AI: the new reader of annual reports
Annual reports have traditionally been written for people such as investors, analysts, shareholders, regulators, employees and journalists. Today, they are increasingly also read and analysed by large language models. This creates a new audience for corporate reporting and makes the way information is structured increasingly important.The annual report is becoming an AI interface
AI INPUT Generative AI is becoming a structural part of how corporate disclosure is consumed.
Generative AI is increasingly used to find and summarise information from annual reports. As a result, investors, journalists and other stakeholders may encounter information through an AI generated answer before they read the annual report itself. This means that AI is increasingly becoming an intermediary between companies and the people using their corporate information. 3
A 2025 study of ChatGPT’s answers to corporate-reporting questions found that 58.5% of the citations the model produced linked directly to annual reports, making them the largest single source category in its responses. 3
This shows that annual reports are an important source of information for generative AI. How easily AI can access and interpret a report therefore matters. The FRC notes that AI tools increasingly use the machine readable XHTML version of annual reports and recommends making these files directly accessible, rather than only providing them as part of a download package. 12
AI performs better with HTML reports
A 2025 study by the University of Applied Sciences St. Pölten, HHL Leipzig Graduate School of Management and reporting agency Nexxar examined how ChatGPT uses annual reports when answering questions about companies. The study compared how effectively ChatGPT finds and uses information from annual reports published in different formats. 3
AR CITATION SHARE
HTML VS PDF
ANSWER ACCURACY
SOURCE DISPERSION
The findings have a clear implication. When an annual report is available in HTML, AI is more likely to use the report itself as a source. When information is only available in a PDF, AI is more likely to rely on other sources, such as press releases, news articles or analyst reports. Making annual reports easily accessible to AI therefore increases the likelihood that the company’s own reporting is used as the primary source.
Structured data improves AI accuracy
The format of corporate reporting also affects how accurately AI can extract financial information. A study covering 5,000 annual reports compared AI results using plain text, HTML and XBRL.
The researchers measured how often AI returned an incorrect financial figure. The overall error rate was 18.24% for plain text and 15.75% for HTML. With XBRL, the error rate fell to 9.19%.
The study also looked at the types of errors AI made. One important source of mistakes was scale. For example, AI could identify the right number but fail to recognise that it was reported in thousands or millions. XBRL largely eliminated these scale errors because this information is included in the structured data.
The findings show that structured reporting does more than make information machine readable. It can also help AI interpret financial information more accurately. 13
|
LINE ITEM |
NO CONTEXT |
PLAIN TEXT |
HTML |
XBRL |
|
Cash |
100% |
10.7% |
8.4% |
6.5% |
|
Inventory |
95% |
14.2% |
9.1% |
5.8% |
|
Total assets |
92% |
12.8% |
7.6% |
4.9% |
|
Operating cash flow |
98% |
15.4% |
10.2% |
6.1% |
|
Hallucination rate |
~50% |
7.8% |
4.2% |
<2% |
Two things stand out. First, the structure of the input matters more than the model. Second, the gap between XBRL and the best alternative is large enough that, for any disclosure where exact numbers are part of the value of the report, structured data is no longer optional.
Consistent information across the report
COHERENCE Reports often disagree with themselves. Structure reduces — but does not eliminate — those disagreements.
Metadata provides context and meaning to digital information. It helps machines understand what information represents, how it is classified and how different pieces of information relate to each other. Good metadata therefore makes corporate information easier to find, compare and analyse.
A 2026 study analysed 700 financial and ESG disclosures from 350 listed companies across five countries. The researchers assessed metadata quality based on completeness, consistency and semantic coherence. These measures were combined into a Metadata Incoherence Score, where a higher score indicates more inconsistencies or missing information.
The results show a clear relationship with reporting format. PDF reports had the highest average incoherence score at 8.43. This fell to 6.39 for HTML and 5.40 for XHTML/XBRL. Reports containing ESG information also showed more metadata inconsistencies than reports without ESG content.
Structured reporting does not guarantee that the information itself is correct. It does, however, make information more consistent and easier for companies, auditors, regulators and AI systems to check and analyse. 14
From a document for humans to a data source for humans and machines
The direction is clear. Annual reports are evolving from documents primarily designed for people to read into structured sources of information that can be used by both people and machines. 15
Old world
-
Annual report as PDF document.
-
Designed for human visual consumption.
-
AI is forced to look elsewhere for numbers.
-
Errors and hallucinations are common.
-
Companies have little control over the narrative AI produces.
New world
-
Annual report as semantic data source.
-
Designed for humans and machines together.
-
AI reads the report directly.
-
Errors drop sharply, hallucinations approach zero.
-
The company’s own filings shape the narrative AI produces.
“When the data is structured, the model reads the report. When the data is locked inside a PDF, the model goes elsewhere — and the company loses control of the narrative.”