1,588 contracts, four sessions — and I slept through most of it.

If you've ever tried to process scanned PDFs at scale, you know the pain. Paid PDF software that chokes on inconsistent formatting. RPA workflows — Power Automate, UiPath — that work until someone changes the template, or a "scan" turns out to be just an image inside a PDF. Older AI models that could extract text, sort of, but couldn't reliably pull dates, terms or references across languages. These tools worked in the demo. They broke in production.

One prompt, no configuration

Recently I pointed an AI agent — one that can run OCR, manage files and keep state across sessions — at 1,588 contracts: scanned PDFs across eight languages, sitting in 470 folders. One instruction: convert every PDF to a structured, searchable markdown file, and rename each with the signing date in a consistent yymmdd format, so every output maps straight back to its source.

Here's what I didn't do. I didn't build a template. I didn't explain the folder structure. I didn't write rules for French dates versus Dutch versus German. I didn't create error-handling logic.

It chose the next step

This is what made it different from any automation tool I've used: it didn't just execute — it decided. Text extraction returned gibberish from a scanned image? It switched to OCR. Couldn't find a signing date in the filename? It tried the document text, then the PDF metadata, then flagged the file for human review — it never guessed. The largest files crashed the process? I asked it to add a skip-and-log step, and it resumed cleanly: no duplicates, no missed files.

An RPA bot follows the rules you write. This agent adapted as it went.

The result: 1,588 PDFs in, 1,588 structured files out. It auto-processed 1,540 and flagged 48 — mostly files where no signing date could be confidently extracted. It never invented an answer. Forty-two hours of compute across four sessions; my time went on queuing the task and checking the next morning.

Two years ago a project like this meant weeks of configuration — mapping every format, building extraction rules per language, testing edge cases one by one. Now those contracts sit in structured files: consistently named, searchable, ready to query like a database.

The documents are the moat

The AI isn't the moat — the documents are. PDFs are containers: they store information but don't share it. Structured files are a foundation everything else builds on. Before you chase AI use cases, fix your document layer. If your work still starts with "open the PDF and find the…", that's where it begins.