WEBINAR RECORDING · 2026-09
AI in Trade Policy Research and Analysis
In this conversation: Frederic Gonzales (OECD); Ana Elena Sancho Calvino (Global Trade Alert); Ralf Peters (UN Trade and Development); moderator Ben Shepherd.
Speakers from OECD, Global Trade Alert and UN Trade and Development describe AI-assisted research, regulatory data collection and access to evidence. Their message is to formalize tasks, sources and expert verification. Q&A exposes confidentiality and source-coverage limits that a better model alone cannot resolve.
Session information draws on the recording and its original description. Affiliations and policy context refer to the time of the webinar.
FROM THE CONVERSATION
Three takeaways.
What they mean in practice.
Prepared from the session recording using automatic transcription and editorial review. Follow the timestamp links for the original wording and context.
Build a controlled workflow around a bottleneck
Watch 06:24 ↗Gonzales separates discovery, extraction, classification and verification rather than treating AI as one chatbot. His internal STRI interface illustrates easier access to recorded evidence, while retaining human responsibility and acknowledging residual hallucinations.
Policy interpretation: Begin with a specific need, record sources and instructions, and make verification explicit before integrating outputs.
Institutional knowledge improves delegated work
Watch 23:44 ↗Sancho Calvino describes bounded requests, reusable skills, taxonomies, style rules and checklists. Her dashboard example retains human control over indicators and validation while AI supplies implementation. In Q&A she stresses that experts must also identify new agencies and interpret poor scans.
Policy interpretation: Encode substantive standards and update the source map; checking generated answers alone is insufficient.
Reliable screening requires iterative expert feedback
Watch 42:51 ↗Peters reports reduced screening errors in one NTM example after collaboration among domain experts, AI specialists and an intermediary IT team. Q&A clarifies that improvements took weeks and months of model trials, expert review and prompt refinement.
Policy interpretation: Budget for repeated evaluation with experts, not just deployment; distinguish screening performance from completed legal classification.
From the discussion
Confidentiality questions broaden from datasets to internal knowledge: speakers discuss controlled access and resources for in-house solutions (47:44–52:23). Asked about absent online regulations, Peters confirms some requirements still need physical collection and some mandatory standards must be purchased (52:24–55:13). AI cannot remove those gaps. Gonzales closes by emphasizing language and legal interpretation as continuing reasons for expert involvement (01:01:00).
