With AI, structured data wins the day
A new academic study led by Professor Ariel Markelevich of Suffolk University, published 3 July, puts artificial intelligence (AI) under the microscope by testing how well it reads financial statements. Spoiler alert: AI does better with a structured diet. Using 5,000 annual reports from the US Securities and Exchange Commission’s EDGAR system, the study compared AI performance across text, HTML and XBRL formats.
The researchers asked a large language model to extract 26 standard accounting metrics from reports spanning 2014 to 2023. The results were telling. XBRL-formatted data significantly outperformed unstructured text and HTML across the board, particularly when metrics came from footnotes or complex company reports. And the kicker? Scaling errors, a common AI pitfall, dropped from over 8% in text to just 0.11% with XBRL.
Why does this matter? Because as AI becomes a bigger player in financial analysis, the quality of its output depends heavily on the input. This study confirms what the XBRL community has long known: structured, standardised, contextualised data is the bedrock of accurate digital analysis.
As AI tools evolve, pairing them with high-quality XBRL data unlocks serious analytical power. Better structure, better outcomes, better insights.
The full study is well worth a read – see here on the XBRL US blog.
