Speaker
Description
Official statistics is rapidly adopting cloud based and open-source self-service production platforms, modern programming languages, and more machine learning in production. In parallel, generative AI is becoming a natural part of development and maintenance work. These shifts are often aimed at improving speed, reproducibility, and reuse, but also increase the need for systematic operational practices that protect quality, trust, and stability.
Promising developments already point in this direction, including Reproducible Analytical Pipelines (ONS), Principles for Statistical Production (BLS), work on Implementation Frameworks within the ESS AIML4OS programme and examples of quality requirements on production systems developed at several NSIs including Statistics Sweden. However, the community still lacks a shared operational level framing that connects such efforts and supports reuse across organisations.
This paper proposes starting a coordinated UNECE effort towards “StatOps”, using an MLOps inspired structure across Principles, Components, Roles, and Architecture, but grounded in end-to-end statistical production. The aim is to capture practical patterns such as version control, automated testing, CI/CD, controlled execution environments, and monitoring, and to make them reusable within the roles and competencies typical for statistical organisations. A further open question is how StatOps should incorporate lessons on where generative AI can augment specialist expertise in statistical production, without weakening methodological responsibility or institutional control.
Key words: StatOps, statistical production, MLOps, generative AI, implementation frameworks, GSBPM