Why Intent Changes Everything: The International Rescue Committee's AI Operating Model for Humanitarian Crises

An IRC worker in Mexico shares Signpost information How do you deploy AI in situations where the people you're trying to support may need information, trust, reassurance, empathy and human contact at one of the worst moments in their lives? That was the question I put to Andre Heller, director of the Signpost project at the International Rescue Committee (IRC). And his answer was interesting — you do it carefully. But most of all, you do it with an intent to extend the reach of human support, not to automate as many of the interactions as possible. Heller’s introduction to the work of Signpost shows what that human support involves: > The Signpost project is essentially a light touch form of social work that's conducted over digital channels, aimed at supporting populations impacted by armed conflict or disaster. We support them to understand what their options are throughout the arc of a crisis — how to access services, what programs they're eligible for, what kind of rights they have and how to exercise those rights. But delivering that kind of support at the scale necessary for fast-moving, large-scale humanitarian crises can turn into a mess without the right structures — something Heller quickly discovered after picking up his role in 2020. Scaling humanitarian support Providing support to displaced populations during moments of crisis is a complex business. First responders on the ground are often struggling with physical needs such as shelter, food, and medicine. But Heller says that information needs can be just as important — knowing who to trust, what to ignore, and what rights and support are available: > What we learned as aid providers during the Syrian refugee crisis of 2015 was that people had a basic need for information — to know what was trustworthy, what was true and what was wrong. Because in contexts like this information can be just as important as healthcare, food and shelter. But Heller says providing this kind of support is not simple. While many of the people Signpost supports are digitally savvy, every crisis still requires its own specific set of social media accounts, online content and support workers. And after taking up his role in 2020, Heller says he realized that the tools and practices put in place during the original pilot could no longer scale — especially once COVID-19 turned digital support channels into critical infrastructure: > The project technology foundations were very clunky — fragmented, expensive, and very hard to scale. And with the pandemic, a lot of the face-to-face work that responding agencies were doing was completely forbidden, making the need for digital work very obvious. He explains this need for firmer foundations led the team to Zendesk’s “Tech for Good” program: > We met Zendesk right after the pandemic struck and their platform helped us tie together social media channels, content management and ticketing so we could make sense of different requests. Because [with] many different language groups operating, we need to get requests to the person who can respond to them — and having a centralized architecture that does this is incredibly powerful. But the benefit went beyond tying systems together. Heller says it also gave Signpost a way of deploying support for new crises in a repeatable way: > Zendesk gave us the infrastructure to connect people, but also a replicable process in which each deployment can be tailored to specific communities. And we also had standard training materials on how to engage with people in a way that's empowering and respectful of trauma, as well as on the tools. It gave us this governable set we could deploy and quickly adapt. In Heller’s telling, therefore, Zendesk is not framed as a tool to queue or deflect tickets, but as part of an operating model for deploying humanitarian support — one which aims to connect people with the right information or human responder as effectively as possible given the language, context and community knowledge required. Adopting AI for purpose As we turn to AI, I suggest to Heller that the consequences of providing the wrong information in a humanitarian environment can be significant, precisely because the people reaching out are already in crisis. Having set up such a careful operating model I want to know how it shapes their thinking about where AI might — and might not — be appropriate. Heller starts by going deeper into the nature of the work itself: > What we've found through doing impact research on our program is that the broadest way to make an impact is through social media channels and a kind of mass broadcasting of information. But the biggest and deepest way to make an impact is through one-to-one sessions, when people reach out and say, hey, I’ve got four problems and I really need help. And he goes on to explain that it was the spike in work triggered by a viral post that brought AI to the forefront of discussions: > We shared a post at a time when people were trying to evaluate their options for asylum outside of Afghanistan. And this post went viral on social media — we received about 50,000 support requests for a team of three people over the course of a couple months. Needless to say, that's very hard to do with that many people. So we thought maybe AI can respond to some of these queries if we get it just right. Heller says the early pilots kept humans in the loop to verify accuracy and showed promising productivity gains — but that the conversation quickly moved on to how far AI could safely go: > We thought that AI could do more if we could figure out how to make sure it was safe and knew the limits of its knowledge. Sometimes things are too sensitive or needs are too acute. So we developed a framework where some queries are sent to humans while others are taken by AI — with an elaborate number of guardrails on either end to de-risk [our] high-risk use cases. I ask Heller if he can expand further on how it works in practice: > When a support request comes in, the AI evaluates the knowledge base to see if it has enough information to answer, and also analyzes the intent to determine whether this is a high-risk case that requires empathy and social skills. If so, AI has no business responding and creates a ticket for a person. Otherwise it responds — but maybe flags that a human should follow up. But while this answers the appropriateness question in a limited way, it still doesn’t answer the deeper part of Heller’s own distinction. Because if one-to-one support creates the deepest impact, how do we protect the empathy, reassurance and judgment that creates that impact when AI sees the request before a human does? Making human-centric AI work I press Heller a little further on this point. Because while he clearly believes that AI “has no business” responding to a wide range of support requests, he is also equally clear that he has to balance the deepest possible benefit against the broadest possible impact. But if many of the people accessing the service are vulnerable, I suggest that they might be looking for human reassurance as much as factual answers — and so I'm curious how he thinks about the trade-off between these two seemingly opposite poles. His answer suggests that the right decision is always contextual: > We like to say there is no need to choose between those two things. The programs we run will always have a pathway to a human, but with a 24-hour SLA. So you might have to wait a little bit. But the AI might be able to give you some information straight away, before someone reaches out. And there are other cases where the need is just information and that's the point. So we try not to have to choose. The strong subtext, at least for me, during the discussion is that it’s a mistake to assume that empathy always means delaying help until a human is available. Heller’s belief appears to be that, in many cases, the most empathetic response is to provide reliable information immediately, so that some