Abstract

The discussion around Artificial Intelligence (AI) and Machine Learning (ML) in official statistics often overlooks a critical prerequisite: the foundational data infrastructure and processes required for these tools to be effective, sustainable, and trustworthy. For smaller and non-advanced National Statistical Offices (NSOs), a rapid leap into advanced analytics is neither feasible nor prudent. This paper argues that a deliberate, phased approach is essential and presents it as a practical blueprint for AI readiness.

Using experience from price statistics production, the paper demonstrates that before implementing AI or ML applications—such as automated price classification, anomaly detection, or inflation nowcasting—NSOs must first strengthen their core statistical systems using existing data and skills. Three foundational steps are highlighted.

First, the paper discusses the implementation of Reproducible Analytical Pipelines (RAP) in Consumer Price Index (CPI) production. Manual and error-prone processes were replaced with a version-controlled, script-based workflow, improving transparency, auditability, and consistency. This reproducible environment creates the structured and traceable data foundation required for any future ML application.

Second, the paper examines the integration and validation of alternative and administrative data sources, including scanner data. Pilot integrations focus on data governance, quality assessment, metadata documentation, and introductory data linkage techniques. Rather than prioritizing automation, this phase emphasizes building institutional confidence and analytical capacity to manage more complex and diverse datasets.

Third, the paper highlights the importance of upskilling staff in foundational data science. Through targeted internal training, price statisticians use historical CPI time series to learn scripting, automated validation rules, and basic time series techniques applied to CPI data, such as trend analysis, seasonal decomposition, and simple forecasting. This approach fosters a hybrid analytical mindset that combines domain expertise with computational thinking, enabling statisticians to critically evaluate and responsibly oversee future AI tools.

The paper concludes that investments in reproducibility, data governance, and human capital are not separate from the AI journey but constitute its indispensable first chapter. For non-advanced NSOs, this foundation transforms AI from a risky experiment into a logical, manageable, and trustworthy progression aligned with the principles of official statistics. The paper contributes practical lessons for NSOs seeking to modernize incrementally while preserving trust in official price statistics.