Optimising reporting for AI / Section 2:
Why AI needs digital reporting
The underpinning reasons why AI prefers digital-first reporting.
The code within a PDF is not sufficiently structured for digital tools to read.
This makes PDF an unreliable source of any content or data, and creates a high risk of error when analysed by all AI tools.
74.9%
of PDFs did not meet a single one of six separate accessibility criteria.
The need for a digital route map
Imagine you needed to travel to an unfamiliar address. Would a map help? The answer is surely “Yes” and it’s the same for LLMs searching for company information. The more directions that LLMs can be provided with, the more accurate their answers. And the HTML format provides signposting information in its code that PDF cannot.
When an annual report is published entirely on the web, there is more going on than meets the eye. Within the coding of each online page are semantic tags which identify to AI tools what the information is about, the topic section headings, navigation tools for content that appears in tables, accompanying metadata and much more. For over 30 years the web has evolved highly sophisticated optimisation (SEO) to ensure the most relevant information is found and served to users. In effect, a route map of the data within every webpage comes as standard within HTML code. People can’t see this digital ‘map’, but Google and AI depend on it.
PDFs don’t come with that map. PDFs are largely intended for sharing and printing, so they appear consistently whoever opens them – not as a data source. In PDFs, all meaning must be reconstructed by any LLM or AI tool asked to read them and errors get made.
Whilst PDFs are a type of electronic format and can be loosely described as “digital”, they are intended for print or to be viewed on a screen by human eyes. As a result the code within a PDF is not sufficiently structured for digital tools to read.
For example, the text reading order may be incorrectly extracted and ‘visual’ images or diagrams are particularly difficult for AI to read with any accuracy. Corporate reports contain many crucial tables of data, but to an LLM the information within PDF tables is simply visual. At a code level there is no explicit relationship between dates, metrics, rows or columns. AI systems are forced to make a ‘best guess’ about these relationships or look for a more convenient source.
When it comes to AI searches, this makes PDF an unreliable source of any content or data, and creates a high risk of error.
The myth that PDFs may be useful for analysing data with AI tools has been debunked in a number of studies, and is well summarised in this article.
An aspect of reporting that is closely related to AI readability is the digital accessibility of content. The problems the PDF format presents for accessibility were set out in this academic study where 74.9% of PDFs did not meet a single one of six separate accessibility criteria, and this is a pattern also seen across reporting PDFs. Using print design software, accessibility compliance is a major challenge, with most published PDFs falling well short of desired levels of accessibility. Remediating this usually requires complex and slow post-publication ‘fixes’. Executives who focus on accessibility and equality legislation often become increasingly aware of the PDF-first process limitations and, as the FRC has outlined, these are addressed by using digital-first reporting software and processes.
Maximising the probability of AI accuracy
When an LLM is prompted to find information, the process it uses is described as ‘probabilistic’. LLMs don’t really know facts, instead they assign probabilities to potential answers based on the data they can easily access and then supply an answer, based on the best probabilities. Incidentally, this is one of the key reasons LLMs are not infallible, but our main point is that with a map to better find and identify the most relevant content, the probability of AI performing accurately improves dramatically.
Summary content placed onto a corporate website is not necessarily approved by the Board, is unaudited, nor is it tagged with XBRL.
Does the corporate website offer a solution?
Partly, but not wholly. One quick way to improve “AI-friendliness” of reporting content may be to repurpose and place selected elements of the annual report within the corporate website. Whilst this is certainly better than a PDF alone, and it improves the availability of key information, it is only a partial solution, and has downsides. It can be costly, time-consuming and often either delays the overall publication of results or arrives after results are announced.
An approach which places some aspects of a corporate report onto the corporate website might reasonably be described as ‘digital-second’, only providing limited information, secondhand, having been derived from a print-first PDF-based process.
Following a digital-second approach means that AI still cannot easily access the entire definitive and trusted version of the report content in a digital format. Summary content placed onto a corporate website is not necessarily approved by the Board, is unaudited, nor is it tagged with XBRL.
Specialist reporting design agencies are shifting their approach to solve this with full digital-first reporting. They understand the differences when making recommendations about digital-first and digital-second options for their clients. This article sets out the position of one such reporting agency.
These are the underpinning reasons why switching to digital-first reporting improves AI discovery and accuracy. Let’s turn to the evidence in practice.