India’s courts faced a backlog of more than 50 million pending cases in 2025. In government schools across the nation, teachers spend months collecting foundational literacy and numeracy data. And for millions of citizens, the laws that govern their daily lives remain inaccessible, written in language they cannot parse, and delivered through channels they cannot reach. These gaps are endemic to systems that are under-resourced and understaffed.
Within this landscape, a few nonprofits are working to introduce artificial intelligence (AI) to streamline these governance processes. Adalat AI, for instance, works with the judicial system to build infrastructure to power paperless courts through automated transcription, digital records, and workflow management. Vowels of the People Association (VOPA) has built an AI-powered foundational literacy assessment tool now used in government schools in Maharashtra under the NIPUN mission. Civis, a Mumbai-based nonprofit, is leveraging AI to make public consultations more accessible and participatory, and has partnered with government departments to synthesise public feedback on laws and policies.
But in addition to building the right products, it is very important to ensure that these models can be trusted, are affordable, and can function in public systems that are already stretched. While each works in a different domain of governance, they have learned important lessons about making AI work within public institutions.
Identifying the right problem
Before any of these three organisations could build anything useful, they had to understand what was actually ‘broken’, from the perspective of the people inside the system bearing the weight of it.
In Adalat AI’s case, the insight came from having practiced law in India’s district courts. Co-founder Utkarsh Saxena had witnessed firsthand what he describes as a logistics problem rather than a law problem. “Justice is not just a question of law,” he says. “It’s actually a question of logistics.” While the rest of the world has moved to Google Calendars and Microsoft Outlook, everything is scheduled in diaries in Indian courts. And because of this system, sometimes summons don’t reach witnesses. Judges also spend a long time writing proceedings by hand. Adalat AI set out to fix this with an secure, end-to-end paperless court infrastructure that allows courts to operate without paper, from the moment a case is filed to the issuance of the final order. Remarking on what helped build trust in the tool, Utkarsh emphasises the distinction between a ‘painkiller’ as opposed to a ‘multivitamin’. “When you’re building credibility, painkillers are a more effective way to get buy-in,” he says. Adalat AI could have built a case management system or paperless filing tools, but those would have been multivitamins—features that are great to have, but are likely to fail in a landscape already riddled with a dozen other logistical and institutional problems. Transcription was a central pain point, and solving it first gave them a strong foundation to build on.
India’s pre-legislative consultation policy requires the government to publish draft policies for public comment, but the people most affected by those policies have historically been the least able to engage with them.
VOPA arrived at a similar logic through a different door. In districts across Maharashtra, teachers were assessing foundational literacy and submitting data upward through paper sheets, Google Forms, and chatbot-based tools that Prafulla Shashikant, who leads the organisation, describes as “inhuman and joyless.” It also introduced a lot of subjectivity in the data collection method. One aspirational district that ranked near the bottom on most development indicators, for instance, consistently appeared near the top of state assessment results. It was obvious that the data had inaccuracies; however, the lack of a shared method, objective tool, or real-time verification resulted in the cycle simply repeating year after year. By the time the data discrepancies were identified, the bureaucracy which was meant to help students had already missed the window for action and students had been promoted to the next grade. “The system never had the real-time data to intervene while it still mattered,” Prafulla says. Through AI, VOPA was able to build a standardised assessment method that removed subjectivity and could operate at state scale.
For Civis, the chief concerns were access and voice. India’s pre-legislative consultation policy requires the government to publish draft policies for public comment, but the people most affected by those policies have historically been the least able to engage with them. The language is inaccessible, the process is unfamiliar, and the infrastructure for participation leaves most people out. Civis used AI to change that equation, translating complex legislative language into easy-to-follow text, creating multilingual access points, and enabling voice-based participation that allows people to respond in their own language through a phone call or voice note. As Antaraa Vasudev—Civis’s founder—notes, “For people who have been largely excluded from policy conversations, being able to answer questions and offer feedback in voice has been transformative.”
What connects all three organisations is that each of them spent time inside the system, understood the actual friction, and only then designed their respective solutions.
What it takes to work with the government
1. The data question
Data, in all three cases, is simultaneously a technical matter, a trust matter, and a political one. It is the foundation on which everything else is built.
For Adalat AI, this was particularly imperative. Court proceedings are sensitive, and the functional requirement from the judiciary was unambiguous: Data could not leave the nation’s sovereign borders. Adalat AI responded by building its models entirely in-house and working only with vendors that maintain Indian servers. They also deployed an encryption system to enhance trust. When a judge creates an account, they receive an encryption key generated on their own device. Even Adalat AI’s own team cannot decrypt this data. “The idea of a trust-less system where everything is public and yet no data gets revealed, was the foundational piece of how we were thinking about security,” says Utkarsh.
This level of protection created an additional constraint: Adalat AI trains its models with zero access to actual user data. Its entire feedback and model improvement pipeline has been built around this reality. The team collects qualitative signals from state leads, identifies where the model is failing, sources publicly available legal texts, works with in-house linguists, and uses all of this information to generate training data. Given the data-sensitive contexts it operates in, the aforementioned guardrails have helped the organisation build and preserve trust.
AI that bypasses government data produces a nice demo that never scales, but AI that embeds inside the existing system can actually reach every school.
Prafulla describes data quality and decentralised access to data as the heart of VOPA’s project. To make assessment data trustworthy, the organisation built in multiple verification layers: supervisor reassessments of a sample of students, independent audio cross-checks, and longitudinal tracking of the same students across assessment cycles. One result of this system has been the ability to catch malpractice that the earlier paper-based process had no way of detecting. In one district, supervisor reassessments showed only a 62.4 percent match with teacher assessments. When VOPA pulled the underlying audio to investigate, they found that in roughly 72 percent of those flagged cases, the child assessed by the teacher and the child assessed by the supervisor were simply different children. But with the implementation of their tool, teacher and supervisor levels matched exactly in 89 percent of cases.
VOPA also made a deliberate architectural choice to build on existing government data infrastructure, and not alongside it. Rather than creating its own student database, VOPA integrated with SARAL for student records, UDISE+ for school identifiers, and SHALARTH for administrative hierarchies. Explaining the rationale driving this decision, Prafulla remarks, “AI that bypasses government data produces a nice demo that never scales, but AI that embeds inside the existing system can actually reach every school.”
Speaking of Civis’s work with the government, Antaraa identifies two distinct anxieties that arose: Whether the data the AI draws on is verified, restricted, and limited to approved government sources rather than the open internet, and whether the outputs it produces are accurate and comprehensive. Satisfying both, she says, is what converts a sceptical government partner into a confident one.
Building trust in the data is one of the most important puzzles to solve when building for the government.
2. Human in the loop
According to Antaraa, AI should never be tasked with doing what a human hasn’t done a thousand times before. She says, “The only way you can create a good quality citizen-friendly summary of a law is if you’ve done it countless times before and then feed the logic to the AI to do it at scale.”
“Human oversight never disappears, it evolves,” says Prafulla. AI surfaces a student’s learning gap; the teacher decides how to respond; a district officer decides where to allocate resources. The data is only as valuable as the human action it triggers. “We are not using AI to replace ground stakeholders,” Prafulla says. “India is a long way from that, and frankly it should be.” Without human oversight, even the smartest system can hamper meaningful change by losing sight of ground realities.
At Adalat AI, human oversight is built directly into the user interface. For example, when a transcription includes a reference to a legal case, the system automatically looks up the citation and inserts it, but marks it in red. The judge cannot download or use the document until they have manually verified the citation. “No system in the world is 100 percent accurate,” says Arghya Bhattacharya, the organisation’s co-founder and CTO. “We built guardrails to make sure that people can’t make mistakes, even if there is a natural tendency to over-rely on AI.”
But human oversight is not only about catching errors. At Civis, when using AI to draft citizen-friendly summaries of laws and policies, it also calls for editorial judgement on whether the summary is logical and clear, what example will make the issue feel real, and what question will prompt participation.
3. Building an AI product versus embedding AI in government
Embedding AI in government goes beyond just building a product. All three organisations draw a sharp line between these two activities. The former involves aligning with government data systems, building trust, managing workflows and costs, and creating structures to ensure continuous support.
According to Prafulla, “The AI is perhaps ten percent of the work. The other 90 percent is trust, training, and institutional alignment.” A public institution comes with constraints and it is accountable to millions of people who cannot afford for the system to fail them.
Adalat AI cautions against the instinct to build one product that fits all. Change inside a government institution is a slow process. The organisation initially only built its speech-to-text tool, and getting adoption for that took significant effort. But having demonstrated they can solve one real problem well, courts are now actively requesting them to build more features. “You meet them on their end first,” Utkarsh says, “and then slowly up level the entire setting.” This shift in appetite has been one of the more encouraging signs of progress. Partners are now asking for tools they would have shown no interest in eighteen months earlier.
VOPA makes a parallel point about how AI should appear to end users. Prafulla says, “The moment they see the word ‘AI’, half the room mentally checks out.” The goal is to give a teacher the information they need about which children in her class are unable to read at the required level. Whether AI was involved in surfacing that information is a detail she should never have to think about. “The embedding should be so smooth that the system itself is essentially unaffected,” he adds.
In Civis’s experience, two practical constraints shape what embedding actually looks like in government systems. The first is cost. Once an AI tool is deployed at scale, it becomes a recurring infrastructure cost. “You don’t want to have to pay a month’s rent of a physical office in your AI bills as well,” says Antaraa. Tools that work as demonstrations but are unaffordable at deployment scale never become infrastructure. The second is alignment with government’s own AI guidelines on open-source requirements, local deployment, and procurement norms. Nonprofits entering this space need to understand those guidelines before building, and should resist the pressure to keep expanding scope. “While you’re going to get asked to do all sorts of things and build all sorts of products,” she says, “make sure that you know your niche and you’re able to double down on that.”
4. User experience as a condition for the work
All three organisations treat user experience as the determining factor in whether a system works at all, which is something to be designed for rigorously rather than added afterward.
For VOPA, the single biggest lesson at scale was language. For teachers and administrators working in Maharashtra, Marathi is a necessity that determines whether a tool gets used at all. VOPA also sent its engineers into classrooms to personally conduct assessments for hundreds of children, and that fieldwork directly improved the product. The organisation also designed the tool with the assumption that users would receive minimal training and should be able to use it as intuitively as they use WhatsApp. Rather than relying on dedicated support teams, they built self-help features and peer-support mechanisms directly into the application.
In Civis’s experience, user experience also extends to the design of participation itself. Voice-based engagement, offered in Indian languages, is imperative to ensure that people find it easy to participate. The Interactive Voice Response format means people can give feedback with a phone call alone. For one of its recent labour consultations, including voice features led to broader participation. Overall interactions increased by 38 percent and first-time participants grew by 27 percent.
The design principle across all three organisations is consistent: The tool has to feel less demanding than the problem it solves.
The challenges that remain
Working with the government is slow, and the pace is structural rather than incidental.
For VOPA, one of the hardest constraints has been institutional decision-making. Approvals take time. Within the public education system, they have to liaise with many stakeholders at once: the State Council of Educational Research and Training, the Education Commissioner, the Education Secretary, and multiple other officials—all of whom are subject to frequent transfers. Each new officer arrives with fresh ideas and requires re-briefing. Prafulla describes this as continuous work, a baseline condition rather than an exception.
At Adalat AI, sustaining a serious in-house technology team inside a nonprofit structure is its own ongoing challenge. “AI engineers are hot commodities in the labour market today,” Utkarsh says. Convincing them to take a different path requires a pitch built around mission, the significance of the problem, and the quality of the technology they will get to build. He describes this as something he underestimated at the outset. Funding norms in the social sector further compound this challenge. Nonprofits are competing for the same talent as private technology companies, but funders are often hesitant to support salaries that reflect market rates for senior technical talent. As a result, building and retaining the specialised teams needed to develop and maintain AI systems can become a constraint in itself.
The infrastructure constraints at Adalat AI are equally concrete. Courts currently lack the ability to procure graphics processing units (GPUs). Hardware also varies enormously across courts, from high-spec touchscreen laptops to rooms with loud fans and no microphones. Adalat AI built its system to work at the lowest common denominator, designing models that function without external microphones and training them on data from real, noisy courtrooms rather than clean studio recordings.
For Civis, the question goes beyond how to respond to government demand to also include how to deliberately build around these demands. As government departments become more familiar with technical capabilities, new requests emerge, right from document analysis to other administrative use cases. These help open doors and build trust, but they also require a clear strategic filter on whether doing so strengthens its core mandate or not. In this sense, the challenge is not simply to resist expansion, but to expand with intention. Navigating all of these challenges require an understanding of the system from within—the institutional culture, political pressures, data constraints, and human beings who will use the tool every day. A clear problem needs to be identified and studied thoroughly before any solution is even thought of. And when starting to work on that solution, organisations must ask themselves a few critical questions:
- Can this solution be integrated with existing government data and workflows?
- Who verifies the AI output?
- What happens when officials change or are transferred?
- Can the government afford to implement it at scale?
- Who is in charge of the data, model, and long-term maintenance?
Ultimately, the gap between a good AI tool and one that becomes public infrastructure is almost never about the technology itself. It is about whether the people who built it recognise that deploying such technology within a public institution is a long-term commitment to a specific system with specific constraints.
—
Know more
- Read the following case studies on Adalat AI, Civis, and VOPA on the Digital Toolbook for Social Impact platform to better understand how they leverage AI in their work.
- Read this article, which captures insights from and barriers to building and deploying AI on the ground.
