Why AI needs digital public infrastructure to deliver for citizens

A mother arrives at a rural health centre with a child who has a high fever and a rash. The nurse needs to check the child’s medical history, confirm eligibility for the government’s cashless programme and decide on treatment quickly. Too often, this still means long queues, paper folders and, worse, mistreatment.

With the right digital foundations, it can feel different. Identity can be verified in seconds. Records can be retrieved with consent. Payments can be made instantly. And a simple, multilingual AI ‘agent’ can guide the nurse through the key steps.

This is the practical promise at the intersection of digital public infrastructure (DPI) and artificial intelligence (AI).

Over the past decade, many countries have invested in DPI: shared digital building blocks that enable public and private providers to deliver services securely at population scale.

In practice, DPI often includes digital identity, instant payments, trusted registries and consent-based data exchange mechanisms that let information move safely across agencies and providers. Examples range from India’s Aadhaar, Unified Payments Interface (UPI) and DigiLocker to Estonia’s X-Road and Brazil’s Pix – different designs shaped by context, but united by interoperability, trust and inclusion.

A second change is now unfolding. AI is moving from experimentation to everyday use, helping us search for information faster, summarise cases, detect patterns and support decisions. Governments are exploring how to use AI to improve service delivery across health, education, skilling and a range of citizen-centric services.

The Global DPI Summit, held in Cape Town in November 2025, and the recently concluded AI Impact Summit, held in New Delhi in February 2026, brought renewed focus on the promise and potential of these two technology paradigms.

Yet the most consequential opportunity does not lie in deploying digital public infrastructure or AI in isolation. It lies in their convergence.

On one side is AI for DPI: using AI to make DPI-enabled services more accessible, more responsive and easier to run. On the other side is DPI for AI: using DPI principles to create the foundations that make AI safer, more locally relevant and more inclusive. Each reinforces the other.

And yet, AI×DPI has not "lifted off” at scale because the convergence only works when AI can run through trusted rails. In many countries, the connective tissue is still weak: data and workflows remain fragmented across agencies, so AI cannot reliably draw on them. And too often, AI is funded as pilot projects rather than treated as shared infrastructure: without reusable components, a common operating model and clear ownership for risk and outcomes, promising demos don’t scale.

Today, AI capability is moving fast but AI impact is moving much more slowly. We celebrate smarter model releases almost every month – but how does that help a farmer, a sick child or a student in a rural district looking for guidance?

AI for DPI is about scaling AI to population scale – improving outcomes and experience on top of the rails that already exist. It shows up in two practical ways.

First, there are cross-cutting AI capabilities that can improve public services that rely on DPIs – identity, payments or data exchange-based services.

For example, India’s Bhashini initiative provides AI-enabled multilingual and voice support and integrates with key national rails (Aadhaar, UPI and DigiLocker), illustrating how language AI can expand the reach and usability of core DPIs.

Similarly, the African Next Voices project has recorded 9,000 hours of speech across 18 languages in Nigeria, Kenya and South Africa, creating AI-ready datasets. Applying DPI principles to language data – common standards, interoperable formats, clear metadata and governed sharing – makes it possible to bring together disparate speech datasets across institutions and countries into reusable, “exchange-ready” data assets.

Evidence is also emerging on how AI can augment – not replace – teachers. A randomised evaluation in Brazil of an AI-powered essay grader, designed to give immediate feedback, showed students improved through tailored feedback, while teachers could spend more time on 1:1 instruction. DPI can enable this to scale in public systems through interoperable learner records and common learning standards, consent-based data exchange, and the ability to write back progress into trusted registries so support is targeted and trackable.

DPI for AI is harder, and less visible, but just as important. It provides the enabling foundations that AI needs in a public context, with three foundations key.

A DPI approach – for example, data exchange based on open APIs – can reduce friction in obtaining context-specific local datasets for public purposes, creating high-quality, “AI-ready” data that reflects local linguistic, cultural and demographic realities. This matters in many low- and middle-income country settings where local data can be limited.

Moreover, Harvard University research shows that many widely used AI models exhibit cultural skew, reflecting the values of a narrow subset of the global population. When benchmarked against survey data from 65 countries, the values and reasoning patterns of leading AI models cluster with those of Western democracies — a poor fit for the diverse populations most DPI systems are designed to serve.

In the absence of strong local data foundations, deploying such models risks importing misaligned assumptions into public services, education and governance, potentially leading to biased decisions in benefits, credit or care. Initiatives such as India’s AIKosh are using a DPI approach to source local contextual datasets for training AI models that support local use cases.

A core element of the DPI approach is federation, interoperability and shareability. The same idea can apply to compute in AI: shared compute platforms, sandboxes and model repositories can enable innovators and public institutions to access affordable, scalable compute without centralising control. India’s Open Cloud Compute (OCC) and the provisioning of graphics processing units for startups and research institutions under the IndiaAI mission are examples of a DPI-aligned approach to compute infrastructure.

DPI is trusted because principles such as privacy and security by design, interoperability and clear accountability are built into the system. The same principles are key in AI usage. Consent, transparency, auditability and accountability need to be built into how data is accessed and how models are trained and deployed. This enables AI systems that are safe, inclusive and trusted at national scale, while equipping governments with tools to guide and oversee AI development.

Together, these three foundations position DPI as the backbone of public-centric AI – enabling AI systems that are representative, resilient and aligned with public values.

For public leaders, the convergence of DPI and AI presents an inflection point – a fundamental shift in how states can deliver services and create public value. Here is what public sector leaders should focus on.

What must be different this time is orchestration. Earlier waves produced siloed apps and one-off AI pilots that struggled to scale. DPI×AI will deliver only if it is treated as a concerted, cross-ministry state-capacity agenda – with clear priority outcomes, clear ownership for operations and risk, and shared building blocks that can be reused across services.

Start where DPI already creates trust – and use AI to remove everyday friction. Choose a few high-volume user journeys that already use DPI building blocks, such as identity, payments, trusted documents, registries or permissioned data sharing. Apply AI in practical ways that make those rails easier for citizens and frontline staff to use.

In primary healthcare, for example, verified identity and consented access to medical records can reduce repeated questions; AI can leverage these DPIs to find relevant information quickly and complete processes faster and more accurately.

Build a small set of shared AI capabilities on top of DPI and reuse them across services. Invest in a few capabilities that strengthen many DPI-enabled services at once – for example, multilingual and voice interfaces, document and form assistance that reduces incomplete submissions, and anomaly detection that protects payments and benefits.

Create the foundations for sovereign AI. Invest in building a shared, scalable public AI backbone that reduces dependence on imported models and lowers the barriers to innovation for local startups, researchers and SMEs, through shared data exchanges and federated compute.

The opportunity before governments is clear: to treat DPI×AI as a core state-capacity agenda. Those that act decisively will not only shape country-aligned intelligent systems, but also deliver better citizen-centric services.

Ultimately, when that mother returns to the clinic, the nurse should spend less time navigating tech systems and more time providing care – because identity and records can be verified with consent, support can be delivered quickly and AI assistance is built into the service in a way that is clear and accountable.