Responsible AI is a practice. It is built, tested, broken, and rebuilt through every decision an organisation makes as it moves from idea to deployment to scale.

The panel hosted by Digital Futures Lab at the AI & Equality Festival of Ideas brought that practice into full view, through the lived experience of five organisations building AI tools in India’s social impact sector.

Moderated by Sasha John, the session gathered voices from ARMMAN, Adalat AI, Quest Alliance, Tattle Civic Technologies, and ARTPARK. Together they represent some of the most active and complex ground for AI deployment globally, working in maternal health, legal aid, youth education, and civic technology, across geographies, languages, and vastly different levels of digital infrastructure.

What emerged across every account was a shared understanding earned through iteration: responsible AI requires staying close to users, close to evidence, and willing to be surprised by both. Amrita Mahale from ARMMAN described how their ANM knowledge assistant, now serving over 8,000 frontline health workers across rural India, took a year and a half of experimentation before it was ready to scale. The knowledge base had to be clinically validated. Early escalation messages, designed as guardrails against out-of-scope queries, had to be rewritten entirely when health workers began reading them as personal failures. Thirty percent of early queries fell outside the intended scope, a signal that the tool had been designed around a use case that did not fully match the reality of an ANM’s working day.

Tanisha Kedia from Quest Alliance offered a detailed account of what responsible design requires before a single line of code is written. Quest ran a trust mapping workshop with learners from across India to understand the emotional and psychological landscape their users were navigating. What came back were questions about comforting grieving friends, navigating family pressure, making impossible choices between education and work. The chatbot Quest was building would either compound the pressures these young people were carrying or help ease them. Understanding that required going to the source.

Astha Khurana from Adalat AI described the work of introducing cutting-edge transcription technology into courtrooms that still run largely on paper and pen. Co-designing with lawyers as proxy users, and on-ground training that accounts for the physical realities of where a microphone is placed relative to a judge, has been as important as any technical decision the team has made.

From an ecosystem perspective, Jigar Doshi of ARTPARK and Dr. Baarish Aggarwal of Tattle Civic Technologies offered a candid assessment of where the sector stands. Open-source tools and safety guardrails exist. The friction lies in adoption, in the bandwidth required to use these tools consistently, and in the structural conditions under which social sector organisations operate: short funding cycles, limited in-house technical expertise, and pressure to deploy faster than the evidence supports. Tattle’s work on customisable safety guardrails for the Indian context, including the specifically high-stakes problem of caste-based personally identifiable information leakage, points toward what grounded, contextual technical support for this ecosystem can look like.

The panel closed on a note that felt both honest and galvanising. The unknown unknowns will always exist. Responsible AI does not mean eliminating uncertainty. It means building the organisational practices, the user relationships, and the evaluation infrastructure to keep learning from it. The organisations on this panel showed what that looks like in practice, and why there is no shortcut to getting there.