The tech
A chatbot lets people use software by having a conversation with it, by text or voice.
- Traditional chatbots follow predefined rules or decision trees. They offer menus and give pre-written answers. They're predictable, but can only handle the questions their designers anticipated.
- AI chatbots use a language model to generate a response to whatever the person has written. Many can also look up documents or search the web before answering.
Many real services combine the two: scripted flows for common requests, with AI handling open questions or helping human staff respond.
Why it matters
- Reaching people on channels they already use. Chatbots can run on SMS, WhatsApp and other messaging apps, so people don't need a new app or a fast connection to get information.
- Supporting frontline services. In health, chatbots can send reminders, answer routine questions, help with initial triage and collect feedback on care, freeing up staff for cases that need them.
- Access to public information. Citizens can ask about government services, report problems and give feedback, in their own words and often in their own language.
- Scale at low cost. A chatbot can answer many people at once, at any hour. That makes it practical for high volumes of routine questions that would overwhelm a small team.
- The risks scale too. An AI chatbot can give a confident wrong answer to thousands of people. Services in health, finance or legal advice need tested content, clear limits on what the bot will answer, and a route to a human when the question is serious.
In the real world
- MomConnect, run by South Africa's National Department of Health, registers pregnant women by mobile phone and sends them weekly messages matched to their stage of pregnancy. Mothers can also ask questions and give feedback on the quality of care they receive at clinics.
- During the Covid-19 pandemic, the World Health Organization launched its Health Alert chatbot on WhatsApp, followed by Facebook Messenger and Viber. Built with Praekelt.org and offered in several languages, it shared accurate health guidance and reached more than 12 million people on WhatsApp within its first month.
- Newer services use language models to answer open questions. Jacaranda Health's PROMPTS service in Kenya now uses its own Swahili-focused model to reply to mothers by SMS, with nurses stepping in when a risk is identified (see Open weight models).
Before building one, test answers with the people who will use it, measure how often it's wrong, and decide in advance which questions it should hand to a person.
All ExplAIners