3 min read.

The vendor contract renews automatically in eleven days. You know that only because you happened to open the right file. There are two hundred more in the folder, all scanned, and you have no idea what is in them.

You have probably already tried the obvious thing. Upload a statement to ChatGPT or Claude, ask a question, see what comes back. Sometimes it works. Often the numbers are subtly wrong, or the assistant tells you it cannot read the file at all. The reason is not that the model is weak. A scanned statement is a photograph of text, not text. A PDF invoice with a line item table frequently arrives with its columns collapsed into one long run of numbers, so the date, the description and the amount blur into a single string that means nothing. Finance runs on precisely the documents that computers read worst.

The fix is a three step habit. None of it involves new software, and none of it involves code.

Step 1: Convert the backlog once, not file by file.

Before AI can answer anything useful, the pile has to become clean readable text. Take a whole batch, statements, invoices, engagement letters, and run it through a converter instead of feeding documents in one at a time as you need them. Clean text here means the structure survives: a statement's headings stay headings, and a table of line items stays a table with its columns intact rather than a scrambled row of digits.

Do a quarter's worth in one sitting. It is less work than it sounds, because you are not reading any of it, you are only changing the format. After that it becomes a two minute habit as new documents arrive. The goal is that conversion stops being a decision you weigh up each time and becomes a step you always take.

Step 2: Give the archive a shape.

Converted files still need somewhere to live. Sort them the way you would actually look something up: one folder per vendor, one per quarter, one per account. If the question you ask is "what did we pay Henderson last year", then there should be a Henderson folder.

This is the step people skip, because it feels like filing busywork with no payoff. It is also the difference between pointing AI at a labelled drawer and pointing it at a cabinet someone has tipped over. Before conversion there was no reason to bother, because the pile was unsearchable either way. Now there is.

Step 3: Ask questions instead of opening files.

Once the archive is text and sorted, the work changes shape. Which invoices from last quarter have no matching purchase order. What are the payment terms across every contract renewing before year end. Where does a closing balance fail to match the following month's opening balance. These are questions you already ask. The difference is that you ask them once, against your own documents, instead of opening forty files and assembling the answer by hand.

One caution worth stating plainly, because it matters more in finance than almost anywhere else. AI gets things wrong, and a confident wrong number is worse than no number at all. Treat the answer as a shortlist, not a conclusion. The value is that it points you at the six files that matter out of two hundred, and because the converted text keeps the original structure, you can open those six and check the figure against the source in about a minute. That is a real audit trail rather than a black box.

None of this needs new accounting software, a migration, or a line of code. It needs one habit: treat converting a document to clean text as a normal part of handling it, the same way you treat filing a receipt. Do that, and the AI tools you are already paying for stop failing on the first scan you hand them.

Start with the worst document you own.

Robin

P.S. Next issue is the researcher workflow I promised last time: an entire literature review moved into Obsidian and NotebookLM, what works, what breaks, and the step that is easiest to skip. Lands in your inbox in two weeks.