Commissioned by Ofcom's Economics and Analytics Group — a demonstration of what agentic AI can deliver in a regulatory economics context.
This report documents the analysis of EU Digital Services Act (DSA) Statement of Reasons moderation data submitted by Discord Netherlands B.V. to the DSA Transparency Database, covering four daily extracts spanning September 2024 to December 2025 and comprising 143,726 moderation records in total.
The report is both evidence of what DSA transparency data reveals about Discord's enforcement practices, and evidence of what agentic AI tools can deliver in a regulatory economics context.
Agentic AI for Regulatory Workflows — Ofcom EAG, March 2026The work has a dual character. First, it is a substantive analytical output: a structured three-phase examination of Discord's moderation behaviour across four time points, producing longitudinal metrics on moderation volume, harm category composition, enforcement method, automation rates, and time-to-action. Second, it is a methodological demonstration: the entire analytical workflow was executed using agentic AI, with full reproducibility via a companion Python script.
Profiling the raw CSV structure, documenting field types, missing data patterns, and the relationship between raw records and analytical constructs.
Deriving key analytical variables — visibility actions, interval days, harm categories, automation indicators — with full traceability to source columns.
Longitudinal summary statistics: moderation volumes, harm category shares, action type distributions, automation rates, and time-to-action across four extracts.
Provision actions grew from 51.0% to 83.0% of total actions across the four extract dates, while visibility actions fell from 29.6% to 7.6% — a structural shift in Discord's enforcement approach. The automation rate showed high volatility across the period: 36.7%, 72.7%, 44.1%, 19.1% — with a four-date mean of 43.2%. This volatility is more revealing than the average, indicating unstable automated enforcement rather than a settled operational baseline.
Primary platform adopted. Multi-model orchestration, persistent workspace management, cloud-isolated code execution. Full pipeline from ingestion to structured report.
Independent replication in April 2026. Results effectively identical to Perplexity Computer across all metrics. Now used alongside Perplexity for EAG's ongoing DSA extract review.
Microsoft Copilot Studio — rejected due to architectural mismatch with CSV processing. Python–Streamlit — developed and tested, but superseded by agentic platforms.
This document — a complete analyst-grade output produced through an agentic AI interface, with full section structure and longitudinal charts.
dsa_discord_analysis.py — a self-contained script replicating all results from the raw CSV, independent of any AI platform.