Skip to main content
World Bank Group, Development Data GroupWorld Bank Group, Development Data GroupCookbooks

Practical guides for AI-ready data

Ten cookbooks for national statistical organizations, each organized around the questions a team asks, with recipes at three maturity levels, scripts that run on a shared example, open standards, and checklists. A team new to the series starts with the reading path for its role; a team with an assessment profile starts from the dimension with the largest gap.

Data products that AI can use

National statistical organizations

Practical Guide to AI-Ready Data Dissemination

Helps a statistical organization make its published statistics findable, retrievable, and understandable by people and by AI systems. Nine chapters cover catalog records, open files and APIs, documented meaning, natural-language access, provenance, evaluation, monitoring of use, governance, and sustainability, each with recipes, worked examples, and a checklist.

9 chapters Dissemination, library, and research data teams

Practical Guide to Extracting Data from Documents

Helps a team recover the statistics locked in its reports and PDFs as structured, verified, and citable data. Nine chapters cover the inventory of documents, text from scans, locating and extracting tables and figures, verification against the page, series across documents, publication, citation, and running the pipeline over a whole backlog.

9 chapters Data managers and dissemination teams in national statistical organizations

Practical Guide to Publishing ML-Ready Datasets

Helps data managers publish the examples a statistical organization produces, such as coded responses and transcribed tables, as datasets for training and evaluating models. Eight chapters cover selection, file form, Croissant records, dataset cards, representativeness, splits, licence and privacy, and versions.

8 chapters

Metadata and interfaces

Metadata curators and their managers in national statistical organizations

Practical Guide to Metadata Curation with Language Models

Helps a curation team use language models on catalog metadata with a curator deciding every change. Eight chapters cover which tasks a model may assist, migration of legacy records, drafting of missing fields, quality scoring, review of suggestions, vocabularies, translation, and improvement from the decisions curators record.

8 chapters Developers and dissemination teams in national statistical organizations

Practical Guide to Serving Official Statistics to AI Agents

Helps developers and dissemination teams expose official statistics to AI assistants through the Model Context Protocol. Nine chapters cover the scope of agent access, the design of the tools, the context an assistant needs, provenance in every response, access control, safety, evaluation, distribution to clients, and operation, built on the Data360 MCP server.

9 chapters

Models inside the office

Methodologists and production managers in national statistical organizations

Practical Guide to Language Models in Statistical Production

Helps methodologists and production managers use language models inside statistical production, from questionnaire design and open-text coding to editing, anomaly explanation, and release commentary. Seven chapters keep a person at every point of decision and document model use in the quality report.

7 chapters IT and data science leads in national statistical organizations

Practical Guide to Small and Open Models for Statistical Offices

Helps IT and data science leads run small and open-weight models on the organization's own servers for its narrow tasks. Seven chapters cover when a small model is enough, candidates and licences, hardware and serving, adaptation, comparison with larger models on the organization's own suite, security, and cost.

7 chapters Microdata teams and methodologists in national statistical organizations

Practical Guide to Synthetic Data for Sharing

Helps microdata teams produce synthetic files that applicants, students, and developers can work with while the real file stays protected. Seven chapters cover what synthetic data are for, the choice of method, utility and disclosure risk, linked tables and panels, documentation and release, and the place of synthetic files in the access workflow.

7 chapters

Measurement

Teams that deploy AI components in national statistical organizations

Practical Guide to Evaluation Suites for Statistical AI

Helps a team build, run, and maintain the evaluation suites that decide whether a search, an assistant, an extractor, or a drafting tool is good enough to publish and safe to change. Nine chapters cover what to evaluate, question sets and labels, the measures for each kind of component, graders, gates on changes, evaluation in production, and reporting.

9 chapters Dissemination teams and management in national statistical organizations

Practical Guide to Monitoring Data Use

Helps a dissemination team see where its data are used, in research, policy, news, and AI answers. Nine chapters cover what counts as use, access and citation counts, collecting documents, finding and matching mentions of the data, reporting use with its caveats, and acting on the results.

9 chapters

Reading paths by role

Dissemination lead
  1. Findable records and search by meaning
  2. Source, citation, and verified numbers
  3. What an agent interface should do
  4. What counts as use of the data
Microdata curator
  1. Model drafts under a curator's decision
  2. When a synthetic file helps
Methodologist
  1. Where a model may assist production
  2. Coding with a threshold and a re-coded sample
  3. Test questions and labels
  4. Representativeness of training data
IT and data science lead
  1. When a small model is enough
  2. Sizing and running a model server
  3. Gates on every change
  4. Keeping the interface safe
Management
  1. The AI-use policy and the component register
  2. The statement of model use
  3. What models cost to sustain

New cookbooks follow the authoring guide, which sets the structure and writing rules and names the scaffold and checker.