Three ways India is using AI for development

You can listen to this podcast on Spotify, Apple Podcasts, or wherever else you get your podcasts. You can also watch this conversation on YouTube.

Organisations in India are hard at work trying to turn the promise of AI into better outcomes across a range of different settings, including schools, courtrooms and farmers fields.

In this episode of Ideas in Development, Utkarsh Saxena, Claire Cullen and Niriksha Shetty discuss how their organisations, Adalat AI, Youth Impact, and Precision Development, are already deploying AI in India and what they’ve learned during the process.

So far in our series of podcasts on AI, we’ve talked a lot about the big picture. Whether the industrial revolution is a good comparison for AI, the work needed to make Africa AI-ready, and how AI could upend labour markets. But what about here and now?

One shared lesson stood out to me from talking to Utkarsh, Claire and Niriksha – the AI model is rarely the hardest part. In India’s courts, classrooms and agriculture, what really seems to determine whether an intervention gets of the ground is the organisational plumbing around it. This includes the nitty gritty implementation questions of trust-building, workflows, incentives, training, and language involved in partnering with public systems in India.

“I’d assumed that the model was the hard part, but actually I think it’s a lot of the like human, pipeline, infrastructure, relationships part that’s hard”. Claire Cullen

“Technology has only been one part of the problem. And it’s often not the hardest part of the problem”. Niriksha Shetty

AI for development on the ground

It’s tempting to think that AI for development is simply a question of organisations accessing the best model. But experiences in India repeatedly highlight that progress depends on the systems that surround the model.

In public institutions, shipping an AI tool is far more than just releasing an app. If you’re trying to change routine behaviour in settings where time is scarce, capacity varies widely, and mistakes can be costly, you need to meet users where they are. This means integrating with existing processes, and patiently building relationships with the people who will actually use the tool.

Utkarsh Saxena describes Adalat AI’s approach as twin engine. One engine is the technology, and the other is the partnerships and implementation work that make the technology usable day-to-day.

“Adalat AI is a twin engine venture. Only one engine is all this AI… The other engine of the venture is programming and partnerships.”

Can AI help to fix India’s broken justice system?

The episode begins in India’s courtrooms with Utkarsh Saxena who co-founded Adalat AI, a legal tech social venture focused on reducing delays and backlogs.

“There are 50 million cases stuck in courts, it’s going to take us 300 years to clear the backlogs per some estimates”. Utkarsh Saxena

This has direct impacts on economic development. Slow moving court systems undermine contract enforcement and property rights. It also has implications for human rights.

“India’s prison population comprises of 70 to 80% under trials who are not in jail serving a sentence, but just trapped there waiting for their bail to be heard”. Utkarsh Saxena

To combat this, Adalat AI’s approach is tackling the administrative bottlenecks that grind courts down. India’s courts, like many across the Global South, remain paper-based, which causes enormous frictions. Searching for documents and recording evidence consume huge amounts of time that judges and staff simply don’t have.

One vivid example is the shortage of stenographers in India, which results in judges having to handwrite proceedings themselves, turning judges into scribes and slowing cases dramatically. To fix this, part of Adalat AI’s work is building an AI-powered legal transcription tool, built to understand legal jargon and work across multiple Indian languages in real time. The aim is to speed up the production of the written record so courts can process more cases, more reliably.

Since courts are a sovereign, highly confidential setting, they can’t simply pipe proceedings into third-party APIs like GPT, Gemini, or Claude. Adalat AI’s security architecture rests on three pillars: data localised in India, no third-party APIs, and an encryption framework where keys are decentralised across local court machines. This requires fine-tuning models, annotating datasets with judges and court staff, and building language-specific capabilities (including legal phrases in languages like Tamil): these steps require time and money

“It’s almost 15 to 20 times more expensive to house your own models and to kind of run them on your GPUs because GPUs are expensive” Utkarsh Saxena.

This has important implications for government capability and scale. Utkarsh argues that just as governments built IT teams after the dot-com boom, they need AI teams and compute capacity after the AI boom: data centres, GPUs, and the operational ability to run models internally. For Adalat AI, that affects their strategy: they’re not only building tools, but also thinking about making it easy to hand over their work so Indian courts can take it forward at scale.

AI for education in India: Using AI to scale an intervention we already know works

Too many children finish primary school without being able to read or do basic numeracy. Youth Impact is an evidence-driven NGO working across education and health, and Claire Cullen joined us to discuss how they are using AI.

Youth Impact’s core educational approach involves targeting instruction to children’s level, which evidence shows is an extremely effective intervention. This involves identifying whether a child is at letter, word, or paragraph level (or addition versus multiplication level), then teaching them accordingly at the level they are at. So in India, they’re working with the state of Karnataka, which is keen to scale proven interventions, but where teacher time remains a major bottleneck.

So Youth Impact is developing a voice-to-voice AI bot that can run quick diagnostic assessments. The AI assesses where a child is (e.g. addition vs division) and shares that information with teachers, enabling more targeted instruction.

The project is still at an early stage. Claire notes that it’s working well with older children in grades four and five, but the model struggles more with younger children’s voices In response, they are collecting more voice data for training, and in the interim, caregivers are asked to repeat answers so the system can pick up their adult voice.

I found Youth Impact’s thinking on how ‘future-proof’ this intervention is interesting. Claire believes that even if better models arrive, much of their work including building government relationships, household trust, curriculum alignment, language compatibility (Kannada), and the operational system that makes phone-based tutoring actually work, is model-agnostic. The tech can be swapped as model’s improve, but the human pipeline takes time.

AI in agriculture: Personalised advice at scale

“More than 2 billion people live in smallholder farming households” Niriksha Shetty

In the final part of this episode, I talked with Niriksha Shetty, CEO of Precision Development (PxD), which works to connect smallholder farmers to practical innovations.

PxD acts as a bridge between farmers and useful innovations, delivering better seeds, better forecasts, and better finance in ways farmers can actually use. A flagship example is a voice-based advisory service built with the state government of Odisha, reaching around seven million farmers with actionable guidance via mobile phones.

AI enters PxD’s work in two main ways:

1. Better prediction and relevance. Niriksha highlights that forecasting critical events like the onset of the monsoon with longer lead times is crucial for decisions about what to grow, when to plant, whether to borrow, and how much labour to invest. They have worked with several partners to disseminate AI-based monsoon onset forecasts to 38 million farmers across India.
2. Cheaper, scalable personalisation. PxD is using GenAI to generate advice that’s localised and personalised. Their ambition is to produce localised messages at scale tailored to specific constraints and engagement patterns.

Early testing surfaced issues with accent quality and numbers – a serious risk when advising fertiliser quantities. The response has been to put guardrails in place and keep humans in the loop.

Like our other guests, PxD isn’t anchored to a single model, and most of the important work involves building systems for testing, A/B experimentation, learning, and delivery partnerships, which will allow them to scale as the tech improves.

What these three India case studies teach us about scaling AI for development

Taken together, these three profiles offer a useful set of takeaways for anyone thinking about AI for development programmes:

• Start by identifying a binding constraint.
• Localisation: Basics like language, accents, curricula, and context-specific norms are all potential barriers to navigate.
• Security and governance: In sovereign settings like courts, confidentiality requirements impact on unit economics, infrastructure needs, and deployment models.
• Evaluation and iteration: All three organisations treat learning as continuous, through RCTs, pilots, A/B tests, and rapid feedback loops.
• Partnerships are the product. Whether it’s courts, state governments, or farmer networks, scale comes from institutions which means building relationships is crucial.

“Actually a lot more work needs to be done on building those partnerships with the actual stakeholders… meeting them where they are at so that the technologies are the right level and is actually internalized in the workflows.” Utkarsh Saxena

That AI model’s are increasingly capable is great for potential applications in development settings. But this episode shows that the ultimate development impact of AI still critically depends on implementation.