Zambia · Lusaka · See the Zambia profile · See the Lusaka profile
Evidence: Quasi-experimental Top 25% 73/100 · Ask Evidence Copilot about this practice
Zambia's national statistics agency and the IGC tested seven LLMs against 1,059 human-coded Labour Force Survey responses: the best models beat enumerators by up to 17.8 percentage points on industry codes, at under $10 a year, projected to save over 170 working days annually.
Zambia's quarterly Labour Force Survey (LFS), run by the Zambia Statistics Agency (ZamStats) with the Ministry of Labour and Social Security, interviews 10,400 households a year. Enumerators assign each respondent a four-digit occupation code (436 possible ISCO codes) and industry code (419 possible ISIC codes) from verbal descriptions collected in the field, a process the survey team describes as time-consuming and prone to error.
In 2025, ZamStats partnered with the International Growth Centre (IGC) and the University of Oxford to test whether large language models could code these responses better and faster than human enumerators.
Researchers built a retrieval-augmented-generation pipeline with few-shot prompting and benchmarked seven LLMs, including GPT-4 Turbo, GPT-5 mini and GPT-5 Nano, against 1,059 LFS responses that ZamStats headquarters staff had independently coded as ground truth. ZamStats and IGC also built an open, no-code interface for the pipeline, targeting full implementation by the end of 2025.
Every LLM tested outperformed human enumerators by a statistically significant margin (p<0.001). The best-performing models reached 60-63% exact-match accuracy on occupation codes (11.8 percentage points above enumerators) and 57-66% on industry codes (up to 17.8 percentage points above enumerators). Classifying an entire annual LFS sample costs under $10 with GPT-5 Nano, and the team calculates the approach could save roughly 43 working days a year for headquarters staff and 130 working days a year for field enumerators, using conservative time estimates.
The authors note real limitations: the 1,059-response sample is too small to compare LLMs against each other with statistical confidence, the enumerator 'ground truth' itself contains coding errors, and results are demonstrated only for Zambia's LFS so far, though the team argues the method could scale to Zambia's four-million-household census and to labour surveys in other countries.
National / regional programmes
Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.
Reusable artefacts from this practice — as published by their sources.
Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.
Where this practice's information was retrieved from, and when.
Mexico
Mexico's statistics institute INEGI is scaling BERT/FastText-based AI coding from its household income survey to the national employment survey, formally …
India
India's National Statistical Office launched a public Model Context Protocol server letting AI assistants query official economic and social statistics …
Sierra Leone
Sierra Leone's National Medical Supplies Agency and University of Pennsylvania researchers built a decision-aware ML tool that raised essential-medicine consumption …
Lithuania
Lithuania's State Data Agency runs a Palantir Foundry data lake connecting ~500 government systems, with 900+ open tables and 238 …
Open full copilot Grounded in cited practices — always check the sources.