From Learning to Earning: What Peru Is Testing With an AI Career Coach

A World Bank programme in Lima is testing whether a large language model can deliver the personalised half of career guidance at scale. It launched in March 2026 across 110 schools and is being evaluated as a randomised controlled trial. The results are not published yet, and the connectivity constraint is unsolved.

Published 2026-08-20 · By Shahzad Asghar

A World Bank programme in Lima is testing whether a large language model can do the part of career guidance that has never scaled: the personalised part. It is called Eligiendo Mi Camino, it launched on 16 March 2026 across 110 public schools, and it is being evaluated as a randomised controlled trial rather than announced as a success. That last detail is the one worth paying attention to.

I spend my working life on the gap between what a system does in a demonstration and what it does in an operating environment. This programme is interesting because it is honest about which side of that gap it is currently on.

What Eligiendo Mi Camino actually deploys

The programme puts two tools into fifth-year secondary classrooms. One is an AI mathematics tutor built on roughly 4,000 curriculum-aligned questions covering four competencies, sixteen topics and eighty subtopics. The other is an AI career coach.

The career coach is named Gallito, after the Gallito de las Rocas, the Andean cock-of-the-rock. It covers 142 occupations across twelve economic sectors, and it draws its salary and employment data from Peru's Ministry of Labour rather than from the model's training data. That distinction matters more than it sounds, and I will come back to it.

The scale is 110 public schools, 6,600 students and 400 trained teachers. Anthropic supplies the model, with uDocz and Microsoft as delivery partners.

Why personalised guidance has never scaled

The problem Gallito addresses is not that students lack information. It is that generic information does not change what students do.

A careers leaflet listing median salaries by sector is cheap to produce and largely inert. Intensive human mentoring does change behaviour, and it is priced accordingly: a counsellor can hold a few hundred conversations a year, which is why the students who most need the conversation are the least likely to get it.

The Peruvian context makes the stakes concrete. The World Bank reports that seventy per cent of students finishing secondary school in Peru do not go on to higher education, and that across Latin America fifty-six per cent of children cannot read a simple text by age ten. A guidance system that only reaches the already-advantaged compounds both numbers.

What the design gets right

Three choices in this deployment are worth copying, and they are architectural rather than clever.

It grounds the model in an authoritative dataset. Salary and occupation data comes from the Ministry of Labour. The model conducts the conversation; it does not supply the facts. Anyone who has watched a language model confidently invent a wage figure understands why this separation is the whole design.

It is structured, not open-ended. The career coach runs eight defined steps across twelve sessions of ninety minutes. That is a curriculum with a model inside it, not a chatbot with a hopeful prompt.

It sits inside existing school infrastructure. No new devices, no parallel programme, no dependency on a household owning anything.

What is not yet known, and why that matters

Here I have to be careful, because a great deal of commentary about this programme is running ahead of the evidence.

The randomised controlled trial covers all 110 participating schools and is led by Ezequiel Molina and Carolina Lopez. It is running. As of the World Bank's July 2026 update, roughly 4,500 students across 85 schools had taken part.

What the World Bank has not published is outcome data. There are no effect sizes, no utilisation rates, no measured change in enrolment intent or in students' accuracy about wages. Figures of that kind are circulating, and I could not source any of them to the World Bank's own materials. Until the trial reports, the honest description of this programme is a well-designed experiment in progress, not a demonstrated result.

That is not a criticism. Running a properly powered randomised trial instead of publishing a case study is the more rigorous choice, and it is rarer than it should be.

The constraint nobody has solved yet

Gallito runs against a cloud-hosted model. That is a reasonable decision in metropolitan Lima and a hard limit everywhere else.

The students furthest from good career guidance are, with grim reliability, the students furthest from reliable connectivity. Any system whose personalisation depends on a round trip to a data centre serves the periphery worst, which is the same failure pattern I have documented in the Last-Mile AI Framework: the pilot succeeds where the infrastructure already works, and the deployment fails where it does not.

The architectural answer is small language models running locally on modest hardware, with the authoritative dataset held on device and synchronised when a connection appears. That is now technically plausible in a way it was not two years ago. It is not yet the default, and the gap between a cloud demonstration and an offline deployment is where most of the remaining engineering sits.

What to watch when the results land

Four things will tell you whether this generalises.

Whether the effect survives at the tail, in the schools with the worst connectivity and the least teacher capacity, rather than only in the average. Whether stated intent converts into actual enrolment, which requires follow-up beyond the school year. Whether the grounding in Ministry of Labour data holds up when salary figures move. And what the per-student cost actually is at scale, which nobody has published.

If the results hold, the finding is not that AI can give career advice. It is that the expensive, personalised half of guidance can be delivered for something close to the cost of the generic half. That is a public-sector procurement question as much as a technical one, and it belongs alongside the governance work any ministry would need to do before adopting it.

Frequently asked questions

What is Eligiendo Mi Camino?

Eligiendo Mi Camino is a World Bank programme in Lima, Peru, that places an AI mathematics tutor and an AI career coach into fifth-year secondary classrooms. It launched on 16 March 2026 across 110 public schools, reaching 6,600 students and 400 trained teachers, and it is being evaluated through a randomised controlled trial.

What is Gallito, the AI career coach?

Gallito is the programme's AI vocational coach, named after the Gallito de las Rocas, Peru's national bird. It holds a structured conversation with a student across eight steps and twelve ninety-minute sessions, covering 142 occupations across twelve economic sectors, using salary and employment data from Peru's Ministry of Labour.

Has the programme been proven to work?

Not yet. The randomised controlled trial is under way across all 110 schools and the World Bank has not published outcome data. Effect sizes, utilisation rates and enrolment changes attributed to the programme are circulating but cannot be sourced to the World Bank's own materials. The correct description today is a well-designed experiment in progress.

Why does connectivity limit this kind of programme?

The career coach depends on a cloud-hosted model, so every personalised exchange requires a working internet connection. The students who most need guidance are usually those with the least reliable connectivity, so the design serves the periphery worst. Moving the capability to a small language model running locally, with the occupational dataset held on the device, is the route to rural deployment.

Which AI model powers the programme?

Anthropic is the model provider, with uDocz and Microsoft as delivery partners. The model conducts the conversation while the occupational and salary data comes from Peru's Ministry of Labour, which keeps wage figures out of the model's generative output.

---

*Sources: World Bank, Eligiendo Mi Camino programme brief, and Artificial Intelligence in Action in Latin America's Schools: Evidence from Peru, July 2026.*

*Shahzad Asghar is Head of Data and Digital Solutions at UN-ESCWA, with more than twenty years building AI and data systems across United Nations operations. He created the Last-Mile AI Framework and writes on AI governance, humanitarian AI and information technology audit in the age of AI.*

← All articles