region: one with implications not only for workers, but also for firms’ productivity and the region’s capacity to sustain knowledge-intensive activities. Complementary labor formation, particularly dual education and supervised practical experience, is critical to ensure that AI adoption does not happen at the expense of young workers’ opportunities. The IDB has long supported initiatives toward that goal, most notably through finishing schools, where private firms and governments co-design programs for young workers to gain direct experience while learning the skills that AI will eventually complement rather than replace.
From this perspective, finishing schools are more than just employability programs; they are institutional mechanisms to preserve the transmission of tacit knowledge in contexts of rapid technological change. Evidence synthesised by the IDB on finishing school programs for the global services industry points to a consistent pattern: outcomes improve when training is co-designed with employers, anchored in real tasks rather than simulated exercises, and evaluated on demonstrated performance rather than credentials alone.
This is precisely the design that matters in a labor market where experience increasingly serves as a signal of value, alongside or instead of formal qualifications.
A realistic road map for AI skills
What should LAC countries do at this stage? Three directions stand out.
First, rethink education to prioritize learning by doing. If AI increasingly handles codified content, the comparative advantage of human-centered training may lie in cultivating nuanced, context-dependent judgment through practice. That means more apprenticeships, residencies, mentoring, supervised work experience, and structured learning environments—such as dual training and finishing schools—that intentionally embed practice into skill formation, not just more information.
Needless to say, this is not a call to substitute formal education, but to rebalance it at the margin, expanding experiential components where AI most directly reduces the scarcity of codifiable knowledge. The case rests on converging evidence: the classic returns-to-experience literature, the bank’s own operational record with finishing schools, and early findings on how AI automation of entry-level tasks is already compressing the channels through which tacit expertise is transmitted. Waiting for definitive causal proof carries its own risk: if those learning channels erode faster than policy adapts, the cost may prove difficult to reverse.
Second, make employment records portable and accessible. Right now, young workers carry credentials as their calling cards to the labor market. If practical experience is what increasingly matters, they should be able to document and carry it with them: a record of what they did, what they solved, and where they learned. Most countries have fragmented, private employment records inaccessible to workers themselves. Making these records portable and public would help workers validate their hands-on knowledge the way credentials validate formal education. This is a straightforward policy change with large upside.
Third, consider subsidies for early-career experience. Young workers without prior training are initially less productive but generate positive spillovers by learning on the job. As AI narrows entry-level openings, the economic case for public support of early-career experiential programs becomes stronger.
If we delegate too much of the trial-and-error, practice-based dimension of work to the algorithm, we risk depriving ourselves of the very experiences through which hard-to-document know-how is formed. The result could be a short-run productivity gain coupled with gradual erosion of the human capabilities on which that productivity ultimately depends, with substantial negative economic and social externalities.
The AI debate in our region should go beyond adoption, infrastructure, or regulation to ask what kind of human capital we are building.
Are we giving workers the practical experience to develop it? How much are we willing to pay today to give future workers a chance?