Exploratory Study: Leveraging AI to Enhance Expert Survey Workflows in Data Governance Assessments

In the Global Data Barometer project, we sought to use existing Large Language Models (LLMs) and Deep Research Agents to automate the process of answering expert questionnaire questions and collecting evidence. This report presents the results from three sets of experiments.

Experiment 1: Automated Questionnaire Answering

This experiment tested the Data Protection Indicator. We built a custom workflow using Dify to process pre-prepared DPL (Data Protection Law) framework documents for various countries. We then used an LLM (specifically, Gemini 2.5 Flash) to answer questions for each indicator based on the content of those documents. The resulting data includes each answer generated by the model, which we then compared with the answers provided by human researchers in the second edition of the GDB.

Experiment 2: Automated Evidence Search

This experiment focused on the Data Protection and Data Sharing indicators. We used Deep Research Agents from Gemini, ChatGPT, and Perplexity to search for relevant framework evidence. Our data records the evidence found by each agent, as well as the results of our manual verification of the accessibility of each evidence link.

Experiment 3: Automated End-to-End Questionnaire Answering

This experiment involved automatically searching for evidence and answering all questions. We tested two GOVERNANCE indicators—Data Protection and Data Sharing—which are based on textual evidence, as well as two Availability Indicators—Public Procurement and Land Tenure. The evidence for the Availability Indicators is more heterogeneous, requiring the inspection of specific data, website functions, and other non-textual sources. Our data records the answers provided by each agent and compares them with the original answers from the human researchers in GDB 2.0.