Why do AI document conversions silently lose data?
Because language models produce plausible output, not provably complete output: if a document is too long, the output is cut off; complex tables get shortened or shifted; footnotes and annexes are dropped. The result still looks finished – the error only surfaces when someone needs the missing data.
Typical failures
- Output limits: long documents are cut off mid-text.
- Tables: rows go missing, columns merge, tables spanning pages fall apart.
- Layout: multi-column text is read in the wrong order.
- Small print: footnotes, annexes and margin notes are removed as noise.
- Scans: poor originals and handwriting cause misreads.
- “Corrections”: the model fills in or smooths values that were never in the original.
Why nobody notices
There is no error message. The output is cleanly formatted, the JSON is valid, the text reads well. Spot checks rarely find rare gaps – and those are often the expensive ones: a missing clause, a missing amount.
The consequences
Whatever is missing after conversion is missing everywhere downstream: in internal search, in AI assistants and agents, in reports. An AI system then answers wrongly, and nobody sees why.
What helps
Not a better model alone, but verification against the original. How that works is covered in How can you verify that AI extracted all data from a PDF?