Dispatches
Essays··8 min read

Mexico's AI School Needs a Payroll Test

Mexico's public AI school has graduated 3,001 of its first 12,000 starters and measures its success in certifications. The Inter-American Development Bank's new regional report puts the wage-outcome spread between good and bad scenarios at twenty-five percentage points — determined entirely by whether workers enter formal employment. Linking enrolment records to the social security registry, before the second cohort graduates, is cheap and is the measurement the programme actually owes.

How many of the young Mexicans who sat down for their first class at the country's free public AI school in February are now in a job that pays into the social security system?

It sounds like a simple question, the kind a ministry should be able to answer from a spreadsheet. It matters more this week than it did last week, because on 16 November the Inter-American Development Bank published the 2026 edition of its flagship Development in the Americas series, titled From Digitalization to Artificial Intelligence: Turning Promises into Productivity, and the bank's own summary of it puts the gain from broad AI adoption at 5.1 percent of regional GDP over a decade, with wages 4.2 percent higher. The same release says these technologies are "not 'plug-and-play'" and calls for targeted training and income support for displaced workers. Two months earlier, previewing the work, the bank was blunter. Reuters' Rodrigo Campos reported on 21 September that wages could rise 2.3 to 5.3 percent if workers move into jobs in expanding sectors, and fall 13.5 to 20.9 percent if they do not.

Look at those two ranges the way a teacher looks at a mark sheet. The spread between the good outcome and the bad one is roughly twenty-five points of wages, and the variable that decides which side of it a country lands on is worker movement. Mobility, in practice, means a person finishing a course, getting hired, and staying hired somewhere the new skill is paid for. Every regional government now has an AI training programme. Very few are measuring the one thing the IDB's model says determines whether the programme was worth it.

Mexico is the most interesting case on the continent, because it built big and published its numbers.

The Centro Público de Formación en Inteligencia Artificial opened in November 2025 under the Agencia de Transformación Digital y Telecomunicaciones (ATDT), offering free tracks in AI, data analysis, cloud, cybersecurity and software, co-certified by large technology companies. At launch the agency set a target of 25,000 certified graduates a year. President Claudia Sheinbaum's Second Government Report, delivered on 1 September and summarised by Mobile Time Latinoamérica, gives the first cohort's arithmetic: more than 25,000 applicants, more than 12,000 who started classes in February 2026, 3,001 graduates and 1,532 still active at the report's cutoff. The second cohort opened in May with room for 30,000. When it formally launched at the Biblioteca Vasconcelos on 16 October, La Crónica reported 23,000 participants and 17 partner companies, from Google and Microsoft to Tata Consultancy and Ericsson. ATDT head José Peña Merino restated the goal: 25,000 certifications by the end of this year.

Put the published figures side by side and the shape is clear enough.

First cohort (Feb 2026 start)Reported
Started classesmore than 12,000
Graduated by report cutoff3,001
Still active1,532
Neither graduated nor active (my subtraction)at least 7,467

Roughly one in four starters had graduated by September. A bit under four in ten were either finished or still going. If the second cohort of 23,000 completes at the same rate, it adds something under 6,000 graduates, and the year closes with perhaps 9,000 people through the door. The 25,000 figure can still be met, because the target counts certifications, and a graduate can hold more than one. One first-cohort graduate quoted in the Crónica piece earned two. That is a perfectly sensible way for an education programme to count its output. It is a poor way for an economy to count its return.

I want to be careful here, because the strongest objection to what I am about to argue is a good one, and it comes from the best evidence we have.

David Card, Jochen Kluve and Andrea Weber pooled more than 200 evaluations of active labour market programmes and found that average impacts are close to zero in the short run and become more positive two to three years after completion, with larger gains for programmes that build human capital. On that evidence, judging a school nine months after its first class opened is premature, and any column-writer demanding employment outcomes this year is asking the wrong question too early. There is a second line of defence. Online courses have always leaked students. Justin Reich and José Ruipérez-Valiente, looking at six years of MIT and Harvard open courses, wrote in Science in January 2019 that the bane of open online courses, low completion, "has not improved over 6 years". Against that baseline, a free five-month programme that graduates a quarter of its starters is doing well.

I accept both points on the facts. Neither changes what I would ask Mexico to do now.

The Card result is an argument about when to read the outcome, and that is exactly why the measurement has to be designed at the start. You cannot discover in 2028 what happened to the February 2026 cohort unless you recorded, in 2026, who they were, what they earned before, and how to find them in the administrative data afterwards. The MOOC comparison flatters the wrong peer group. The CPFI was never sold as an open course for the curious. It was sold, by the government and by its corporate co-certifiers, as a workforce pipeline answering an employer shortage. Its comparison is a training programme, and Latin America already has one of the best-studied training programmes in the world.

Colombia's Jóvenes en Acción paired classroom instruction with an internship inside a firm. The randomised evaluations by Orazio Attanasio, Costas Meghir and colleagues, followed up to eight years later in formal-sector records, found that those offered the programme were more likely to work in the formal sector, 36.8 percent against 33 percent, with formal earnings about 12.2 percent higher, and more likely to work at large firms. Those are modest numbers. They are also measured, and the measurement is what made the programme defensible to every finance ministry that later copied it. Notice the design: the employer was inside the course, and the outcome was read from the payroll, not from the certificate.

This is where the macro lens and the classroom meet. Mexico's labour market is not one in which a new skill flows smoothly to a new job. INEGI's monthly employment survey for September, published on 22 October, put labour informality at 54.3 percent of the employed, 32.6 million people, and found that 30.5 percent of the unemployed were aged 15 to 24. In that setting, "moving into expanding sectors" is a crossing from informal to formal work for a large share of the people the school is meant to serve. That crossing is the precise event the IDB's wage ranges hinge on, and it is the event Mexico's social security registry already records, worker by worker, every month.

There is a second reason the crossing matters, and it comes from inside the IDB itself. In an April working paper, Too Fast to Adjust, Eduardo Levy Yeyati models what happens when firms adopt AI faster than the retraining pipeline can absorb displaced routine workers. Some of them become discouraged and leave the labour force permanently, and the welfare-maximising adoption speed rises with retraining capacity. Read plainly, the paper says a country's retraining throughput sets how fast it can safely automate. Throughput is people placed. Certificates are an input to that number.

The demand side points the same way. The IDB's November release finds that 63 percent of digitally engaged firms in the region use generative AI, and 82 percent of those rely on off-the-shelf tools or light adaptations. Regional firms mostly need people who can configure, secure and run bought tools, the cloud, data and cybersecurity tracks the CPFI already teaches. That is good news for the school's curriculum. It also means the payoff shows up as a hire at a mid-sized firm in Guadalajara or Monterrey, which is precisely what a certification count cannot see.

My position is a flat one. Reporting certifications as the headline outcome of a national AI workforce programme is a mistake, and Mexico should stop doing it before the second cohort graduates. The fix is cheap by the standards of the rest of the plan; the same Second Government Report lists the Coatlicue public supercomputer at an investment of more than MX$6,000 million by the end of the term. Linking enrolment records to the social security registry, with consent collected at sign-up, costs a sliver of that. Publish, for each cohort, the share in formal employment at six, twelve and twenty-four months, set against a comparison group of applicants who were not admitted or did not start. Card and his co-authors would tell you to wait three years for the answer. Fine. Start the clock this month.

The first cohort's graduates are real people with real offers. One told the Crónica reporter she was hired without prior work experience after earning two certifications; another earned an advanced certificate at 42. Those stories are worth celebrating, and they are also the reason to count properly, because the programme deserves evidence stronger than its best anecdotes.

The harder question sits with the larger number in the table. At least seven thousand people who started in February are neither graduates nor still enrolled. Some found work before finishing. Some returned to an informal job they never left. Some may try again with the third cohort. Which of those is the common case, and what would Mexico change about its school if it knew?


Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.

Cartouche
Mexico's AI School Needs a Payroll Test · Dispatches, 17 November 2026 · T. Singh