
SignpostChat AI: IRC Combines Artificial Intelligence and Human Oversight in Humanitarian Support
SignpostChat AI, a new artificial intelligence initiative from the International Rescue Committee (IRC), is designed to help people affected by crises access information about shelter, food, safety and other essential services more quickly. Rather than replacing humanitarian caseworkers, the system uses AI to handle suitable enquiries while escalating sensitive situations to human specialists.
The initiative builds on the IRC’s Signpost digital information service, which has helped 20 million people worldwide access timely, verified information about assistance during crises. Cisco, a longstanding supporter of the programme, is backing the development of this latest AI-enabled service.
Building AI into an established humanitarian information service
For people displaced by emergencies or arriving in unfamiliar locations, finding reliable information about available assistance can be difficult. Questions about where to sleep, how to obtain medical attention or where to find food may require immediate answers.
The IRC leads the consortium behind Signpost, which was developed to connect people in crisis with relevant, verified information. According to Cisco, the service has reached people in more than 20 countries and operates in 12 languages.
Cisco’s involvement dates back to 2015, when funding from the Cisco Foundation helped launch the programme’s precursor. Subsequent support has contributed to the development of its infrastructure, including service maps that can be accessed offline and unified data analytics.
The latest development, SignpostChat, introduces AI into the platform to help caseworkers manage growing volumes of enquiries and extend the service’s reach.
Cisco says its partnership with the IRC reflects a shared commitment to responsible AI. Its approach is based on six core principles, although the source material does not list those principles individually.
Why SignpostChat AI keeps human caseworkers involved
Until now, trained Signpost caseworkers, many with lived experience relevant to the communities they serve, have personally responded to questions submitted through the platform. While this approach provides people with credible information, the number of enquiries a team can handle is limited by its available personnel.
Briana Orr, the IRC’s Senior Advisor for Responsive Information, highlighted this constraint: a caseworker can answer only a finite number of questions during a working day.
AI offers a way to increase response capacity. However, humanitarian enquiries can involve vulnerable people and highly sensitive circumstances. An incorrect or incomplete automated response could expose someone to further harm.
The IRC therefore designed SignpostChat around a human-in/on-the-loop model. AI agents handle enquiries when it is appropriate for them to do so, while human caseworkers retain responsibility for situations requiring specialist judgement or intervention.
A system designed to recognise when not to respond
The development process involved several stages of testing and refinement.
Initially, the IRC used simulations, AI-assisted evaluations and red teaming exercises to identify potential weaknesses and establish safeguards. Red teaming involves deliberately testing a system to uncover vulnerabilities and potentially unsafe behaviour.
The team then developed and tested a triage methodology to determine how the system should assess and route incoming enquiries. Specialists also refined the agents’ communication, including their language choices, and trained them to recognise signs of distress.
Incoming messages are placed in a shared queue. The AI agent identifies the subject of each enquiry while assessing its potential risk.
Depending on the circumstances, the system can:
- Provide an initial response and flag the enquiry for subsequent human review.
- Escalate a sensitive enquiry immediately to a caseworker.
- Handle routine enquiries when the system determines that an automated response is appropriate.
According to Orr, the aim is to ensure that AI agents answer questions when it is safe to do so and transfer conversations to a person when it is not.
SignpostChat can provide immediate responses through text and voice, including in underrepresented languages. This allows human moderators to concentrate on enquiries requiring greater attention, rather than making AI the final authority on every situation.
SignpostChat AI pilot in Mexico reports faster initial responses
The IRC began piloting SignpostChat in Mexico in July 2026. The service operates there under the name InfoDigna and supports a small Mexico-based team responding to people seeking shelter, food and other assistance.
Luis Alberto Tejeda Carrasco, who led the pilot’s initial phase with the IRC Mexico Signpost team, said one of the most noticeable changes has been the speed of initial responses.
Before the pilot, the team could take several hours or, in some cases, days to respond to all incoming messages. Since the AI agent went live, most enquiries have received an initial response in under 10 seconds, according to Carrasco.
The system also identifies cases involving greater vulnerability, allowing human caseworkers to prioritise people who need additional support.
Helping a family find shelter in Mexico City
Carrasco described an early example involving a mother, grandmother and child who had recently arrived in Mexico City. The family feared for their safety and faced homelessness, without resources or a support network to rely on.
Using information provided by the AI agent, they were able to connect with one of the city’s largest and safest shelters on the same day.
The family subsequently sent feedback through the service, expressing relief and gratitude that they no longer had to sleep on the street.
Carrasco said the experience illustrated the purpose of the hybrid model: combining human expertise and AI to respond to more requests while keeping people’s needs and safety central.
He also emphasised the principle guiding the initiative: AI must strengthen, rather than undermine, the equity, safety and dignity of the people receiving assistance.
What comes next for SignpostChat
The IRC is continuing to document lessons from the Mexico pilot. These findings are expected to inform further deployments of SignpostChat and other AI-enabled initiatives within the organisation.
One area under consideration is the earlier integration of advanced red teaming and model stress-testing into the development process. The objective is to identify potential risks earlier and prepare the system to expand without compromising the human-centred support on which the service depends.
For Cisco, the partnership demonstrates how AI could contribute to humanitarian assistance when its development is accompanied by appropriate safeguards and oversight.
As SignpostChat evolves, its central design principle remains that technology should extend the capabilities of humanitarian workers rather than replace their judgement. In crisis situations, faster responses matter, but the reliability and safety of the information provided are equally important.