IRC Signpost AI Answers Refugees in Seconds With Strict Safety Rails

When a family flees a war or a disaster, one of the first things they lose is reliable information. How do I register for asylum? Where can my child see a doctor? Which border crossing is open today? A wrong answer can cost days, money, or safety. The International Rescue Committee, one of the world's largest humanitarian organizations, has spent a decade building Signpost, a digital service that answers exactly these questions for people in crisis. Now it is using a purpose built AI platform called Signpost AI to answer them instantly, without handing vulnerable people over to an unguarded chatbot.

The stakes explain the caution. In March 2026, the IRC published a detailed account of its two year effort to deploy AI responsibly in humanitarian settings. Early pilots showed both promise and limits: the AI systems met quality standards roughly 75 percent of the time and increased staff productivity by nearly 70 percent, but every response still needed a human check before it could be sent safely. That gap between impressive and trustworthy is what Signpost AI was engineered to close.

From 20 Million People to a Platform

Signpost began in 2015 as a digital communications platform that co-created trusted, localized information with displaced communities and connected them to IRC social workers for personalized support. Since then it has served more than 20 million people across roughly 30 countries, growing into a network of more than 300 staff. Research consistently showed that the personalized assistance drove real impact. It also showed why it could not scale: there are simply not enough trained responders to answer every question from every displaced family in every language.

Rather than plugging in a generic chatbot, the IRC built Signpost AI, an agent building platform designed with partners including Google.org, Zendesk and Cisco. Teams design each agent the way they would write a job description: defined roles, responsibilities, safety protocols, guardrails, workflows and knowledge boundaries. Every agent draws only on verified Signpost content, a retrieval approach that keeps answers anchored to information humanitarians have already checked. Every deployment is monitored, and every use case is evaluated for safety, cost effectiveness and real world benefit.

Ten Seconds Instead of Three Days

The results are becoming visible in the field. In Mexico, the IRC launched SignpostChat, known locally as InfoDigna, in July 2026. Before AI support, people seeking help with asylum procedures, shelter or legal referrals often waited hours or days for a responder to reply. With the AI agent handling first responses, most inquiries now get an initial answer in under 10 seconds, freeing human staff for the complex cases where judgment truly matters.

Signpost AI enabled agents have now piloted in eight countries, supporting frontline staff and, in some contexts, engaging users directly. The same platform now powers agents across sectors: from protection and education to refugee resettlement and anti human trafficking work. A related IRC virtual assistant called ALMA helps refugees and other newcomers in the United States navigate documentation, employment and complex bureaucracies where confusion and delays can have lasting consequences.

Escalation Rules Come First

What distinguishes this effort from most chatbot deployments is that the IRC publishes its guardrails alongside its wins. Agents are red teamed before launch, tested for performance in low bandwidth environments and in dozens of languages, including those underrepresented in training data. Strict escalation rules define when an agent must hand a conversation to a human, and the IRC has treated those rules as the precondition for moving beyond pilots, not an afterthought. The pilot consortium model, which brought together the IRC, Mercy Corps, Internews and local partners, was credited with making the contextual grounding and trust possible.

The organization is also candid about what it does not yet know. Its March 2026 report states plainly that the jury is still out on whether current programs will drive the impact hoped for, and commits to publishing failures alongside successes so the wider humanitarian sector does not have to relearn the same lessons. Funding support from the Mastercard Center for Inclusive Growth and data.org through the AI to Accelerate Inclusion Challenge helped carry this work from experiment to working systems.

Boringly Reliable Beats Brilliant

The IRC's own framing of the lesson is that AI becomes useful in humanitarian work not when it is clever, but when it is boringly reliable and fits into the systems people already use. A displaced parent does not need a dazzling conversational model. They need a correct answer about where to take a sick child, in their own language, on the phone they already own, in seconds.

At a moment when global humanitarian needs are rising while budgets shrink, Signpost AI offers a template for how the aid sector can adopt AI deliberately: grounded in verified information, bounded by escalation rules, monitored by humans, and measured against real outcomes for the people it exists to serve.