DPI-AI Framework:
Building AI-Ready Nations through Digital Public Infrastructure

About this paper

This paper is intended for digital leaders in government, development partners, donors, and practitioners who are shaping local, national and global digital agendas, as well as for anyone interested in societal transformation, poverty alleviation, and developmental impact using technology.

Its aim is not to promote AI adoption for its own sake, but to provide a clear and practical way to align emerging AI capabilities with the foundational rails of DPI towards development outcomes, enabling more adaptive, coherent, and publicly governed digital systems over time. In this vision, the value of AI emerges not from its autonomy but from how it integrates dynamically with the foundational rails of DPI to create adaptive, intelligent, and publicly governed digital infrastructure.

Executive Summary

This paper presents the DPI–AI Framework as a practical way to think about how Artificial Intelligence (AI) can be integrated into public digital systems through Digital Public Infrastructure (DPI). It is written at a moment when AI capabilities are advancing rapidly, while many governments are still grappling with fragmented systems, legacy architectures, and uneven institutional capacity. Rather than proposing AI as a standalone transformation, the paper explores how existing DPI foundations can provide structure and coherence for the use of AI in the public sector.

Many of the ideas discussed in this paper are not new. Modularity, shared services, workflow orchestration, and human oversight have long been part of public sector digital ambitions. What has changed is the maturity and accessibility of AI, combined with the attention it now receives from political leaders, institutions, and the market. This convergence creates pressure to adopt AI quickly, often before there is clarity on how it should interact with existing public systems.

The DPI–AI Framework positions AI not as a new layer within DPI, but as an external and interoperable set of capabilities that connect to DPI through shared standards, governance mechanisms, and safeguards. It focuses on foundational DPI rails such as digital identity, data exchange, and payments.

At the core of the framework are three interrelated elements:

• AI Blocks are modular units of AI capability that can be invoked as callable functions. They perform sector-specific functions (identity verification, eligibility checking) and foundational tasks (translation, OCR, summarisation).
• DPI Workflows provide the orchestration layer that coordinates AI Blocks with existing DPI systems, policy rules, data flows, and human oversight.
• Public Agents are AI-enabled interfaces that interact with citizens and public servants, drawing on AI Blocks through DPI Workflows.

The value of AI emerges not from its autonomy but from how it integrates dynamically with the foundational rails of DPI to create adaptive, intelligent, and publicly governed digital infrastructure.

Introduction

In a pivotal scene from the 1999 film The Matrix,[1] Neo undergoes an instantaneous transformation. A diskette uploads knowledge directly into his brain. He opens his eyes and declares: "I know Kung Fu." For a generation raised on the cusp of the internet revolution, this symbolised a dream of frictionless learning and limitless upgradeability.

Two decades later, the fantasy remains fiction. But we are entering an era where advanced statistical models — particularly Agentic AI based on Large and Small Language Models[2] — can assist in generating information, summarising documents, responding to user input, and even tailoring services at scale. These systems, however, do not "understand," "reason," or "think" in the human sense. They operate by predicting patterns in data, not by grasping meaning or engaging in cognition.[3]

There is no need for a new mental model for AI in government. The DPI approach is already set up to integrate AI, as long as we carefully consider safeguards, emphasising cross-sectoral, governance, and procurement implications. AI does not require a separate framework but rather adjustments within the existing DPI model. From this perspective, AI naturally fits as just another modular building block within DPI, reinforcing the principle of minimalism and reusability.

Why act now?

The cost of inaction is already visible. Governments that postpone building AI-ready digital infrastructure risk deepening dependency on proprietary systems and locking public functions into closed platform models whose costs increase over time. Without interoperable and publicly governed AI capabilities, ministries duplicate investments, procurement becomes fragmented, and crisis responses slow down.

The greater cost, however, is institutional. When governments lack open and modular infrastructure, they lose the ability to choose their own digital paths, to adapt technology to local contexts, and to govern it on their own terms.

Digital sovereignty is not achieved through isolation — but through capability. It depends on building the competence to design, reuse, maintain, and evolve shared public systems independently.[6]

This framework does not assume that AI should be prioritised ahead of foundational DPI work. In many contexts, the immediate focus remains on clean registries, reliable APIs, last-mile connectivity, and minimal institutional capacity. The purpose of the framework is to make AI integration coherent when the time is right — not to rush it.

What is Digital Public Infrastructure (DPI)?

DPI is an approach to digitalisation focused on creating foundational, digital building blocks designed for the public benefit.[4] This approach combines open technology standards with robust governance frameworks to encourage private community innovation to address societal scale challenges such as financial inclusion, affordable healthcare, quality education, climate change, and access to justice.

It is based on five technology architecture principles:[5]

1. Interoperability
2. Minimalist, reusable building blocks
3. Diverse, inclusive innovation by the ecosystem
4. A preference for remaining federated and decentralised
5. Security and privacy by design

At an implementation level, this approach takes shape through a set of foundational, interoperable building blocks: identifiers and registries; data sharing; AI/ML models; trust infrastructure; discovery and fulfilment; and payments.

Government, at its core, is a system of registries (records that are created, updated, or read — identity, land, benefits, credentials) and rules (laws and policies that are executed — eligibility, consent, approval logic). AI Blocks can read and write registries. DPI Workflows encode and execute rules. Public Agents interpret citizen intent and trigger the right combination of both.

What is AI in the public sector?

AI refers to computer systems that learn from data and make context-based judgements to perform tasks such as understanding language, answering questions, executing tasks, recognising patterns, or generating text and media. Unlike traditional software that follows fixed rules, AI systems adapt their outputs based on patterns learned during training and ongoing reinforcements.

The most visible examples today are Large Language Models (LLMs), which are powerful models trained on vast amounts of data, and Small Language Models (SLMs), which are lighter and designed to run on smaller devices or at lower cost. For public sector applications — particularly in low-resource environments — SLMs offer significant advantages in cost, latency, and sovereignty.

Agentic AI refers to systems that use AI to both provide information and take actions: connecting with third-party systems, invoking AI blocks, or running workflows to complete tasks such as searching databases, submitting forms, or scheduling appointments. Within the DPI-AI Framework, Public Agents are the primary expression of agentic AI in the public sector.

The DPI-AI Framework

The DPI–AI Framework is a design approach that extends Digital Public Infrastructure by connecting AI as an external, interoperable layer. It enables modular AI capabilities — Public Agents, DPI Workflows, and AI Blocks — to interact with DPI systems through shared standards, governance, and safeguards.

The framework supports adaptive and accountable public service delivery while maintaining openness, transparency, and human oversight. It deliberately avoids prescribing specific technologies or models. Instead, it offers a shared mental model that helps governments, development partners, and ecosystem actors reason about where AI fits within a DPI-based architecture.

The +1 Approach: Governments don't need to replace legacy systems to become AI-ready. Layer new capabilities alongside what exists — wrapping legacy systems with interoperable APIs and DPI blocks. Every registry cleaned, every API published, every governance process formalised is a direct enabler of AI-ready public infrastructure.

AI Blocks: Units of AI as callable functions

To embed intelligence into public systems, we need units of AI that act like callable functions. In the DPI-AI Framework these units are AI Blocks.

These blocks follow the same logic as DPI components. AI Blocks are discrete, auditable modules that encapsulate a single capability. Each one can be invoked by a Public Agent or by another system, much like a function call.

Think in terms of ingredients and recipes. AI Blocks are the ingredients. DPI Workflows are the recipes that combine ingredients with policy and process to deliver a service. Ingredients remain reusable across many recipes. Recipes remain transparent, auditable, and easy to revise as rules and needs evolve.

An AI Block has four properties that mirror DPI principles:

1. Minimalist. One clear purpose with a bounded interface.
2. Composable. Designed to interoperate with other Blocks and workflow engines through open APIs and common data formats.
3. Reusable. Applicable across agencies, sectors, and regions without rework.
4. Governable. Observable, testable, and aligned with policy, with built-in audit trails and privacy controls.

Foundational AI Blocks

General-purpose capabilities reusable across many sectors. They provide the baseline tools — the multimodal and local language primitives that are prerequisites for an inclusive and AI-ready nation:

• Translation and transliteration into local languages
• Speech-to-text and text-to-speech
• Optical character recognition (OCR)
• Text summarisation
• Image and video recognition

Sector-Specific AI Blocks

Tailored to specific sectors or workflows. They embed policy and operational logic into callable functions:

• Eligibility verification for social protection programs
• Identity verification integrated with the digital ID platform
• Vaccine certificate validation in health services
• Clinical decision support in hospitals
• Personalised tutoring support in education

Safeguards as Callable AI Blocks

Safeguards can themselves be AI Blocks — callable functions that check for bias, validate consent, detect anomalies, or flag low-confidence outputs for human review. By treating safeguards as modular, replaceable components, governments can evolve their risk management without rebuilding entire workflows.

Deep dive: AI Blocks →

DPI Workflows: The Orchestration Layer

While AI Blocks provide the modular capabilities, what turns them into usable public services is how they are orchestrated. This is the role of DPI Workflows — structured, auditable recipes for service delivery. They connect Public Agents, AI Blocks, and core infrastructure such as digital identity, payments, and data exchange into coherent flows.

Each workflow defines a clear sequence of steps, along with the data, conditions, and safeguards that guide how a service is delivered. For example, a social protection public agent may invoke a workflow that verifies identity, checks eligibility through an AI Block, and initiates a payment using a government payment rail.

Just as importantly, DPI Workflows can themselves become reusable assets. Published in open repositories, workflows could be adapted and reused by other governments or agencies, much like open-source code. Treating workflows as templates allows countries to borrow from each other's playbooks, install them with minimal effort, and adapt them to local policies and contexts.

Generic DPI Workflow Template (YAML)

Deep dive: DPI Workflows →

Public Agents: AI-enabled interfaces

Small Language Models (or Large Language Models depending on scale, budget or context), optimised for efficiency and contextual relevance, can be deployed within ministries, municipal offices, or service centers. When coupled with DPI, they power the creation of Public Agents.

Public Agents are AI-enabled assistants that may take the form of software-based agents, AI-assisted public servants, or hybrid arrangements combining both. They interact with citizens, officials, or service providers within defined rights-based boundaries, and activate DPI Workflows. In all cases, Public Agents are not autonomous actors but accountable extensions of government capacity.

Public Agents can support tasks such as:

• Act as on-demand experts, embedded in messaging platforms, web portals or mobile apps
• Handle citizen queries using publicly governed AI blocks such as eligibility verification, grievance filing, or identity assistance
• Operate in low-resource or multilingual environments, using lightweight language models adapted with local data and civic logic

As modular capabilities mature, there is a natural progression toward multi-agent orchestration — multiple agents collaborating by sharing context, coordinating across workflows, and dynamically decomposing complex tasks.[8] DPI provides the shared digital rails to meet these requirements, offering common protocols, registries, and governance frameworks.

Deep dive: Public Agents →

Architecture Principles and Design Patterns

Six Principles — The "Why"

1. Interoperability First. AI Blocks connect through open standards — not proprietary integrations. Any block can be swapped without rebuilding the workflow.
2. DPI as the Source of Truth. AI outputs inform decisions; registries remain authoritative. No AI block directly modifies authoritative records without a governance step.
3. Minimal Footprint. Each AI Block does one thing well. Complexity is managed at the workflow level, not embedded in blocks.
4. Human Oversight by Default. Any low-confidence output, exception, or safeguard violation routes to a human. Automation is earned, not assumed.
5. Sovereignty and Replaceability. Model provenance, hosting options (sovereign cloud / on-prem / external API), and replaceability conditions are declared in every block spec.
6. Govern Data, Not Just Models. Data provenance, consent, and stewardship are foundational. A government cannot train a trustworthy model on ungoverned data.

Seven Patterns — The "How"

1. Wrap before replace. Before rewriting legacy systems, expose them via interoperable APIs. This is the +1 Approach.
2. Publish block specs as open standards. Every AI Block should have a machine-readable spec declaring inputs, outputs, confidence thresholds, and replaceability.
3. Design for the lowest-bandwidth channel first. USSD and SMS as minimum viable channels; WhatsApp voice as primary for last-mile contexts.
4. Confidence thresholds over hard decisions. Every AI Block output includes a confidence score and a defined fallback path.
5. Human escalation is not a failure. It is a designed feature. Build escalation paths as first-class workflow steps, not exception handlers.
6. Share workflows as templates. Publish successful workflow configurations in open repositories for cross-government reuse.
7. Consent is a DPI block. Treat user consent as a callable, auditable function — not an implicit assumption or a static checkbox.

From Digital Public Goods to AI Blocks

The DPG ecosystem — which includes open-source software like OpenCRVS, Mojaloop, OpenIMIS, and DHIS2[7] — provides a natural foundation for AI Blocks. Existing DPGs can be extended to expose AI-callable interfaces. New DPGs can be purpose-built as AI Blocks from the outset.

AI Blocks themselves can qualify as Digital Public Goods when they: are open-source or governed by open standards; declare their training data provenance; embed privacy and consent controls; and are designed for replaceability rather than lock-in. This creates a reinforcing cycle: DPGs enable AI, and AI capabilities become DPGs.

Use Cases: Making the Vision Tangible

The framework is illustrated through five use case archetypes, each showing how AI Blocks, DPI Workflows, and Public Agents combine to deliver a specific public service.

🛡️ Access to Social Services and Benefits

A citizen authenticates via WhatsApp, submits a benefit claim in their local language using voice input. The Public Agent invokes a DPI Workflow: speech_to_text() → translate() → identity.verify() → ai.eligibility_verify() → [if confidence < 0.85 → human_escalate()] → payments.disburse(). Human caseworker reviews borderline cases. Audit trail created at every step.

📝 Civil Registration and Legal Services

Hospital staff submit a birth notification form. The workflow invokes: ocr.extract(form_image) → identity.verify(parent_ids) → registry.check_duplicate() → [if anomaly → human_review()] → registry.create_record() → certificate.generate(). Human registrar retains final authority on every registry write.

🌾 Livelihoods and Economic Participation

A farmer sends a WhatsApp voice note in a local dialect. The agent transcribes, translates, extracts structured fields (name, location, crop, land size), verifies identity against an agricultural registry, and submits a benefit application — with USSD as fallback for low-connectivity zones.

📚 Education and Human Development

A student authenticates, the Public Agent assesses their current level, generates personalised content via personalized_tutor(), and tracks progress. Teachers are notified when at-risk patterns are detected. Student data is governed under strict purpose limitation — only mastery scores, not full interaction logs.

🚨 Crisis Response and Urban Services

A disaster is declared. A pre-built, tested workflow activates: disaster.declare(area_polygon) → gis.identify_population() → registry.match_eligible() → fraud.screen(disbursement_list) → human_authority.approve(batch) → payments.bulk_disburse(). Speed comes from pre-built infrastructure, not from removing oversight.

Critical note: The Crisis Response workflow must be built and tested before a crisis. This is not a use case to stand up during an emergency.[9]

Challenges Ahead: Safeguards, Procurement, and Governance

Procurement

Most government procurement frameworks were not designed for modular, callable AI capabilities. They favour large platform contracts over interoperable building blocks.[10] Reforming procurement to support AI Blocks as reusable public goods — with clear specs, open APIs, and replaceability requirements — is a prerequisite for the framework to scale.

Institutional Capacity

Governing AI requires skills that government often lacks: AI literacy, data governance expertise, workflow engineering, and the ability to write and audit block specifications. Building this capacity — in-house and across the ecosystem — is as important as the technical infrastructure itself.

Data Quality

Clean registries are a prerequisite, not a given. AI Blocks that read from inaccurate, incomplete, or biased datasets will produce inaccurate, incomplete, or biased outputs. Every registry cleaned is a direct investment in AI readiness.

Connectivity and Inclusion

Last-mile delivery requires offline-capable workflows, USSD fallbacks, and voice-first interfaces. Designing for the lowest common denominator channel is a governance principle, not a technical afterthought.

Language and Cultural Context

Foundational AI Blocks for local language translation, speech-to-text, and transliteration are prerequisites for inclusion. Without these, AI-enabled services will serve only those who speak dominant languages — a structural exclusion risk.

Conclusion

The DPI-AI Framework is not a call to adopt AI at all costs. It is a call to be ready — to build the modular, governed, interoperable foundations that allow AI to be integrated when the moment is right, on terms set by governments and citizens rather than by vendors.

The three elements — AI Blocks, DPI Workflows, and Public Agents — offer a practical vocabulary for this work. They connect the principles of DPI (minimalism, reusability, interoperability, sovereignty) with the realities of AI deployment in resource-constrained, multilingual, and institutionally complex environments.

The task ahead is not to wait for a perfect AI system, but to build the infrastructure that would allow any capable AI system to operate within public governance. Every registry cleaned, every API published, every workflow documented, and every safeguard designed is a step toward that goal.

The question is not whether governments will use AI. It is whether they will use it on their own terms, within their own governance structures, and in service of their own development goals.