Info Pa'lante Colombia: Framing a Case Study in Scaling Responsible AI

Colombia hosts more Venezuelan migrants and refugees than any other country in the world. Driven out by economic collapse, violence, and political repression, more than 2.9 million Venezuelans have sought safety and opportunity in Colombia — many arriving with little knowledge of their legal rights, available services, or how to navigate a complex and often overwhelming system. Information gaps don't just cause confusion; in this context, they cost people access to healthcare, legal protection, work, and safety.

What Info Pa'lante Is

Info Pa'lante is the IRC's Signpost information platform for Colombia, launched in September 2020 with support from Tripadvisor, Google.org, Zendesk, and NetHope. Its name — Colombian slang for "forward" or "keep going" — reflects its purpose: to give displaced Venezuelans the information they need to move forward in their new country with confidence and agency.

The platform provides verified, up-to-date information on legal rights, asylum and regularization processes, access to health and social services, employment, and safety — delivered through WhatsApp, Facebook Messenger, and an online chat interface, and complemented by a service map covering more than 300 verified service points across Colombia. A team of trained human moderators responds to queries in real time, ensuring that the most vulnerable users receive personalized, trusted guidance — not just a list of links.

Info Pa'lante operates at a scale that is exceptional even within the Signpost global network:

  • 2.3 million users reached in Colombia

  • 15,000 complete two-way user journeys completed annually

  • 65% website engagement rate — among the highest in the Signpost portfolio

The Challenge AI Is Designed to Solve

Info Pa'lante's human moderator model is what makes the platform trusted and effective — but it is also the primary constraint on its scale. Moderators can only respond to so many queries. Routine questions — eligibility for Temporary Protected Status, how to access health services, what documents are required for enrollment in schools — consume significant time that could be redirected to the complex, high-stakes cases that genuinely require human expertise and empathy.

As demand continues to grow and global aid funding contracts, the gap between what Info Pa'lante's team can handle and what its users need is widening.

Responsible AI Deployment

The Signpost AI platform is being deployed with Info Pa'lante to introduce a hybrid human-AI model: an agentic AI that handles the high volume of routine queries independently, freeing human moderators to focus on the cases that need them most.

The agent will be capable of digitally onboarding new users, responding to standard queries from the knowledge base with accurate, culturally appropriate information, and escalating complex or sensitive cases — protection concerns, urgent legal situations, mental health needs — directly to human staff, with full conversation context preserved.

The result is not AI replacing the human element that makes Info Pa'lante trusted. It is AI extending the reach of that human element to far more people, at far lower cost, without sacrificing the quality or safety that vulnerable populations depend on.

The potential to serve significantly more Venezuelans with the same or fewer resources — while maintaining the quality, trust, and safety standards Info Pa'lante has built — represents exactly the kind of humanitarian efficiency that responsible AI can deliver.

Getting there

In the following articles in the Info Pa’lante series, we will explore the AI safety and performance testing process, document lessons learned about AI deployments with a focus on partnering, and at a later stage, a publication on progress of the pilot post-launch with real world impact figures and deeper analysis of cost and quality.

Previous
Previous

Stress-Testing AI at the Humanitarian Frontier: Findings from Signpost Colombia's First AI Simulation Round

Next
Next

How Evidence Shaped aprendIA: