01What happens.
AI reads a limited amount of a file before it stops and answers. What it did not read is missing, and it does not report the gap; it completes the picture from what it has. A slow, restricted or throttled connection lowers the limit. This is how AI works, and it is one route to hallucination and misinformation.
02Why it matters.
If you own the annual report, the material the board is most careful about sits at the back: the notes, the contingent liabilities, the going-concern statement, the pay. Those are the pages a part-read file loses.
An investor asking AI about a provision or a related-party transaction receives an answer built without the page that contains it.
The fault does not sit with AI. A machine fed half the information cannot be blamed for building the other half, and the company that published the file in that form is the one whose name is on the answer.
03The case.
AAAnow read FTSE 100 annual reports as AI reads them. Some files were large enough that the machine read only part of them. The finding is real and the companies are not identified.
The report in question was a design-led document: full-page photography, a long strategic narrative, and the financial statements and notes at the back. The file was large, as such reports are, and nobody had thought of its size as a reporting matter, because size had never been one. It opened on screen, it printed, and it was published in the same form to the regulatory news service and the investor site.
Read by a machine, the file reached the machine's limit before the financial statements. The chief executive's review, the market overview and the divisional narrative were read in full.
The balance sheet, the notes and the remuneration report were beyond the limit and were not read. Asked about the company's pension deficit, AI answered from a reference in the strategic report and stated a figure. Asked about a provision disclosed in the notes, it stated one.
Asked about the chief executive's pay, it returned a number. Each answer was confident, each was built from the front of the report, and each filled in what the back would have said, had the machine reached it.
On a slow or restricted connection, which is what many of the routes AI uses encounter at a corporate site, the limit is lower and the unread portion larger.
The same report on a different day produced a different answer, because the machine had read a different amount of it.
The reporting team had checked the content, the design agency had checked the rendering, and the company secretary had checked the file against the printed proof. File size and page count are found by 2 (TWO) checks in the machine's reading of the file, and neither check had been part of producing the report, because nobody had been asked whether a machine could finish it.
04Exposure.
An answer about the notes to the accounts, built without the notes, is attributed to the company. The board approved a report that a machine could not finish, and the figures the machine supplied in place of the unread pages circulate under the company's name. The report and its form are the board's, and the board answers for both.
The cost is correction of figures the company never stated, and the confidence of the investor who acted on them. It is larger for the matters at the back of the report, because those are the matters where an invented figure does the most damage, and where the investor had most reason to expect the company's own words.
05What AI Readiness changes.
From the inside there is no sign of where the machine stops reading. AAAnow reads the report as AI reads it and evidences how much of the file the machine reaches, which sections fall beyond the limit and what AI answers in their place. With that evidence the company can decide how the report is published so that the whole of it can be read 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 results table AI read had its figures in the wrong columns
The same report, a different fault.
An old annual report answered as this year's after the page that dated it was removed
The report that is read, and read as current, long after its year.
07References.
- Oversized files read only in part: AAAnow, FTSE 100 annual report review, 2026, anonymised.
- That AI completes what it has not read is stated as how AI works and as one route to hallucination and misinformation; it is not attributed to a third-party source.
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.