Always confirm what was actually read before working with what came back.
Do now: Upload one document and ask what it received — page count, sections, anything unreadable.
Files carry more than their contents: metadata, hidden columns, other windows in the screenshot.File handling is where silent failures live. A misread number, a truncated document or a photograph carrying its location are all invisible in the output, and all consequential.
Confirm receipt before anything else
Ask first: how many pages did you receive, what are the section headings, and what could you not read. This takes one exchange and catches truncation, failed attachments, unreadable scans and password-protected sections — all of which otherwise produce fluent answers about material the tool never saw.
For spreadsheets, ask for the column names, the row count and the data type of each column. Compare that against the file. Hidden columns, merged cells, filtered views and multiple sheets are extremely common and routinely misread.
For images of text, ask it to read the values back to you before interpreting them. Handwriting, low light, receipt printing and unusual layouts produce confident misreadings, and numbers are the values most likely to be wrong and least likely to be questioned.
One exchange confirming what was received prevents the most expensive class of file error.
Prepare the file before it leaves your device
Remove what is not needed. Delete the columns containing personal identifiers, crop the photograph to the part that matters, extract the three relevant pages rather than sending the whole report. Minimisation is faster than it sounds and it is the only control that works if everything else fails.
Check what travels with the file. Documents carry author names, edit history and comments. Photographs can carry location and device details. Screenshots capture whatever else was on screen, including the notification that arrived at that moment and the browser tabs above it.
Watch spreadsheets particularly. A hidden column, a second sheet, or a filtered view that looks like ten rows and contains four thousand are all easy to send without realising. Open the file as a stranger would before uploading it.
Spot-check the extraction
Pick three values from different parts of the file — one near the start, one in the middle, one near the end — and verify each against the original. Errors cluster at boundaries, so checking only the first table is the least informative thing you can do.
Ask how many rows or items were used, and compare against how many the file contains. Silent dropping of rows the tool could not parse is common and produces analysis of a subset presented as analysis of the whole.
Keep the original file unchanged and work on copies. Any cleaning is a series of decisions, and at least one of them will need revisiting.
Check the beginning, the middle and the end. Errors cluster where the reading ran out.