Relief Applications

Humanitarian chatbots: when to build one and how

A chatbot can give people reliable information in their own language, on an app they already use, at any hour. It can also give a wrong answer to someone in a hard situation, with nobody watching. This guide sets out when a humanitarian chatbot is worth building, the main design choices, and the safeguards we put around one.

The short answer

A humanitarian chatbot is worth building when many people ask the same questions, the answers can be written down and checked, and your users are already on a messaging app. Choose a rule-based design where answers must be exact, and add a language model only inside limits you set. Design for low literacy with menus and buttons, hand over to a person for anything sensitive, test with the people who will use it, and check what the chatbot stores before launch.

When is a chatbot the right tool?

A chatbot suits a specific kind of need. Look for these signs.

  • Many people ask the same questions, such as what documents they need, where a service is, or what their rights are.
  • The answers can be written, checked and kept up to date by your team.
  • Your users already use a messaging app, and a phone call or a website would be harder for them.
  • Questions arrive outside office hours or in more languages than your staff speak.
  • You want to collect structured information, such as a survey or a complaint form, in a conversation people find easy.

It is the wrong tool when the questions are different every time and need judgement, when a wrong answer can cause serious harm and nobody reviews it, or when your users have no phone, no connection or no electricity. A chatbot adds to a service. It does not replace the people who run it.

Should the chatbot follow rules or use a language model?

Most of the design depends on this choice. The two are not opposites. Many good chatbots use both.

Rule-based
How it works: The user picks from menus and buttons, or the bot matches fixed keywords. Every answer is written in advance. Strengths: Predictable and easy to check. Suits legal, medical and safety information. Works well for people who read little. Risks: Cannot answer what you did not plan for. Needs the content kept up to date by hand
Language model with limits
How it works: A large language model writes the reply from your approved content, within rules set through prompt engineering. Strengths: Understands free-form questions and varied wording. Can reply in several languages. Risks: Can make up an answer if it is not held within limits. Needs testing and monitoring
Either, with handover to a person
How it works: The bot answers what it can and passes the rest to staff. Strengths: Gives people a path when the bot cannot help. Risks: Needs staff time and a clear response promise

Our approach is to let you decide what the model may say. For the Migrant Recruitment Advisor chatbot, built for the ILO and ITUC, we added a large language model and set boundaries on its replies through prompt engineering, so that its answers stay within what it was meant to cover. Where exact wording matters, a fixed answer is safer than a generated one.

Which messaging app should it run on?

Go where your users already are. The right app depends on the country and the community, so ask them rather than assume. A single chatbot can often run on several apps from one set of content. The Migrant Recruitment Advisor runs in 8 languages on WhatsApp, Facebook, Telegram and Line, with a visual menu of countries and forms in the chat. A separate chatbot for the ILO collects data on working conditions through Facebook Messenger.

Remember what this means for data. A messaging app is a third party. What people type passes through it, and each platform has its own rules on business accounts and automated messages. Read those rules before you choose, and say plainly in the chatbot what it keeps and what it does not.

How do you design it for people who read little or have little time?

A chatbot is not a website in a chat window. People use it on small screens, often in a rush, sometimes with limited reading skills. Good design removes typing.

  • Use buttons, menus and icons in place of long text. The working conditions chatbot we built for the ILO was designed around visual interaction and less writing.
  • Keep each message short, with one idea. Put the most useful line first.
  • Say in the first message that it is a chatbot, what it can help with and what it cannot.
  • Offer a clear way to reach a person at any point.
  • Write in the languages and the register of the people who will use it, and have speakers check the text. Do not rely on machine translation alone for sensitive information.
  • Let people leave and come back without losing their place.

For users who cannot see well or move easily, a spoken interface can work better than text. Our own MediCal project is an example: patients check a medicine calendar by speaking to Alexa, and a website keeps the calendar in step. See the MediCal case study.

What safeguards should be in place before launch?

Deploying AI in humanitarian work carries risks of its own. The CDAC Network’s briefing note Risks in humanitarian AI sets out the main categories, and its SAFE AI framework treats the participation of affected communities as part of governance. The list below is a checklist to work through before launch.

A defined scope
The chatbot answers about what it was built for and says so when it cannot help
Approved content
Answers come from content your team has written or checked, with a named owner and a date
Limits on a language model
Prompt rules that set what the model may say, and test questions that try to make it go beyond them
Human handover
A clear route to a person for sensitive, urgent or unclear situations, with someone who answers
A safeguarding path
What happens when a user reports harm or danger, and who is told, written down before launch
Data minimisation
The chatbot collects only what it needs, stores it for a stated time and no longer, and explains this to the user
Monitoring
A dashboard of what users ask and how the chatbot answers, reviewed regularly, with the prompts and the content adjusted
A way to switch it off
A named person who can take it offline quickly if something goes wrong

These fit into your wider AI rules. If you do not yet have any, start with our guide to an AI policy for NGOs.

What about data protection?

A chatbot talks to people, so it processes personal data even if you never ask for a name. Phone numbers, messages, locations and the content of a question can all identify someone, and what people ask can be sensitive in itself. Before launch, decide what you collect, why, where it is stored, who can read it and when it is deleted. Check whether the language model service you use keeps or learns from what is sent to it.

For a chatbot that serves vulnerable people, a data protection impact assessment is usually the right step before launch. Ask your data protection officer early. See our guide on whether an NGO needs a DPO.

How do you build one, step by step?

  1. Review the use case. Look at the workflows, the data sources and the people concerned. Write down the ten questions you expect most.
  2. Prototype and let your own team try it. Their comments are the cheapest quality test.
  3. Test with the real users, in their language, on the app they use. Watch where they stop.
  4. Launch with the safeguards above in place, including monitoring.
  5. Improve after launch. Review what users ask and how the chatbot answers, and adjust the prompts and the content with your team.

This is the sequence our AI and interactive tools service follows. A chatbot often feeds a platform, so we can also build the dashboard or the system behind it.

How do you know it works?

Count what matters, and not only the number of conversations. Useful measures are how many people reach an answer without help, how often the chatbot fails to understand, how often users ask for a person, which questions are most frequent, and what users say about it. Review a sample of real conversations every month, with the privacy rules for doing so written down. The questions people ask that you did not expect are your next round of content.

Questions and answers

What is a humanitarian chatbot?

A chatbot that gives people in need information or collects information from them, usually on a messaging app such as WhatsApp or Messenger. Examples are answering questions about rights and services, or running a survey in a chat.

Should a humanitarian chatbot use AI?

Only where it helps. A rule-based chatbot with fixed answers is easier to check and is often the better choice for sensitive information. A language model can be added inside limits you set, with testing and monitoring.

How do you stop a chatbot making up answers?

Keep the model to approved content, set limits through prompt engineering, test with questions designed to push it past them, monitor real conversations, and give a route to a person. For information that must be exact, use a fixed answer.

Which messaging app should we use?

The one your users already use. Ask them. One chatbot can often run on several apps, as the ILO Migrant Recruitment Advisor does on WhatsApp, Facebook, Telegram and Line.

Do we need a DPIA for a chatbot?

Often, yes, where it serves vulnerable people or handles sensitive information. Ask your data protection officer before you launch.

Can a chatbot replace staff?

No. It handles common questions and collects information. Staff handle what needs judgement, sensitive situations and complaints.

Talk to us about a chatbot for your users

We review the use case, prototype it with your team, launch it with safeguards and improve it with you afterwards. Tell us who it is for.

Keep reading

This guide is general information, not legal or technical advice. Reviewed on 11 November 2026 by the team at Relief Applications. The CDAC Network guidance is described from its own public page and may have changed.