No room for error

Artificial intelligence can accelerate learning for those starting out or take away the work where learning happens. Costa Rica needs to know which of the two things is happening in its labor force.

In 2008, when I was studying journalism, I arrived at the Costa Rican Social Security Fund to do my University Community Work (TCU). They assigned me to a still nascent unit that managed the institution's web portal and other digital assets. In a couple of months I learned to wrestle with HTML, CSS and tables, because that's how pages were built back then.

In January 2009 I returned as an intern, this time to the audiovisual communication unit. In my first week I had already been sent to Cinchona after the earthquake. By the time I finished the internship I had produced several dozen deliveries of the radio micro "Five minutes with you", had been an assistant on the "Health for all" program and had even recorded the executive president of the moment in the studio. None of that would have happened without Caro, Doxy and Patri, professionals with years of craft who took in the green kid who had just arrived.

It was like going from 0 to 100 in eight seconds: I dove headfirst into professional reality, but with someone there to correct me. A good part of what I did those months, from building a web page to producing a radio segment, today is solved by a professional with a free application or a $20 subscription. I wonder where I would have learned if that institution had done the math that many companies do today.

My story has nothing exceptional about it; that's how a good part of a generation entered the labor market. What changed is the calculation. Job postings still ask for experience from people who haven't had anywhere to get it yet, and now an AI agent drafts proposals, answers inquiries and writes code for a fraction of what it costs to train an intern. For the company, the savings are obvious. For someone looking for their first job, the question is different: where are they going to learn what they'll later be expected to know?

Those entry-level tasks served two functions at once. Reviewing documents, answering simple inquiries or correcting errors produced something useful and, at the same time, taught the trade under the gaze of someone with more experience. The institution that received me paid, in time and patience, for still-imperfect work, and in return trained future professionals. If AI delivers that result instantly, receiving a beginner stops looking like good business. The problem starts when all companies do the same math and expect someone else to train experienced staff.

Data from the United States is worth looking at carefully.In August 2026, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen updated their Stanford study with payroll records through June. Among workers aged 22 to 25 in occupations exposed to AI, employment was 19% below where it would have reached if it had followed the trajectory of their less-exposed peers. The gap was explained mainly because less hiring was happening, rather than because there were more exits. In the economy as a whole, the authors found no widespread displacement.[1]

The researchers themselves warn that these are early indicators that do not allow us to speak of causality. Some differences already existed before generative AI and were reduced when taking into account educational level. Nor does it mean that one in five youth positions have been lost. What it does suggest should concern us: an economy can maintain a good portion of its employment and, at the same time, narrow the entry door.[1]

In Costa Rica this hypothesis must be posed without blaming technology for any labor deterioration. According to INEC, in the June-August 2026 quarter there were 157,000 unemployed people, about 20,000 more than a year earlier, a statistically significant increase. The unemployment rate, of 6.7%, did not vary significantly, nor did the employed population. The survey portrays the labor market, but does not allow us to know what part of its behavior has to do with AI.[2]

Still, access to work already demands a generational reading. In April-June 2026, 55.7% of unemployed people were under 35 years old. At the same time, 89.1% of those looking for work had worked before, which recalls that unemployment also hits people with experience. It's worth reading these numbers carefully: they describe who is unemployed, not the youth unemployment rate or how many entry-level positions exist.[3]

Technological exposure, on the other hand, already has a name and surname. In February 2026, PROCOMER characterized a sample of 86 ICT companies: 76% used generative AI or digital agents and 72% exported services, with the United States as the main destination. The sample does not represent the entire economy, but it places the discussion in Costa Rican companies that are already adopting these tools and selling right in the market where changes in hiring are beginning to be seen. Waiting for proof of widespread damage to act would be arriving late.[4]

On the global level we also need to be precise. The International Labour Organization estimated in 2025 that one in four workers had an occupation with some degree of exposure to generative AI, but only 3.3% of world employment fell in the highest exposure category. For the ILO, the most likely scenario is that work will transform before it disappears. That an occupation is exposed means that some of its tasks can change, and that change can alter what a position teaches even if no one loses their job.[5]

The same technology can also work in favor of those getting started. Brynjolfsson, Danielle Li and Lindsey Raymond studied the arrival of an AI assistant to 5,172 customer service representatives. Queries resolved per hour increased 15% on average, and those with less experience gained speed and quality; the authors also found signals of learning. It's a particular environment, but the tool did something similar to what Caro, Doxy and Patri did with me: shorten the curve for newcomers. With one important difference: the assistant suggests, but doesn't hire anyone.[6]

For Costa Rica, the first step is to measure what happens at the employment entry point. Aggregate figures are not enough: data is needed on hiring by age and occupation, the experience required by job openings, and the degree of technological adoption in companies. We should see if entry-level positions are falling while requirements are rising, and compare this with external demand and other factors. Productivity can rise while opportunities shrink, and if we only measure the first, we'll be left with half an evaluation.

Training, for its part, requires agreements with the private sector, and Costa Rica doesn't start from scratch: the TCU, internships, dual education and professional practices already create space for those starting out. The country can expand dual education and paid practices that combine AI, supervision and real-world problems. ILO Recommendation 208 serves as a reference for quality apprenticeships, with structured training, pay and social protection. In September 2026, moreover, the MEP reported on exchanges with Mexico on curriculum, certification and links with companies. On that basis, an adaptation for digital and professional occupations can be proposed.[7][8]

The idea would be for apprentices to have to explain their decisions, verify what the tool produces and face the cases it solves poorly. Companies should recognize the time their people dedicate to mentoring, which almost never appears in any indicator, and evaluate what skills each person acquired. If there is public support, it should be tied to verifiable results in training and job placement, because subsidizing positions where no one learns would be repeating the problem and, on top of that, paying for it.

Automating entry-level tasks can save money today. Tomorrow, however, companies may be left without people capable of reviewing their systems, resolving exceptions and taking on responsibilities. Costa Rica has bet on competing with talent, and that talent is formed in large measure within the job. I was that spark plug kid who needed help, and every specialist started out much the same way. If the market only opens the door to those who already have experience, in a few years it will wonder, very surprised, why it can't find it.

Sources

1. Stanford Digital Economy Lab. Brynjolfsson, Chandar and Chen.Canaries in the Coal Mine: Six Facts about the Recent Employment Effects of Artificial Intelligence. August 2026 update.
2. INEC. Continuous Employment Survey: June, July and August 2026. Published October 1, 2026; copy of the INEC document hosted by Delfino.
3. INEC. Continuous Employment Survey Infographic: April, May and June 2026.
4. PROCOMER. Costa Rica consolidates exporting ICT sector with massive adoption of artificial intelligence and global vocation. February 25, 2026.
5. International Labour Organization. Generative AI and Jobs: A Refined Global Index of Occupational Exposure. Working paper 140, May 2025.
6. Brynjolfsson, Li and Raymond. Generative AI at Work. Study published inThe Quarterly Journal of Economics, 2025.
7. International Labour Organization. Guide for Policymakers on the Quality Apprenticeships Recommendation, 2023 (No. 208). December 2024.
8. Ministry of Public Education. Mexico and Costa Rica strengthen their cooperation in dual education. September 16, 2026.