What if artificial intelligence could help us see poverty before it happens? For the first time, to our knowledge, AI is being used to measure household vulnerability—not just by analysing data behind a screen, but by directly engaging with people through household surveys. In partnership with IDB Lab, we've developed a groundbreaking method that does just that—piloted in Colombia between January and February 2025, using our pioneering avatar-based survey platform.

Background and origins

The idea originated in 2024, when we began exploring with IDB Lab how AI could support more accurate and nuanced assessments of vulnerability, especially in relation to income. We envisioned an "Income Plus" approach: a method that not only identifies individuals and households experiencing income poverty, but also distinguishes between those who are vulnerable to falling into poverty and those who are relatively resilient.

This approach stems from a critical gap we observed—while income is often the primary lens through which vulnerability is assessed, it does not capture the full picture. We wanted to go further: to understand how other social, structural, and household-level factors—often deeply intertwined with income—can amplify or mitigate vulnerability.

We asked ourselves: Would households with a disabled family member show greater vulnerability to income shocks, even if their current income appears stable? What about people from historically marginalised groups, such as those of Afro-descendant or Indigenous descent—are they more likely to face income volatility or exclusion from social protections, and how can this be measured more holistically?

Throughout Q4 2024, we immersed ourselves in academic research. We reviewed established tools such as the Household Vulnerability Assessment Tool (HVAT) and the Socioeconomic Vulnerability Index (SeVi), drawing inspiration from their strengths while identifying gaps that a more AI-driven approach could help fill.

We were particularly intrigued by the possibilities that emerge when AI conducts household surveys directly. Could AI avatar-based surveys provide a more neutral, stigma-free environment—where people feel more comfortable sharing sensitive information? Would respondents disclose more about topics like disability, unemployment, racial discrimination, or income instability if they knew they were speaking to a non-judgemental digital interviewer rather than a human?

Existing evidence suggests this might be the case. For instance, a study by Lucas, Gale M. et al. (2017), conducted at the University of Southern California and published in Frontiers in Robotics and AI, showed that service members disclosed more symptoms when speaking with a virtual human interviewer than when filling out an anonymous self-report form. Even though both tools were designed to reduce stigma, the AI-based interaction led to deeper and more honest disclosures.

Designing and piloting the Economic Resilience Index (ERI)

Together with Yuri Soares and Maria Claudia Ventocilla from IDB Lab, we partnered with UNI2, a Colombian microfinance institution that offers microcredits through an inclusive lens. We recognised that traditional surveys often fail to capture the full complexity of vulnerability, which is why we wanted to ensure that the survey not only gathered quantitative data but also provided a platform for participants to share their personal experiences and insights.

In collaboration with IDB Lab, we developed a 21-question survey that combined multiple-choice questions with open-ended prompts, giving participants the opportunity to speak freely—something that is not always possible with the traditional written questionnaires commonly used in this type of study. Our method, shaped by the ideas we had been developing throughout the previous year, was ultimately named the Economic Resilience Index (ERI).

The ERI categorises households into three levels:

• Resilient household - score of 70 or above
• Vulnerable to income poverty - score between 30 and 70
• Already in poverty - score of 30 or below

We conducted the survey using a female digital avatar capable of speaking Spanish with a Colombian accent. With the support of UNI2—particularly Sebastián Ayalde, Director of Social Development—we distributed the survey via WhatsApp links to as many UNI2 microcredit beneficiaries as possible. A total of 177 individuals completed the full survey.

Key findings from the pilot

Income and food security were the most pressing issues. Most respondents earned below the Colombian minimum wage, and often far less than the average income in other countries with similar GDPs, such as Mexico or Chile.

Over half reported food insecurity in the past week, with 8% experiencing it daily. Although 80% had worked recently, most did so informally, and only 22% had access to workplace benefits like paid leave or social protection.

Education levels were relatively high—most had completed secondary school or a technical diploma—but income returns were limited, especially for women, who were largely absent from higher education levels. One in four respondents cared for an older person or someone with disabilities, adding to household pressures often overlooked in traditional assessments.

Diversity also mattered. More than a third of respondents identified as Indigenous, and nearly a quarter as Afro-Colombian—groups often facing structural exclusion and fewer opportunities. While most households had access to basic services and reported feeling safe in their neighbourhoods, nearly a third had experienced major shocks in the past year, such as illness, accidents or job loss, often relying on informal work or family support to cope.

When applying our method, this population was classified with an overall score of 51, placing them in the category of vulnerable to income poverty.

Our plan now is to continue piloting this method in other Latin American countries, with the support of IDB Lab. We believe AI has the potential to reveal a more complete and human picture of what vulnerability means—and, in the future, to become a powerful tool to anticipate vulnerability even before it fully arrives.