evidoria

← Retour à la liste

Bonne pratique Importée

ClassifAI et Survey Assist — le pipeline IA de l'ONS britannique qui a divisé par deux le temps de codage des professions

Royaume-Uni · Newport · Voir le profil de Royaume-Uni

Preuves: Observationnel / avant-après Top 7% 87/100 · Demandez à Evidence Copilot à propos de cette pratique

L'ONS britannique a créé ClassifAI, un paquet open source qui code le texte d'enquêtes en classifications officielles, divisant par deux le temps de codage des professions sur une grande enquête emploi, et pilote Survey Assist sous registre public de transparence.

-50% %
Reduction in occupation-coding processing time on ASHE
~90,000 households
Households potentially covered by Survey Assist once deployed (annual, once deployed)
~2.25 million instances
Survey instances Survey Assist could cover annually (annual, once deployed)
ClassifAI et Survey Assist — le pipeline IA de l'ONS britannique qui a divisé par deux le temps de codage des professions

Détails

Maturité
En expansion
Promoteur
UK Office for National Statistics (ONS) - Data Science Campus
Période
2023-present
Mots-clés
official statistics, government data science, labour statistics

Contexte

National statistics offices spend significant staff time manually coding survey respondents' free-text answers into standard classifications such as the Standard Occupational Classification (SOC) and Standard Industrial Classification (SIC), a bottleneck as survey volumes grow. The UK's Office for National Statistics (ONS) Data Science Campus built ClassifAI, an open-source Python package that combines semantic search over previously coded examples with retrieval-augmented generation.

Objectifs

ClassifAI aims to assign free text automatically to classifications including SOC, SIC, COICOP and their international equivalents (ISCO, ISIC). Building on the same pipeline, the ONS is separately piloting 'Survey Assist' on the Transformed Labour Force Survey, using AI-generated follow-up questions to sharpen SIC/SOC coding.

Activités

ClassifAI, according to the ONS's own account, is now coding labour market data in production, including a large-language-model application used on the Annual Survey of Hours and Earnings (ASHE). Survey Assist, which could cover around 90,000 households (about 2.25 million survey instances a year) once deployed, was published on 30 October 2025 under the UK's Algorithmic Transparency Record Standard (v3.0), documenting its pilot status, data flows, human clerical review and a planned evaluation against gold-standard coded data.

Résultats

The ONS's own account states that the large-language-model application built on ClassifAI halved processing time for occupation coding on ASHE. Survey Assist, the newer follow-up-question tool, has not yet published accuracy or consistency results, as testing was still under way through 2025.

Conclusions

The Survey Assist Transparency Record sets a materially higher public-transparency bar than most comparable AI deployments in this dataset, but the well-evidenced part of the impact claim remains limited to the ASHE processing-time figure, with Survey Assist itself still unproven in production.

Mise en œuvre
Coût indicatif
Moyen (50 k€–500 k€)
Délai jusqu’aux résultats
Moyen (1–3 ans)
Personnel et compétences
ONS Data Science Campus (in-house team, Newport)

Conditions de réussite

  • Open-source (MIT-licensed) codebase enabling reuse and scrutiny
  • Retrieval-augmented generation over a library of previously coded examples
  • Publication under the UK Algorithmic Transparency Record Standard (v3.0)
  • Human clerical review retained alongside the AI-generated codes

Modes d’échec courants

  • Survey Assist has no published accuracy or consistency results yet; testing was still under way through 2025

Habituellement financé par

National / regional programmes

Pistes de financement indicatives pour des pratiques de ce type — vérifiez toujours les appels en cours et les règles d'éligibilité de chaque programme.

Kit de réplication

Artefacts réutilisables de cette pratique — tels que publiés par leurs sources.

Vous portez cette pratique ? Revendiquez-la — les porteurs vérifiés obtiennent un contact public sur la page et peuvent proposer des corrections.

Sources de données

D'où proviennent les informations de cette pratique, et quand.

Pièces jointes

Pratiques similaires qui pourraient vous être utiles