AI-supported Skills Elicitation and Formal Employment in Argentina

We run a randomized evaluation with rejected applicants to “Tu Empleo”, a youth employability program run by Fundacion Empujar. Through Fundacion Empujar, we will reach out to 1601 individuals above age 18 who are not admitted to the program but are actively looking for employment. The study will proceed in three individually randomized steps:
1) About half of the study participants will receive access to Brujula, an online AI-based chatbot, that identifies and categorizes skills from past work experience to create a user-tailored curriculum vitae (CV), including informal work. The other half completes a control online task. We track clicks, usage, and collect CVs generated with Brujula.
2) Two to four weeks after interacting with the tool we will administer a survey inquiring about job search behaviour, self-confidence, and willingness to pay for formal jobs attributes.
3) One to three months after receiving access to the chatbot, we will survey participants a second time to inquire about progresses in their job search and, if they found employment, their labor market outcomes. These will include but might not be limited to earnings, working hours, job amenities, and will have a particular focus on formality.

Participants in both groups can enter two lotteries, each giving a random chance to win one of 5 vouchers of 50USD each: any participant completing the conversation with the chatbot or the control task in step one and completing the first follow-up survey can enter the first lottery; any participant completing the second follow-up survey can enter the second lottery.

2026-03-13

2026-06-30

Brujula AI Usage, Job Search Behaviour, Labor Market Outcomes (in particular formality)

Brujula AI Usage: we want to quantify the use of AI tools on both the extensive and intensive margin, to be intended as the likelihood job seekers will open interact with the AI tool, and the quality of this interaction. Quality will be measured in terms of time spent interacting with the tool, if the skill report was produced, and potentially through skill reports textual analysis.

Job Search Behaviour: we want to estimate the causal effect of interacting with the tool on job search behaviour, with specific regard to search within the formal sector. In particular, we will measure time spent searching for formal employment, reservation wages for formal and informal employment, willingness to pay for formal employment characteristics.

Labor Market Outcomes: we want to estimate the causal effect of using the tool on labor supply. In particular, we will quantify differences in probability of finding employment and time spent looking for one, the "quality" of the found employment, measured both in terms of wages and other job amenities. We will pay particular attention on whether treatment improves opportunities in the formal labor market.

Intergenerational mobility, Brujula skills report, self-confidence, beliefs about job opportunities

Intergenerational mobility: we will gather information on whether parents were ever employed in the formal labor market in the first follow-up survey and use them to estimate if more disadvantaged youth (or whose family had less experience with the formal sector) can benefit more from using Brujula.

Brujula skills report: we will use text analysis to extract relevant information about treatment group's past experience on the labor market.

Self-confidence: we will measure if treatment changes confidence in own skills, confidence in ability to write a CV, confidence in ability to explain skills to potential employers, confidence about future job prospects, especially in the formal labor market.

We run an individual-level, stratified randomized evaluation with rejected applicants to “Tu Empleo”, a youth employability program run by Fundacion Empujar. Through Fundacion Empujar, we will reach out to 1601 individuals above age 18 to complete each step of the study. The goal is to explore if AI-powered tools can make job searching efforts more efficient, improve labor market outcomes, and function as low-cost and easy-to-access opportunities equalizers.

The experiment entails one intervention, where the treatment is assigned to half of the sample via stratified randomization, based on age (above/below 24 years old), education level (completed secondary education), parents' current emplyment status (any parent currently employed in a formal job), and rejection-phase to Tu Empleo, i.e. before or after detailed demographic and socioeconomic information are collected by Fundacion Empujar.

The intervention consists of giving the treatment group access to Brujula, a Spanish-speaking chatbot, powered by large language models, that functions as a personal career assistant. Brujula identifies and categorizes skills from any past experience (including those on the informal labor market) and creates auser-tailored skills profile, which can be used in as or in addition to a CV. The control group completes an online control tasks, that emails at testing cognitive skills broadly defined. The control tasks aims at making effort levels equal and allows to disentangle the effects Brujula's unique support from other forms of signals.

Then, participants will receive two follow-up surveys:

1) Survey 1, 2-4 weeks after treatment: we will ask participants to share information about their job search, confidence in own abilities, in writing a CV, in explaining skills to potential employers, and job opportunities, job preferences, reservation wages, and run discrete choice experiment to measure willingness to pay for formal attributes.

2) Survey 2, 1-3 months after treatment: we will ask participants to share information about their job search since the last survey, about their current job characteristics if applicable (including but not limited to earnings, work hours, type of contract, formal or informal job, sector of employment), updated reservation wages and willingness to pay for formal employment attributes.

Participants in both groups can enter two lotteries, each giving a random chance to win one of 5 vouchers of 50USD each: any participant completing the conversation with the chatbot or the control task and completing the first follow-up survey can enter the first lottery; any participant completing the second follow-up survey can enter the second lottery.

Primary outcomes include click-data on interactions Brujula, job search behaviour, self-cofidence, beliefs about job prospects, labor market outcomes on both the intensive and extensive margin. We will pay particular attention at if Brujula can improve opportunities on the formal labor market and act as cost-effective and easy-to-access opportunities equalizer.

Not available

Stratified randomization based on demographics collected by Empujar, done in office by a computer (R) with a set seed for reproducibility.

Individual

No

N/A

1601 individuals

802 individuals treated, 799 individuals control

The MDE is expressed at the individual level and standardized (units = standard deviations). Under our pre-analysis assumptions (two-sided α = 0.05, power = 0.80, 802 individuals treated and 799 individuals in control group, individual-level randomization, planned sample 1601), the standardized minimum detectable effect is 0.14 standard deviations. In other words, we are powered to detect a treatment effect equal to 0.14 × SD of the outcome (0.14 × √[p(1−p)] for binary outcomes).