When Mani Devi, an Accredited Social Health Activist (ASHA) in rural Rajasthan, saw the underweight infant, she knew something was wrong—but not how serious it might be, or what advice to give.
So she reached for her phone and opened WhatsApp: In Hindi, she typed a question to a new tool called ASHABot: What’s the ideal weight for a baby this age?
The chatbot—trained in Hindi, English, and a hybrid known as Hinglish—responded within seconds: a baby that age should weigh around 4 to 5 kilograms. This one weighed less.
The bot’s answer was clear and specific. It encouraged feeding the baby eight to 10 times a day, and it explained how to counsel the mother without causing alarm.
That, she said, was one of the many encounters with ASHABot that changed the way she does her job.
The tool is part of a quiet but significant shift in public health, one that blends cutting-edge artificial intelligence with on-the-ground realities in some of India’s most underserved communities.
ASHABot, launched in early 2024, is what happens when a generative AI model akin to OpenAI’s ChatGPT or GPT-4 is not only trained on the broader internet, but is connected to a knowledge base containing India’s public health manuals, immunization guidelines, and family planning protocols. It takes voice notes when prompted and provides answers that help the ASHAs serve patients.
Built by the nonprofit Khushi Baby (opens in new tab) using technology developed and open sourced by Microsoft Research, the bot has been transforming how some of the country’s ASHA workers do their jobs. These women are the glue between India’s rural households and the health system, responsible for everything from vaccination records to childbirth counseling. But they receive just 23 days of basic training and often work in settings where doctors are distant, supervisors are overburdened, and even mobile signal is unreliable.
“ASHAs have always been on the front lines,” said Ruchit Nagar, co-founder and CEO of Khushi Baby and a Harvard-trained physician. “But they haven’t always had the tools.”
Nagar’s relationship with ASHAs goes back nearly a decade. In 2015, he launched Khushi Baby with the goal of digitizing health data in underserved communities, often designing tech systems that were locally grounded. The idea of ASHABot emerged in late 2023, during a summit with stakeholders in Rajasthan.
At the time, Khushi Baby was working with Microsoft Research on a separate AI project—one that used eye images to detect anemia. But the buzz around large language models, especially ChatGPT, was rising fast. Nagar and his collaborators began to ask whether this technology could help ASHAs, who often lacked real-time access to quality, understandable, medically sound guidance.
“ASHAs were already using WhatsApp and YouTube. We saw an inflection point, new digital users ready for something more,” said Nagar, now a resident at the Yale School of Medicine in New Haven, Conn.
So they began building.
Microsoft researcher Pragnya Ramjee joined the project around that time, leaving a design job at a hedge fund to focus on technology with social impact. With a background in human-centered design, she helped lead the qualitative research, interviewing ASHAs in Rajasthan alongside a trained translator.
“It made a huge difference that the translator and I were women,” she said. “The ASHAs felt more comfortable being open with us, especially about sensitive issues like contraception or gender-based violence.”
Ramjee and the team helped fine-tune the system in collaboration with doctors and public health experts. The model, based on GPT-4, was trained to be highly accurate. When it receives a question, it consults a carefully curated database—around 40 documents from the Indian government, UNICEF, and other health bodies. If the bot doesn’t find a clear answer, it doesn’t guess. Instead, it forwards the question to a small group of nurses, whose responses are then synthesized by the model and returned to the ASHA within hours.
The goal, Ramjee said, is to ensure the bot always stays grounded in reality and in the real training ASHAs receive.
So far, more than 24,000 messages have been sent through the system and 869 ASHAs have been onboarded. Some workers have used it only once or twice. Others send up to 20 messages in a single day. Topics range from the expected—childhood immunization schedules, breastfeeding best practices—to the unexpected.
“They’re asking about contraception, about child marriage, about what to do if there’s a fight in the family,” Ramjee said. “These aren’t just medical questions. They’re social questions.”
One woman came to Mani Devi saying she’d missed her period for two months but wasn’t pregnant. The bot provided Devi with information that gave her the confidence to assure the patient she had nothing to worry about.
The responses come in both text and voice note, the latter often played aloud by ASHAs for the patient to hear. In some cases, voice responses about long-acting contraception help persuade hesitant women to begin treatment.
There is no question the technology works. But the team is quick to emphasize that it doesn’t replace human knowledge. Instead, it amplifies it. ASHABot illustrates how LLM-powered chatbots can help bridge the information gap for people, particularly those with limited access to formal training and technology, said Mohit Jain, principal researcher at Microsoft Research India.
“There is a lot of debate about whether LLMs are a boon or a bane,” Jain said. “I believe it’s up to us to design and deploy them responsibly, in ways that unlock their potential for real societal benefit. ASHABot is one example of how that’s possible.”
– Mohit Jain, Principal Researcher, Microsoft Research India
Of course, the chatbot isn’t perfect. Some users still prefer to call people they know, and the big question of scaling remains. The team is exploring personalization options, multimodal support like image inputs, and parallel LLM agents to ensure quality assurance at scale.
Still, the vision is expansive. As of now, ASHABot is only used in Udaipur, one of the 50 districts in Rajasthan. The long-term goal is to bring ASHABot to all one million ASHAs across the country, who take care of about 800 to 900 million people in rural India. The potential ripple effect across maternal health, vaccination, and disease surveillance is immense.
Nagar, who has traveled to India twice yearly for the last 10 years to research the needs of ASHAs, said there are still “many things yet to explore, and many big questions to answer.”
For ASHAs like Mani Devi, the shift is already real. She says she feels more informed, more confident. She can talk about previously taboo subjects, because the bot helps her break the silence.
“Overall, I can give better information to people who need help,” she said. “I can ask it anything.”