Make a pile of PDFs answerable

A folder of PDFs is not data. It is a hundred documents you can open one at a time, which means any question spanning more than one of them is a job you schedule rather than a question you ask.

The change here is small and the difference in what you can ask is not.

Upload the pile

Drag them into Documents. Statements, reports, signed contracts, scanned invoices, board packs. They are parsed and indexed as they land.

Do not sort them first. Sorting is a job you are doing because the pile is unaskable, and it stops being necessary once it is not.

Ask the question you could not ask before

The test of whether this worked is not whether it can summarise one document. It is whether it can answer something that lives across several.

Across all the supplier statements in this folder, which suppliers did we
pay more to this year than last, and by how much? Show your working per
supplier and name the document each figure came from.

That question has no answer in any single file. There is no search that finds it. It is the kind of thing that would previously have been an afternoon with a spreadsheet.

Always ask it to name the document each figure came from. Numbers extracted from a scan are the single most likely thing to be quietly wrong, and a figure with a source is checkable in ten seconds while a figure without one has to be trusted. If it will not tell you where a number came from, do not use the number.

Check three figures by hand

Pick three, open the source, and confirm. You are testing the parsing, not the reasoning, and parsing is where scans fail: a column misread, a total taken from the wrong row, a negative dropped.

If all three are right, the rest are probably right. If one is wrong, look at whether that document is a scan rather than a real PDF, because that is usually the pattern.

Turn the recurring questions into a job

Once you know which questions you ask every month, they stop being questions.

On the second of each month, read any new statements added since last
month and update the supplier spend summary document. Flag anything that
moved more than 20 percent against the prior month, with the source.

What good looks like

The pile stops being a filing problem and becomes something you interrogate. The first real sign is that you ask it something you would not have bothered to find out before, because finding out used to cost an afternoon.

What this does not do is turn a scan into a spreadsheet you can trust blindly. Treat extracted numbers as sourced claims to spot-check, not as a finished ledger.