For the first time ever, the nonprofit sector has an equal foot in the door to its corporate counterparts when it comes to adopting AI. The social sector is actively having conversations about how AI can serve people, what its limitations are, how to develop use cases grounded in community realities, and how to make it genuinely accessible. At Kaapi, we have been privy to numerous such conversations as a part of our AI Cohort Program, which is designed to help nonprofits adopt AI-powered tools. Having now run two rounds of the cohort, we have realised that the sectoral context is the most primary piece of the puzzle, and current narratives on AI, which are focussed on business use cases, fail to capture that nuance. There is a need for a detailed picture of what is working, what keeps breaking, and what the sector still does not fully understand about its foray into the AI space.
What nonprofits are actually building with AI
A starting point to building AI in the social sector is having a clear idea of the problem being solved. The very first question an organisation needs to ask itself is: “Do we really need AI to solve this challenge?” Across verticals of education, gender, health, or policy, AI usage depends on the communities being served, geography, technological capacities, funding availability, and much more. But we have noticed that despite these differences, a question that is often at the forefront is: can this tool help humans reduce the amount of time this work takes? Given that nonprofits run on thin staffs and even thinner budgets, optimising the workload and time spent is a critical need.
Take Reap Benefit, a Bengaluru-based nonprofit that mobilises young people (called ‘Solve Ninjas’) to tackle local civic problems, including waste, water, and sanitation. Over the years, they have built a database of 47,000 citizen-led actions taken across India. But every new Solve Ninja still depends on a human mentor for guidance, leading to high response times. The organisation is now exploring whether AI can bring that response time to under an hour by using their existing action database to deliver basic guidance instantly, freeing mentors for the cases that need genuine human intervention.
Another organisation, Avanti Fellows, which supports underserved students preparing for competitive exams such as IIT-JEE and NEET, built a tool for student mentors that generated draft feedback summaries. Instead of writing individual feedback for each student from scratch, mentors could edit and send AI-generated responses. When they initially piloted it, the AI was generating a lot of logical inconsistencies and glitches. These fixes took time, and it served as evidence that instead of wanting a perfect tool in three to five months, it is more productive and realistic to build iterations that may take time but account for all user feedback.
We actively discourage a waterfall approach to development, where teams spend months building something and then take it to users for the first time. Instead, nonprofits are encouraged to test the smallest possible version of an idea with real users early on. The aim is to create the simplest proof of concept, send it out, seek feedback, and make improvements iteratively.
What are some learnings?
Through training nonprofits, a few key learnings that have emerged for us:
1. Don’t start from scratch
Every nonprofit building an AI product has to make the same decisions: which model to use, how to handle voice, how to keep the system secure, and how to minimise errors and hallucinations. It is thus beneficial to build interoperable frameworks and shared knowledge. Without these, each organisation is forced to solve the same problems from scratch, burning through their limited time and budget.
When similar needs appear across multiple nonprofits, there is value in building a single, open-source version of the solution that can be used for different organisations in varying contexts. For example, during our first AI cohort, both SNEHA and IPE Global joined with the intention to develop a machine learning (ML) model to predict high-risk pregnancies. They subsequently collaborated on early-stage thinking and problem-solving, before optimising their models for the contexts they work in.
2. WhatsApp chatbots as entry points
WhatsApp has become the default front door for nonprofits looking to adopt AI. WhatsApp chatbots don’t require any new downloads, and work on a trusted and familiar interface. Many nonprofits in our cohort have built their AI products this way, whether for a fieldworker or a student.
Chatbots are able to retrieve information from a curated knowledge set and share it with users. As a result, chatbots emerge as a quick solution in cases where last-mile delivery is a challenge. Users can often also engage with the chatbot in their own language.
3. AI adoption needs strong internal ownership
The nonprofits that moved the fastest were not always the ones with the biggest teams or budgets, but the ones with strong internal ownership. AI projects require continuous iteration, testing, feedback collection, and decision-making. Without a committed internal champion who can work closely with users, mentors, and technical teams, it becomes difficult to move from experimentation to actual adoption.
The challenges that persist
Given the low-resource nature of the sector and the ever-changing landscape of the AI world, the following challenges shape the extent to which nonprofits can engage with AI:
1. Limited tech bandwidth
Nonprofits should strive to build stronger internal ownership of technology and AI initiatives while continuing to leverage external technical expertise where needed. In our experience, the organisations that progressed most effectively were often those where programme teams, leadership, and external tech partners worked closely together with shared ownership of the problem and solution.
This becomes especially important in AI projects where products require continuous iteration to improve accuracy, respond to user feedback, and adapt to changing field conditions. Having internal teams actively involved in decision-making, testing, and feedback loops helped organisations move faster and ensured that tools remained closely aligned with programme realities and user needs.
For example, an education nonprofit in one of our cohorts worked closely with external technical experts to build an AI-based teacher support chatbot, while their internal programme and operations teams regularly tested responses, identified gaps, and prioritised improvements based on classroom feedback. This combination of strong external support and active internal involvement helped the product evolve much more effectively over time.
2. Fast-changing landscape
The AI landscape is rapidly evolving. Features such as retrieval-augmented generation (RAG), which were considered advanced a year ago, are now becoming foundational as newer and more sophisticated features continue to emerge. As a result, tech capacities can rarely be considered “built” in a permanent sense. An organisation may believe it has developed the necessary systems, skills, and workflows to effectively leverage AI, only to find that what meaningful capacity looks like has been redefined. Rather than a one-time investment, building tech capacity needs to be understood as an ongoing process of experimentation and adaptation.
Large language models (LLMs) also update constantly. This means that organisations building tools on top of a particular model may need to continuously adapt, refine, or optimise parts of their systems as models, capabilities, pricing, and performance change over time.
3. Behaviour change
Building technology is often the easier part; the real challenge lies in driving sustained adoption and behaviour change. For a solution to become part of a user’s daily workflow or decision-making process, it must solve a real and immediate problem in a way that feels intuitive, reliable, and valuable. Users do not adopt technology when it is simply innovative, but when it makes their work more convenient or effective.
This is especially important in the social sector, where frontline workers, communities, and organisations already operate under significant time and resource constraints. Even well-designed tools fail if they require additional effort without creating clear value. Successful tech solutions therefore need to centre user behaviour—understanding their habits, incentives, and existing workflows—to build trust.
Driving adoption requires continuous iteration based on real user feedback, strong onboarding and handholding, and designing experiences that naturally fit into how people already work and interact with technology. Over time, meaningful behaviour change happens when users begin to see the technology not as an additional task, but as a genuinely helpful support system that improves outcomes in their everyday work.
For example, an ASHA worker may spend the entire day visiting households, supporting pregnant women, conducting follow-ups, and managing community health needs. At the end of a long day, she may still be expected to manually enter large amounts of data into a complex reporting system. In such a context, an AI-enabled tool that can automatically pre-fill 50 percent of the required information based on voice notes, previous records, or conversational inputs can significantly reduce her workload and time. The value of the technology, in this case, comes not from the sophistication of the AI itself, but from meaningfully reducing friction in her everyday work.
4. Funding
When funding AI, it is very important to have an experimental mindset. In this emergent landscape, the importance of learning what does not work is at par with understanding what does. However, funding models are often built in a way that rewards immediate impact.
There is also a growing expectation that AI will actually compress the time between investment and impact. But behaviour change, adoption, and iteration through real-world feedback cannot be expedited. Having short impact windows for AI projects are in effect setting them up for failure.
AI in the social sector is a long game. While the technology can move fast, the people it is intended to serve, and the people using it, do not. Closing that gap is where the real work lies, and it requires sustained investment and an appetite for failure.
—
Know more:
- Learn how the social sector can navigate the AI overwhelm better.
- Learn how to use AI to advance impact.
- Learn more about the limits of AI in social change.
