New framework argues government AI depends on strong DPI

The Centre for Digital Public Infrastructure says governments must build trusted digital infrastructure before AI can scale across public services.

As governments rush to adopt generative AI, the Centre for Digital Public Infrastructure (CDPI) argues many risk repeating the mistakes of the blockchain era by treating AI as a shortcut rather than building the digital infrastructure it depends on.

Daniel Abadie had seen the pattern before. Years after blockchain was promoted as a shortcut to modernizing government, he believes generative AI is generating similar expectations that countries can leapfrog foundational digital infrastructure.

The DPI–AI Framework, developed over multiple years, was created to challenge that idea. Abadie and the Centre for Digital Public Infrastructure created it to bring home the idea that without strong digital public infrastructure, AI cannot function reliably inside government systems. “You can’t skip the basics,” Abadie argues. “AI needs the rails before it can run.”

Abadie leads digital transformation efforts from his base in Buenos Aires, Argentina, and is head of technology and partnerships at the Centre for Digital Public Infrastructure (CDPI). He’s been a consultant for both the public and private sector, and has held the role of undersecretary of digital government under Argentina President Mauricio Macri (2015–2019).

The CDPI head, after years of experience embedded in various government departments, says that rebuilding every legacy system is not financially possible. Governments work within a budget, so they must modernize strategically. The framework is designed to help people see DPI as a direct path to AI maturity.

Governments may begin with AI and discover they need DPI to progress, or they may already be building DPI and later connect it to AI capabilities. Additionally, without shared data protocols and clear consent rules, AI models will have nothing stable to stand on.

“It’s about how you architect the technology,” Abadie says, adding that he’s not “married to technology” and that it’s a short-to-medium-term vision paper they’ve created that’s really designed for those with more limited resources, advising them on how to move forward.

Three layers of government AI

To make things practical the framework breaks government AI into three layers: AI blocks, DPI workflows and agents. AI blocks include capabilities such as voice‑to‑text, optical character recognition (OCR) and digital public goods (DPGs) like MOSIP for digital identity verification.

DPI workflows are layered above them and define how these blocks are combined for specific use cases. Agents — the “chefs,” in Abadie’s analogy — execute the workflows. The orchestration layer between them reduces hallucinations and ensures human‑controlled decision points.

The framework is deliberately technology‑agnostic. Governments can use commercial or open‑source models and swap them out without locking themselves into a single vendor or cloud provider.

The Centre for Digital Public Infrastructure plans to support governments as they adopt and adapt the framework. There’s an upcoming webinar, and on the CDPI’s website numerous resources, including a nine-step implementation guide.

A key ambition is accessibility. The framework supports countries with limited budgets. Abadie points to Trinidad and Tobago, which is already using AI agents for service delivery despite its smaller economy. The goal is to embed AI into productive government processes; to reduce service delivery times, improve efficiency and smooth bureaucratic friction.

Tools like voice‑to‑text, which India has deployed across its DPI with further work on its numerous dialects, can help those who speak minority languages or who struggle with text. Verifiable credentials that do not require smartphones can widen access. “Exclusion usually comes from how services are designed,” Abadie says, “not from the technology itself.”

Moving beyond theory

To move beyond theory, the framework includes design tools and an implementation sandbox. Governments can prototype services using full‑code or no‑code DPGs such as OpenFn, making it easier to scale small concepts into national systems.

In practice, AI agents act as interpreters rather than data stores. In a tax payment scenario, for example, the agent verifies identity through DPI blocks, triggers the relevant workflow and connects to the payment gateway — without holding the data itself.

The framework also outlines a nine‑step implementation process: from self‑assessment and use‑case definition to workflow creation, metrics, deployment and iteration. Abadie warns that skipping steps risks excluding citizens.

Countries are already approaching digital transformation in different ways. Rwanda, India, Colombia, Kazakhstan, Trinidad and Tobago have each taken distinct paths, proving there is no single model. To support continued learning, Abadie announced a series of webinars in September, held with 50-in-5 and the Digital Public Goods Alliance, leading up to the Global DPI Summit in Bali next year.

The message is simple, even if the framework is a bit tricky to wrap one’s head around, at least on the first go. The series of forthcoming webinars could be vital to explaining and educating policymakers and stakeholders, moving from concept to practice. But the core is this: AI could potentially transform government, and service delivery, for the better, but only if the foundations are in place. DPI is the road to get there.

Article Topics

Centre for Digital Public Infrastructure (CDPI)  |  digital government  |  digital public infrastructure  |  DPI-AI Framework  |  generative AI  |  government services