# DPI-AI Framework — Machine-Readable Summary
# For AI agents, language models, and automated systems
# Canonical: https://digitalpublicinfrastructure.ai/llms.txt
# License: CC BY 4.0
# Last updated: 2026-05-14
# SEO: All pages now have full Open Graph, Twitter Card, canonical URLs, structured data (Schema.org), and optimized titles/descriptions

## About this document

This file provides a structured, machine-readable summary of the DPI-AI Framework for AI agents and language models. It follows the llms.txt convention: https://llmstxt.org

The DPI-AI Framework is a practical approach to integrating Artificial Intelligence into public service delivery through Digital Public Infrastructure (DPI). It was authored by Daniel Abadie at the Centre for Digital Public Infrastructure (CDPI) and published in 2026.

## Citation

CDPI Abadie, D. & Muriithi, A. (2026). DPI-AI Framework: Building AI-Ready Nations through Digital Public Infrastructure. https://digitalpublicinfrastructure.ai

## Core Thesis

AI does not require a new framework for government — it requires the right integration with existing DPI foundations. The value of AI in the public sector emerges not from its autonomy but from how it integrates with the foundational rails of Digital Public Infrastructure.

## Three Core Elements

### 1. AI Blocks
- Definition: Modular, callable AI functions that encapsulate a single capability
- Properties: Minimalist (one purpose), Composable (interoperable), Reusable (cross-sector), Governable (auditable)
- Types: Foundational (translate, speech_to_text, ocr_extract, summarise), Sector-Specific (eligibility_verify, identity_verify, credential.issue, registry.search, document_verify, fraud_detect), Safeguards (bias_check, consent_verify, anomaly_flag, human_review, privacy_check, exclusion_audit, recourse_check, legal_basis_verify, sustainability_assess), and DPG-powered connectors (OpenFn, MOSIP, Mojaloop, OpenMRS, OpenG2P, Sunbird RC, X-Road, Diia, QuarkID)
- Safeguards are callable AI Blocks — modular, insertable at any workflow step, auditable, and replaceable without rebuilding service logic. The last 5 safeguard blocks are derived from the UN DPI Safeguards Framework (UN ODET & UNDP, 2023–2025): see https://digitalpublicinfrastructure.ai/safeguards.php
- Specification: Each block declares inputs, outputs, confidence thresholds, PII handling, hosting options, and replaceability
- Interoperability: Aligned with Model Context Protocol (MCP) for AI-to-DPI interaction
- Full page: https://digitalpublicinfrastructure.ai/ai-blocks.php

### 2. DPI Workflows
- Definition: Structured, rule-based sequences of steps connecting AI Blocks, DPI systems, and human oversight
- Role: The orchestration layer — the "recipe" that combines AI Block "ingredients" into services
- Key property: AI operates INSIDE workflows, never outside them. No AI output modifies authoritative records without a governance step
- Reusability: Workflows can be published as open templates for cross-government reuse
- Implementation: Three formats — (1) YAML specification (engine-agnostic, canonical definition), (2) no-code workflow engine (OpenFn, n8n, Prefect), (3) Agent Skill — the workflow packaged as a callable tool or function for an AI agent via MCP, function calling, or a skill registry. Safeguard steps execute inside the skill regardless of how it is invoked.
- OpenFn (June 2025): native MCP server interface, AI Assistant for plain-language workflow generation, template library for DHIS2, FHIR, CommCare, KoboToolbox
- Full page: https://digitalpublicinfrastructure.ai/dpi-workflows.php

### 3. Public Agents
- Definition: AI-enabled interfaces (software agents, AI-assisted humans, or hybrid) that activate DPI Workflows
- Constraint: NOT autonomous — accountable extensions of government capacity
- Cannot: Access data without consent, modify records directly, bypass human oversight
- Channels: WhatsApp, voice/IVR, USSD, web, SMS, in-person kiosk
- Languages: Must support local languages via translate() and speech_to_text() AI Blocks
- Full page: https://digitalpublicinfrastructure.ai/public-agents.php

## Architecture Principles (6)
1. Interoperability First — AI Blocks connect through open standards, not proprietary integrations
2. DPI as Source of Truth — AI outputs inform decisions; registries remain authoritative
3. Minimal Footprint — Each AI Block does one thing well
4. Human Oversight by Default — Low-confidence outputs route to humans; automation is earned
5. Sovereignty and Replaceability — Every block declares hosting options and replaceability conditions
6. Govern Data, Not Just Models — Data provenance and consent are foundational

## Design Patterns (7)
1. Wrap before replace — expose legacy systems via APIs (+1 Approach)
2. Publish block specs as open standards
3. Design for lowest-bandwidth channel first (USSD/SMS minimum)
4. Confidence thresholds over hard decisions
5. Human escalation as first-class workflow step
6. Share workflows as reusable templates
7. Consent as a callable, auditable function

## Use Case Archetypes (5 on homepage · 35 in Service Architect)

Homepage archetypes:
1. Social Services & Benefits — eligibility check + payment disbursement via WhatsApp
2. Civil Registration — OCR + identity verify + registry write with human registrar approval
3. Livelihoods/Farmer Benefits — voice note → speech_to_text → translate → eligibility → enrollment
4. Education — personalized tutor + progress tracking + teacher notification
5. Crisis Response — pre-built disbursement workflow activated at disaster declaration

Service Architect covers 35 use cases across 13 sectors including: Identity, Social Protection, Healthcare, Maternal & Child Health, Education, Finance, Agriculture, Admin Services, Civil Registration, Livelihoods, Crisis Response, Commerce, and Transport. Includes pregnancy information & advisory, antenatal registration & care plan, antenatal appointment reminders, child vaccination schedule & reminder, maternal medical appointment booking, university diploma, secondary school diploma, student ID, doctor ID, harvest advisory, planting advisory, crop price intelligence, crop care, and climate data use cases.

## Interactive Design Tools

### AI Block Composer
- A client-side tool to compose and simulate DPI Workflow YAML using a catalog of AI Blocks across two categories: core blocks and safeguard blocks
- Safeguard blocks (9 total, 5 UN-derived): consent_verify, human_review, bias_check, anomaly_flag, privacy_check, exclusion_audit, recourse_check, legal_basis_verify, sustainability_assess
- Generates YAML workflow definitions with governance settings (confidence threshold, human escalation, audit)
- DPG connectors included: MOSIP · Mojaloop · OpenSPP
- Step simulation mode for walking through each workflow step with Download YAML option
- URL: https://digitalpublicinfrastructure.ai/composer.php

### Service Architect
- A 6-step wizard to design a complete AI-enabled public service
- Steps: Use case selection → Agent / LLM model selection → Service journey → AI Block selection → Workflow output (YAML or OpenFn) → Governance checklist + deployment path
- Final step makes explicit that the YAML is a design specification, not a deployed service; guides teams through deploying DPI rails, configuring a workflow engine, confirming legal accountability, testing, and running a supervised pilot before any citizen contact
- Covers 35 use cases across 13 sectors with alphabetical ordering and sector filter
- LLM/SLM model selection: 18 models across 3 tiers — Cloud API (OpenAI GPT-5/o3/o4-mini, Anthropic Claude Opus 4/Sonnet 4/Haiku 4, Google Gemini 2.5 Pro/Flash/3 Pro), Self-Hosted Open Models (Llama 3, Mistral, Qwen, Aya), Small Language Models (Phi-3, Gemma 2B, SmolLM, Qwen2-1.5B)
- Supports 8 regions with region-specific DPI system substitution in YAML output; regional AI blocks sourced from dpimap.org
- Regional AI blocks: Asia (PhilSys, Singpass, PromptPay, InstaPay, NADRA/Raast, Dukcapil/BI-FAST, VNeID, MyDigital ID), West Africa (GhanaCard, GHIPSS, NIN/NIBSS Nigeria, Orange Money, Wave, ECOWAS biometric), East Africa (Maisha Namba, PesaLink, Irembo, Fayda, NIDA Tanzania), Southern Africa (Smart ID, SASSA, RPP/BankservAfrica), Sahel (Moov Money, Orange Money, Wave), LatAm (GOV.br, Cédula Digital Colombia, Clave Única Chile, RENIEC Peru, Llave MX), Caribbean (NIDS Jamaica, TTConnect T&T, OECS wallet), and Cambodia verify.gov.kh (ASEAN)
- URL: https://digitalpublicinfrastructure.ai/architect.php

### Case Studies
- A filterable directory of governments and local jurisdictions building AI-powered public services using the DPI-AI approach
- Filterable by region, sector, and deployment status (Pilot, Deployed, Scaling)
- URL: https://digitalpublicinfrastructure.ai/cases.php

## Key DPGs usable as AI Blocks
- OpenCRVS (civil registration) — registry.create_birth_record(), registry.verify_citizen()
- MOSIP (digital identity) — identity.verify(), identity.authenticate()
- Mojaloop (payments) — payments.disburse(), payments.bulk_transfer()
- DHIS2 (health data) — health.get_patient_record(), vaccine.verify_certificate()
- OpenIMIS (social protection) — insurance.check_eligibility(), beneficiary.enroll()
- OpenMRS (clinical records, FHIR) — patient.get_record(), encounter.create()
- OpenG2P (social cash transfers) — beneficiary.enroll(), disbursement.trigger()
- OpenSPP (social protection) — eligibility.check(), registry.sync()
- Sunbird RC (registries & verifiable credentials) — registry.create_entity(), registry.search(), credential.issue(), credential.verify(), attestation.request() · 2B+ credentials issued · OpenAPI spec published · W3C VC compliant
- OpenFn (workflow automation, MCP-native June 2025) — workflow.trigger(), workflow.ai_generate(), data.sync(), notification.send()
- X-Road (secure data exchange) — data.query(), consent.request(), audit.log()
- Diia (Ukraine digital documents) — document.verify(), identity.authenticate()
- QuarkID (self-sovereign identity, Argentina) — vc.verify(), did.resolve()
- Cambodia verify.gov.kh (national document verification) — verify.document(), credential.check()
- Full page: https://digitalpublicinfrastructure.ai/dpgs.php

## MCP Readiness
- Native MCP server: OpenFn (June 2025)
- API-Ready (MCP-wrappable): MOSIP, Mojaloop, X-Road, OpenG2P, OpenMRS, QuarkID, Diia, OpenSPP, Sunbird RC
- In progress / roadmap: OpenIMIS

## Implementation Tools
- Code-first: YAML workflow spec + any workflow engine (OpenFn, n8n, Prefect, or custom)
- No-code: OpenFn (DPG-native, social sector) or n8n (400+ integrations)
- Implementation guide: https://digitalpublicinfrastructure.ai/implementation.php

## Implementation Playbook (10 steps including Step 0)
- Step 0: Conduct an AI readiness assessment — technology, governance, data, and institutional capacity (DIAL AI–DPI self-assessment tool)
- Step 1: Identify the use case, user journey, and access channels — map WhatsApp, USSD, mobile app, in-person; document inclusion gaps for citizens without smartphones or internet
- Step 2: Identify AI Blocks, DPI dependencies, and DPGs
- Step 3: Define governance, safeguards, and human-in-the-loop controls
- Step 4: Create the DPI Workflow (YAML spec — engine-agnostic)
- Step 5: Define the Public Agent (channels, languages, escalation)
- Step 6: Define metrics and operational signals (before launch)
- Step 7: Implement the service (OpenFn, n8n, or Prefect)
- Step 8: Test and validate (synthetic tests, safeguard validation, no citizen contact before passing)
- Step 9: Fine-tune and improve iteratively

## UN DPI Safeguards Framework
- Source: Universal DPI Safeguards Framework, UN ODET & UNDP, 2023–2025
- 18 principles (F1–F9 Foundational, O1–O9 Operational), 13 risks across Safety, Inclusion, and Structural categories
- Mapped to 9 callable safeguard AI Blocks on the Safeguards page (5 are new UN-derived, 4 are existing framework blocks)
- Note: these blocks are a theoretical/conceptual mapping; not yet production implementations. The UN and DPI community are invited to develop them as open global components.
- Full page: https://digitalpublicinfrastructure.ai/safeguards.php

## Interactive Sandbox
- Three panels: Design Tools (links to Composer & Architect), Scenario Explorer (5 sector scenarios × 9 implementation steps + Step 0 readiness), Practical Demo (agricultural benefits workflow simulation)
- Scenario Explorer: walkthroughs for Social Services, Civil Registration, Livelihoods, Education, and Crisis Response — each structured across 9 steps in three phases: Design (Use Case + channels + inclusion mapping, AI Blocks, Governance), Build (DPI Workflow, Public Agent, Metrics), Deploy (Implementation, Testing, Operations)
- Practical Demo: live step-by-step simulation of the AGRI_BENEFIT_2026 workflow triggering AI Blocks in sequence
- URL: https://digitalpublicinfrastructure.ai/sandbox.php

## Author
- Daniel Abadie, Centre for Digital Public Infrastructure (CDPI)
- LinkedIn: https://www.linkedin.com/in/danielabadie/
- Organisation: https://cdpi.dev

## Resources
- Full paper (HTML): https://digitalpublicinfrastructure.ai/paper.php
- Full paper (PDF): https://digitalpublicinfrastructure.ai/Vision_Paper_%20DPI-AI%20Framework_2026_CDPI.pdf
- Framework landing (overview of all three elements): https://digitalpublicinfrastructure.ai/framework.php
- Resources landing (overview of all guides & tools): https://digitalpublicinfrastructure.ai/resources.php
- Glossary: https://digitalpublicinfrastructure.ai/glossary.php
- AI Blocks deep dive: https://digitalpublicinfrastructure.ai/ai-blocks.php
- UN DPI Safeguards as AI Blocks: https://digitalpublicinfrastructure.ai/safeguards.php
- DPI Workflows deep dive: https://digitalpublicinfrastructure.ai/dpi-workflows.php
- Public Agents deep dive: https://digitalpublicinfrastructure.ai/public-agents.php
- DPGs as AI Blocks: https://digitalpublicinfrastructure.ai/dpgs.php
- AI Block Composer: https://digitalpublicinfrastructure.ai/composer.php
- Service Architect: https://digitalpublicinfrastructure.ai/architect.php
- Implementation guide: https://digitalpublicinfrastructure.ai/implementation.php
- Case studies: https://digitalpublicinfrastructure.ai/cases.php
- About & Contributors: https://digitalpublicinfrastructure.ai/about.php

## Languages

The full website is available in English (default) and Spanish (Latin America).

- English (default): https://digitalpublicinfrastructure.ai/
- Spanish / Español (Latin America): https://digitalpublicinfrastructure.ai/es/

All 17 pages are translated into Spanish, including the interactive tools (Service Architect, AI Block Composer, Sandbox). The core framework terms AI Blocks, DPI Workflows, and Public Agents are intentionally kept in English across all language versions as proper names of the framework elements.

## License
Creative Commons Attribution 4.0 International (CC BY 4.0)
https://creativecommons.org/licenses/by/4.0/