Practical Guide to AI-Ready Data Dissemination
A cookbook for national statistical organizations. Version 0.1 (working draft), October 2026.
This guide is a working draft from the AI for Data – Data for AI program. It collects practical approaches and open standards. It does not represent official World Bank Group guidance or policy and is not legal advice; an organization applies it under its own statistical law, data protection rules, and institutional policies. Comments and corrections are welcome as GitHub issues, and the contribute page explains how to propose a recipe.
People increasingly reach statistics through AI systems: search assistants, chatbots, and analysis tools. These systems read the metadata, the web pages, and the APIs that a statistical organization publishes, and they read them without asking anyone for help. When those sources are incomplete, ambiguous, or hard to retrieve, the systems return the wrong series, mix up definitions, or quote numbers without a source, and they do so under the organization's name.
This guide is organized around nine questions that an organization can ask about its own dissemination system, from "can people and AI find our statistics?" to "can we maintain what we build?". Each chapter explains why the question matters and what a good answer looks like. It then gives recipes, which are tasks with a defined result, the skills and the time they need, the code or the template to do them, and a worked example on the running example.
The recipes build on published standards: SDMX for aggregate data, DDI for microdata, ISO 19115 for geospatial data, Dublin Core for documents, DCAT and schema.org for catalogs, DataCite for identifiers, and the World Bank's own metadata schemas for records. The standards page lists each one, where the World Bank uses it, and where it appears in the guide. The chapters follow one indicator catalog as a running example; the Application by data type page shows how each question and recipe applies to survey microdata, geospatial data, documents, and tables, with a record and a profile for each.
Many of the highest-value steps are ordinary data practice: complete metadata, open structured data, stable identifiers, and documented APIs. These steps help human users and conventional search as well. A chatbot is an optional later layer, and chapters 4 and 5 explain what has to be in place before one is safe to run.
Prerequisites
The recipes assume the following are in place or can be arranged:
- A published catalog of indicators or datasets, on a website the organization controls, with someone who can edit its pages or records.
- One person who can run a Python script (the checks in this guide use the standard library and one package) and one who knows the statistics well enough to write definitions and judge answers.
- For the AI-ready level, a developer who can change the website template or the API; for the AI-native level, a server that can run a small open-weight model, or access to a hosted model under the AI-use policy of chapter 8.
First steps
For a team with limited time, this sequence gives the most useful result first:
- Run the completeness check of chapter 1 on the catalog export. It takes half a day, and its output is the work list for the metadata.
- Write the one-page AI-use policy of chapter 8. It takes a day, and nothing else in this guide is safe to start without it.
- Write thirty known-item questions and score the current search with them, as chapter 6 describes. It takes a day, and every later change is measured against this baseline.
- Publish the series as tidy files at stable addresses, with the source and a citation on each page, following chapters 2 and 5. This serves every user and every AI system at once.
A conversational interface (chapter 4) comes last, after the four steps above are at the AI-ready level.
Self-assessment
The self-assessment takes about five minutes. It places an organization at a level in each chapter and suggests a reading order.
Chapters and questions
| Chapter | Question for a statistical organization | Practical topics |
|---|---|---|
| 1. Findability of statistics | Can people and AI find our statistics? | Metadata completeness, indexing, DCAT and schema.org, semantic search, multilingual metadata |
| 2. Retrieval of data by software | Can AI retrieve our data reliably? | Open formats, APIs, stable identifiers, machine-readable access, MCP |
| 3. Meaning of the data: units, concepts, and classifications | Can AI understand what the numbers mean? | Units, concepts, classifications, method notes, SDMX, DDI, code lists |
| 4. Natural-language access | Can users ask questions in their own words? | Natural-language search, retrieval-augmented generation, conversational interfaces, multilingual access |
| 5. Trust and provenance of answers | Can answers be trusted and traced? | Provenance, citation, numeric verification, grounding, freshness |
| 6. Evaluation | Can we tell whether our AI-ready steps improved anything? | Retrieval benchmarks, hallucination tests, answer accuracy, multilingual evaluation |
| 7. Monitoring of data use | Can we see how our data are used, including by AI systems? | Dataset mention detection, citation tracking, AI visibility |
| 8. Governance and responsible operation | Can we operate AI-supported dissemination responsibly? | Governance, human oversight, security, privacy, model choice, vendor dependence |
| 9. Sustainability and maintenance | Can we maintain what we build? | Architecture, open standards, small and open models, cost, skills, lifecycle |
These questions follow the same path as the program's AI-ready data framework: find the right data, retrieve it with context, interpret it, and verify the result.
Maturity levels
Each chapter describes three levels. An organization can sit at different levels in different chapters, and the first level is a complete, useful state on its own.
| Level | Meaning |
|---|---|
| Foundational | Data and metadata are published in open, documented formats that a person can download and that a crawler can read. |
| AI-ready | Data and metadata are structured, identified, and accessible through documented interfaces, so that software, including AI systems, can use them without manual steps. |
| AI-native | The organization offers AI-oriented interfaces such as semantic search or an agent protocol, with provenance and verification built in. |
For access, as an example:
- Foundational: downloadable CSV files with a metadata page.
- AI-ready: a documented REST API, stable identifiers, and structured metadata.
- AI-native: semantic discovery, a standard agent interface such as MCP, and provenance on every answer.
Running example
The recipes use one small, fictional catalog so that each chapter picks up where the previous one ends. The example organization publishes five series:
| id | title | unit |
|---|---|---|
FS_UNDERNOURISH_PCT | Prevalence of undernourishment | % of population |
ED_PRIMARY_NER_PCT | Net enrolment rate in primary education | % of age group |
HE_U5_MORT_RATE | Under-five mortality rate | deaths per 1000 live births |
LF_UNEMP_PCT | Unemployment rate | % of labour force |
PV_HEADCOUNT_NPL_PCT | Poverty headcount ratio at national poverty line | % of population |
One record has an empty definition on purpose, so that the completeness
report in chapter 1 and the search scores in chapter 6 have something to
find. The catalog file uses the field names of the
World Bank indicator metadata schema
(idno, name, definition_long, measurement_unit, periodicity, and
so on), and the values file uses SDMX cross-domain concept names (SERIES,
REF_AREA, TIME_PERIOD, OBS_VALUE, OBS_STATUS, UNIT_MEASURE). The
catalog, the values, and every script in this guide are in the files list
below. All values are invented.
Scope of the examples
The examples are small on purpose: five series, fifteen test questions,
and single-file scripts with no dependencies beyond the standard library
and jsonschema, so that a reader can run them in an afternoon and see
the shape of each check. They are starting points. Each recipe
ends with a What this does not do note that says where the example
stops and what the full version adds, and points to the program's tools
where one exists. Treat a recipe as done when the organization's own data pass
the organization's own version of the check, with the profile, question set, and
rules that the organization has agreed.
Files
Each file is embedded in the chapter that uses it and can be downloaded here. The scripts use only the Python standard library unless stated.
| File | Used in |
|---|---|
| example_catalog.csv | The running example catalog (World Bank indicator schema field names) |
| example_values.csv | The running example values (SDMX cross-domain concept names) |
| metadata_template.csv | Chapter 1: fields every series record needs, named as in the World Bank indicator schema |
| check_metadata.py | Chapter 1: three-layer metadata check (schema, profile, validity) for every World Bank metadata type; needs jsonschema 4.18+ |
| profile_indicator.json | Chapter 1: fields required at each level for indicators, and the validity rules, for the organization to edit |
| profile_microdata.json | Chapter 1: the same for survey microdata (DDI Codebook based schema) |
| profile_geospatial.json | Chapter 1: the same for geospatial data (ISO 19115/19139 based schema) |
| profile_document.json | Chapter 1: the same for documents (Dublin Core based schema) |
| example_microdata.json | A microdata record in the World Bank schema (fictional labour force survey) |
| example_geospatial.json | A geospatial record in the World Bank schema (fictional district boundaries) |
| example_document.json | A document record in the World Bank schema (fictional report, linked to the series and survey it used) |
| dataset.jsonld | Chapter 1: schema.org Dataset markup |
| openapi_minimal.yaml | Chapter 2: API description |
| mcp_server_example.py | Chapter 2: MCP server (needs the mcp package) |
| verify_numbers.py | Chapter 5: numeric verification |
| eval_questions.csv | Chapter 6: question set |
| score_retrieval.py | Chapter 6: retrieval scoring |
| ai_component_register.csv | Chapter 8: AI component register |
| ai_use_policy_template.md | Chapter 8: AI-use policy template |
Related resources
| Resource | Purpose |
|---|---|
| This guide | What a statistical organization can do, and why, in a sequence of practical steps. |
| Application by data type | How the questions and recipes apply to indicators, microdata, geospatial data, documents, and tables. |
| Standards used in this guide | The published standards behind each recipe, with the ones the World Bank has adopted. |
| AI-readiness assessment | A structured assessment of where an organization stands before planning. |
| From assessment to action | Each assessment dimension mapped to the program's tools, methods, and the chapters of this guide. |
| Documentation | How each program method and tool works. |
| Code | Open-source software to deploy or adapt. |
Citation
World Bank. 2026. Practical Guide to AI-Ready Data Dissemination: A cookbook for national statistical organizations. Version 0.1. AI for Data – Data for AI. https://worldbank.github.io/ai4data/cookbook/ai-ready-dissemination/
Version history
- 0.1 (October 2026). First draft: nine chapters, recipes, templates, self-assessment.