# Answer people in crisis in seconds and pass the hard cases to a caseworker

The International Rescue Committee's Signpost answers questions from people displaced by conflict and disaster. Its AI agent tags each message, assesses risk, replies to routine questions and sends sensitive cases to human caseworkers.

Organisation: International Rescue Committee (Signpost)
Status: Piloted; the IRC says Signpost AI agents have been piloted in eight countries, and SignpostChat has run in Mexico as InfoDigna since July 2026
Place: Kenya, GR, IT, SV, MX, LY
Categories: Accessibility & inclusion, Financial inclusion & benefits
Built with: ChatGPT / OpenAI models, Claude, Gemini
Try it with: Claude, ChatGPT / OpenAI models, Gemini, Grok
Time to try: 30 minutes

## Why it matters

Signpost is a digital information service led by the International Rescue Committee (IRC). According to Cisco, one of its funders, it has helped 20 million people across more than 20 countries and in 12 languages find verified information about assistance. People who have fled conflict or disaster need answers about shelter, food, documents and their rights straight away, and caseworkers cannot reply to every message quickly. In Mexico, where the IRC began piloting the agent as InfoDigna in July 2026, Cisco's account reports that people now get a first reply to most questions in under 10 seconds, where it could take hours or days before.

## How it works

Incoming messages land in one shared queue. The agent tags the topic and assesses the risk. For a low-risk question it can send an automated reply that is flagged for later review; a case too sensitive for an agent goes to a human caseworker at once, and higher-vulnerability cases are flagged so caseworkers can deal with them first. Answers draw on vetted, local information: data.org describes the Signpost AI Information Assistant as a retrieval-augmented tool piloted in Kenya, Greece, Italy, El Salvador, Mexico and Libya. The IRC builds each agent on its Signpost AI platform the way it would write a job description, with roles, safety rules, guardrails and limits on what the agent knows, and it evaluates every deployment for safety, cost and real benefit. The same platform runs aprendIA, a chatbot that supports teachers in crisis-affected north-east Nigeria.

## Who it's for

Humanitarian agencies, NGOs, migration and city services, and any helpline that answers the same urgent questions in many languages with too few caseworkers.

## The prompt

```
You are helping me design a first-reply assistant for a helpline that serves people in a crisis (for example, newly arrived refugees). You will not answer real people. Your job is to draft the rules, test them, and show me where a human must step in.

Our service: [who you serve, where, and in which languages]
Our verified information: [paste the facts you are sure of: addresses, opening hours, eligibility rules, phone numbers]
Our caseworkers: [how many, their hours, how a case reaches them]

1. List the ten questions people are most likely to ask us, using only the information above.
2. Sort them into three groups: routine (a reply from the verified information is enough), needs a caseworker soon, and urgent or sensitive (safety, violence, health emergencies, children on their own, legal deadlines). Explain each choice in one line.
3. For each routine question, write a reply of no more than four sentences, in plain words, quoting only the verified information. Where the information does not cover the question, say so and give the route to a person.
4. Write the hand-over message for urgent or sensitive cases: what the person is told, what the caseworker receives, and how fast.
5. List the words or signs of distress that must always send a conversation to a person.
6. Finish with five test messages, including two tricky ones, and how the rules above would handle each.

Never invent addresses, phone numbers, rules or promises of help. Do not ask for names, ID numbers or exact locations in these drafts. Mark anything you are unsure of for a caseworker to check.
```

## Caveats

The figures come from the IRC and its funders, and no independent evaluation of answer quality was found. The IRC says its earlier Signpost chatbots met its quality standards roughly 75% of the time and still needed human review before answers went out. Sources differ on scope: the IRC says agents have been piloted in eight countries, data.org lists six pilot countries and 9,202 people served, and Cisco's figure of 20 million people covers Signpost as a whole, not the AI agent alone. AP reported in 2024 that the pilots used models from OpenAI, Anthropic and Google; the sources do not say which models run today. A wrong answer to someone in a crisis can do real harm, so any copy of this needs vetted local information, human review and a clear route to a person.

## Sources

The IRC page is the organisation's own account, and the Cisco and data.org pages come from funders, with figures supplied by the IRC. The AP report is independent but covers the 2024 pilot, before the current agent. No independent evaluation of answer quality or outcomes was found.

1. International Rescue Committee: "Signpost AI, an AI agent building platform"; "piloted in eight countries"; earlier systems "met quality standards roughly 75% of the time" and needed human review; aprendIA for teachers in Nigeria. From Promise to Practice: How We're Scaling AI for Humanitarian Good (https://www.rescue.org/irc-responsible-ai-humanitarian-sector)
2. Cisco Blogs (Cisco Crisis Response): "helped 20 million people"; "more than 20 countries and in 12 languages"; "one shared queue"; InfoDigna pilot in Mexico from July 2026; first reply "in under 10 seconds". Connecting people to humanitarian assistance with responsible AI (https://blogs.cisco.com/?p=498761)
3. data.org: Signpost AI Information Assistant, a retrieval-augmented tool piloted in Kenya, Greece, Italy, El Salvador, Mexico and Libya; 9,202 people served. International Rescue Committee (IRC), AI to Accelerate Inclusion Challenge awardee (https://data.org/?p=28181)
4. Associated Press (Thalia Beaty), republished by The Living Library: pilots in El Salvador, Kenya, Greece and Italy, answering in 11 languages; models from OpenAI, Anthropic and Google. AI could help scale humanitarian responses. But it could also have big downsides (https://thelivinglib.org/?p=59526)

Fact-checked: 24 Oct 2026

Page: https://escapevelocitysociety.com/picks/irc-signpost-crisis-information-agent

See Selected Projects: https://escapevelocitysociety.com/picks.md
