Knowledge
How AI reads an annual report – and what this means for your reporting
4 min
read
Investors, analysts, media professionals and other stakeholders are increasingly using LLMs and agentic AI to find out about companies, key figures and developments. For annual reports to be reproduced accurately in AI-generated answers, they need to be accessible as HTML-based online reports, clearly structured and machine-readable.
Published:
How has Company ABC’s revenue developed? And what were the main reasons for this? What does its CO₂ footprint look like? Investors, media professionals, analysts and other stakeholders are increasingly asking agentic AI questions like these. It accesses relevant sources and uses a large language model (LLM) to formulate an answer. For that answer to be reliable, the system must not only find the annual report, but also access its content, recognise relationships and correctly attribute statements.
AI does not read like a human
When responding to a query, an LLM draws on a website’s structured data and uses it to generate an answer. The passages that appear relevant to the specific question are decisive. This changes the requirements for annual reports. A key figure should not stand alone. It needs a semantic structure: the financial year, the unit, the comparative period, the definition and, where relevant, an explanation for any change. The more clearly these relationships are represented in both the content and the code, the lower the risk that a system will shorten statements or attribute them incorrectly.
What does this mean for reporting?
For digital use, HTML therefore offers significant advantages over PDF reports: content is responsive, linkable, searchable and easier for machines to interpret. The contemporary solution is consistent multi-channel reporting. The reviewed content is created and approved once, then transferred to the appropriate output channels. This ensures that statements, key figures and definitions remain aligned across the PDF, online report and other touchpoints.
Online reports do not require an additional manual editorial process. What matters is that approved content is distributed in a structured, consistent way across all channels. This is precisely where L•Sync comes in: the platform connects the central source with online reporting and automatically converts approved content into digital, machine-readable formats.
From report to AI-generated answer in five steps
1. The content must be accessible
It may sound obvious, but this is the essential prerequisite: before an LLM can take content into account, it must be able to access it. Discoverability starts with a functioning technical foundation. The page must be publicly accessible, technically easy to read and must not be blocked for relevant crawlers.
2. Structure makes content understandable
Headings show hierarchies, tables connect values with labels, paragraphs provide context and links establish relationships. Semantically structured HTML helps distinguish, for example, revenue from a margin, the current financial year from a prior-year figure, or a forecast from an actual value.
3. The system searches for relevant passages
When asked a question such as “How has profitability developed?”, the entire report is not loaded. The system searches for relevant sections, key figures, tables and explanations. Content that is clearly labelled, precisely worded and understandable in its own right can be identified more easily as a relevant source. Importantly, tables must also be represented in a structured way in HTML.
4. The LLM formulates an answer
The LLM uses the information it has found to formulate a summarised answer. Companies can increase the likelihood that their reviewed information will be used as a primary source: a fully structured HTML report makes it easier for LLMs to access, interpret and correctly attribute content. However, being cited as a source cannot be guaranteed. LLMs also answer questions using third-party sources – with the risk that statements may be shortened, misinterpreted or out of date.
5. Currency and trust matter
An annual report always relates to a specific period. Terms such as “in the reporting year”, “as of 31 December” or “compared with the previous year” are therefore important reference points. Publication and update dates, information about the responsible parties and links to further sources also support accurate interpretation.
Interested in learning more?
Get in touch with us.
Contact
David Adlhoch
Consultant
T +41 44 268 12 06
contact@lsync.ai