Are your disclosures AI-ready? How to format corporate reports for the machine era

Key Takeaways:

• The audience has shifted: Investors and proxy advisers increasingly rely on AI to analyse corporate disclosures, making machine-readability just as vital as visual design.
• Accuracy depends on document structure and content quality: An effective AI analysis of a remuneration report requires both seamless data extraction (driven by clean document architecture) and clear comprehension (driven by standardised, unambiguous language).
• Simple rules can make reports AI-Friendly: Companies can improve AI reading accuracy by amending tables, headings, charts and footnotes.
• Improving AI friendliness can break more traditional disclosure rules: Decisions will need to either prioritise the human or AI reader.

Guerdon Associates provides an AI readability checklist it applies to remuneration report reviews.

AI as an investor tool

With AI becoming increasingly adopted by businesses, Australian proxy advisers and investors are likely to follow the lead of their US counterparts and develop AI models to extract data to make voting decisions.

The accuracy of the data the AI models extract will heavily depend on the remuneration report’s ability to be read by a machine.

Pre-AI public disclosures are designed for human consumption—favouring visual engagement and unique terminology to get a point across and differentiate a company’s approach from others. This format often creates obstacles for machine readability.

This article describes some of the factors that will impact the AI-friendliness of your remuneration report.

The stages that influence AI model data extraction

There are two factors determining whether AI answers user queries well based on a document.

1. The extraction of document information
2. The comprehension of the extracted information

Extraction relies on a complete data source. If information is not found, AI is highly likely to generate flawed conclusions.

The comprehension step is equally important. The clearer the document messaging is, the less likely AI applies its own logic and statistical assumptions to understand the document and potentially replace actual information with “typical” placeholders.

Extraction depends on the document structure while the comprehension of the extracted information depends on the document content quality. 

A High-Level Checklist for AI Readability

The following checklist is a non-exhaustive sample of considerations for optimising a remuneration report for AI readability:

1. Make headings literal, consistent and nest them correctly: Headings should be descriptive of content and consistent with report styling.
2. Talk to the team involved in the typesetting and publication of your report: They will need to ensure that tagging and final format supports AI’s comprehension. (Ask us about some of the rules, we are happy to talk to them for you!)
3. Avoid merged or split cells in tables: Merged cells break the strict grid structure AI models use to map data points.
4. Do not rely on footnotes to convey meaning to body text. AI often struggles to link superscript markers (like ¹ or *) to the corresponding text at the bottom of the page.
5. Use standard terminology: Stick to standardised acronyms (KMP, STI, LTI, NPAT, TSR, FAR) rather than bespoke terms.
6. Avoid or mitigate the impact of charts: It is hard for AI to extract information from charts so they may need supplementary disclosure.
7. Consider other information sources: Repetition of key messages in multiple sources may assist models to be confident in the information’s veracity.
8. Remember not everyone will use AI: Consider the human readership when making changes.

What this means for companies

Investor reliance on AI for AGM voting decisions will only increase. Remuneration reports that have been rolled over year after year may no longer be appropriate.

Guerdon Associates now has an AI checklist it will be employing on any remuneration reports it reviews. For questions on the checklist, please contact us.