Context

In October 2022, TinkerLabs was approached by UNFPA to conceptualise and design an AI-enabled chatbot, with the intention of improving sexual health awareness amongst adolescents in low-resource regions of India.

As part of the 4-member core team that drove this project, I helped design Just Ask / Khulke Poocho — a WhatsApp chatbot that is currently used by more than 25, 000 adolescents and adults in regions of Madhya Pradesh, India.

Most of the work is under an NDA and, hence, I can only share a brief overview of the process we followed.

Research

The goal of our research, both primary & secondary, was to identify information-seeking patterns amongst Indian adolescents in low-resource settings. By the end, we had a list of factors that triggered the search for information, sources that the adolescents were going to (human or otherwise) and the outcomes of those interactions. Through a behavioural deep-dive, we identified 8 emotional needs that must be met with the proposed solution.

Product & Testing

There are 3 distinct areas to tackle when designing a chatbot — identity (whom or what does it look like), tone (whom or how does it talk like) and features (what all can it do). Through multiple rounds of fighting between desirability (what best meets the needs) and feasibility (what all can be done within the technical limitations of WhatsApp), we proposed a total of 173 ideas (features + tone + identity) that were to be tested with the target audience.

I primarily worked on the conversation flows, trying to figure out how each proposed feature would work. The goal was to design a system that could allow testing to be done in an environment as close as possible to real world situations. Since we couldn’t actually develop an autonomous chatbot, we decided to use a Wizard of Oz system where a comprehensive flow would guide a person from our team to ‘act’ like a chatbot, as we tested with participants who believed they were texting with a ‘robot’.

Since this was meant to be an AI-enabled product, I had to account for use-cases of open-ended texts that a person could send over the chat. This was done by incorporating a simple system of variables and constants in the conversation to allow the tester to flow through the conversation without interpreting the message and trying to come up with a response on their own.

Outcome

After we tested the flows and the identity (using a structure similar to a Semantic Differential Survey) separately, we got a good sense of the things that would work well on the field. Aafter rounds of discussions and scoping down our ideas along with the other stakeholders, UNFPA publicly released the Just Ask Chatbot.

Within 2 months of launching the chatbot, it had more than 25,000+ unique adolescents that interacted with it.

Acknowledgements

Team: Devanshi Sihare, Tanwi Mirajkar, Pratyush Pillai, Sanjana Aggarwal and Sakina Attarwala (all from TinkerLabs).