ONS's Data Science Campus benchmarked ML autocoding for SIC/SOC survey classification, reaching 73–85% match rates across five surveys versus 50–62% for the rule-based tool, and is prototyping ScannerAI to auto-extract data from ~60,000 household receipts submitted yearly.
50–62 %
Rule-based baseline match rate
73% match / 84% accuracy
ML approach match/accuracy, Census SOC data
70% match / 85% accuracy
ML approach match/accuracy, Labour Force Survey
Details
Maturity
Pilot
Promoter
Office for National Statistics (ONS) — Data Science Campus
Period
2023–2026
Keywords
official statistics, public administration, data science, government efficiency
Context
ONS assigns free-text survey responses to SIC/SOC classification codes across major surveys; its Data Science Campus prototyped a logistic-regression machine-learning alternative to the existing rule-based fuzzy-matching tool.
Results
The ML approach reached 73% match/84% accuracy on Census SOC data and 70% match/85% accuracy on the Labour Force Survey, versus a 50–62% baseline match rate, but performance dropped sharply on a fifth survey (60% vs 84%), and the code (ACE2) remains open-sourced research rather than deployed; a parallel ScannerAI receipt-reading pipeline remains an early-phase, untested prototype.
Implementation
Indicative cost
Low (< €50k)
Time to results
Medium (1–3 years)
Staffing & skills
ONS Data Science Campus
Conditions for success
Sufficient gold-coded training data volume for SIC classification
Common failure modes
Model does not transfer cleanly across different respondent populations (60% vs 84% match on a fifth survey)
Commonly funded by
National / regional programmes
Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.
Do you run this practice?
Claim it —
verified implementers get a public contact pathway and can propose corrections.
Data sources
Where this practice's information was retrieved from, and when.
Mexico's statistics institute INEGI is scaling BERT/FastText-based AI coding from its household income survey to the national employment survey, formally …
Baden-Württemberg built F13, the first AI assistant in productive use across a full German state administration, running on sovereign infrastructure; …
