01What happens.
A table is a set of relationships: this value belongs to this heading, this column, this period. A PDF records where each item sits on the page and, unless it is built to, nothing about which items belong together. Read by a machine, the table becomes a stream, and the relationships are guessed.
02Why it matters.
If you run financial reporting, the table you spent weeks reconciling is the table AI reads first, because it is where the numbers are.
When the machine cannot tell which figure belongs to which period, it does not stop; it assigns them and reports the result. The figure that reaches an investor is right, and it is in the wrong place.
The fault runs across page breaks as well. A table that continues under one heading over 2 or 3 pages loses the heading on each page after the first, and the loss compounds with the length of the report.
03The case.
AAAnow read FTSE 100 annual reports as AI reads them. In reports where the tables held accounting data, table flattening was observed in 23% to 27% of relevant reviewed cases. The finding is real and the companies are not identified.
A segment results table is a typical instance. Across the top run the segment names; down the side run revenue, operating profit and margin for 3 years; the table runs over the page break. To a person, the layout is the meaning. To the machine, the table arrives as the segment names, then a run of figures, then more figures from the next page without the segment names above them.
AI, asked for a segment's revenue, has a stream of numbers and a set of labels and assembles an answer. The revenue it reports is a real figure from the table. It belongs to a different segment or a different year. Asked again by a different person, the machine may assemble it differently.
The financial reporting manager who built the table checked it against the ledger and the audited accounts, and it was right. The designer who laid it out checked it on the page, and it was right, and the board approved it. Nobody had a way of seeing that the relationships between the figures had not travelled into the file, because on screen the relationships are visible to a person in the layout.
Where the file is later ingested into another system, an investor's own AI, a data vendor's pipeline, a research tool, the lost structure travels with the file, and each system assigns the figures for itself. The allocation may differ from one system to the next, and none of them holds the table the company published.
Several faults sit behind the one word. Reading order, heading structure, table structure and tagging each contribute, and the checks that recover a document's structure are where they are found.
The company does not need the detail of those checks; it needs to know that the table it published is not the table the machine holds.
04Exposure.
The figure AI reports is attributed to the company, and the company answers for it in front of the investor who used it. A segment revenue reported against the wrong segment is a misstatement in circulation under a company's name, whatever the file says, and the correction has to reach each person who received it. The board's accountability for the annual report extends to how it reads, because the file is the disclosure.
The cost is the correction, and the confidence of the analyst who built a model on the wrong figure and has to rebuild it. Where the same table feeds the company's own internal AI, the cost is inside the company as well as outside it.
05What AI Readiness changes.
What the reporting team cannot see from the inside is whether the relationships in the table survived the export. AAAnow reads the report as AI reads it and evidences, table by table, where the structure fails and what the machine holds in its place. With that evidence the company can rebuild the file before publication so that the values, headings and periods travel together and check the next report the same way.
Your team makes the changes; AAAnow provides the assessment, the evidence, the guidance and the review.
06Related use cases.
The annual report AI read was not the one the board signed
The board-level consequence this fault feeds.
The impact of AI finding what you cannot see
The same report, a different fault.
The organisation's own AI compounds its defective documents
The same table inside internal systems.
07References.
- 23% to 27% table flattening in relevant reviewed cases involving accounting data: AAAnow, FTSE 100 annual report review, 2026, anonymised.
Next step
Talk it through, in your terms.
The briefing sets this case in your organisation’s terms, and the discussion follows your questions.
Observed, FTSE 100 findings, anonymised. No company is identified.