INEGI Rolls Out AI Coding to Cut Manual Classification Load in Mexico's Household Surveys

Mexico's national statistics institute, INEGI, runs three major household surveys that require coding open-text answers into standard occupation (SINCO) and economic-activity (SCIAN/NAICS) categories: the quarterly ENIGH income-and-expenditure survey (40,000 records, 260 coders), the ENOE employment survey (30,000 records, 10 coders), and the five-yearly Intercensal Survey (over 1 million records, 600 coders). INEGI describes the manual coding process as labour-intensive, resource-heavy and prone to human error.

Starting with ENIGH, INEGI's data-science team aimed to reduce manual coding workload for occupation and economic-activity variables while maintaining the quality achieved by traditional methods.

The team built a coding pipeline that combines deterministic keyword-matching algorithms with machine-learning classifiers (BERT and FastText), then routes low-confidence cases to human coders for quality validation. INEGI's official 2025 Annual Research Program, one of 28 projects approved from 52 proposals reviewed by its governing board, formally extends the approach to the full ENOE survey.

At a 50% confidence threshold on ENIGH, presented at a December 2024 UN Statistics Division workshop, the models auto-coded 48.2% of occupation cases while matching human coders' classifications 89.4% of the time; applying the same pipeline to the full ENOE survey reached 84.8% occupation-code agreement and 76.7% economic-activity agreement with manual coding.

INEGI's own presentation cautions that model performance is 'strongly linked to the quality and curation of the training data,' and its next step is to build a larger, dedicated ground-truth database and test large language models directly, beyond the BERT/FastText classifiers used so far; as of the 2025 program, INEGI has published its efficiency target but not yet a post-rollout accuracy figure for full production use on ENOE.