How the workflow runs
- A PDF arrives. It lands in a Google Drive folder, comes in as an email attachment or gets uploaded through a simple form.
- Text is pulled out. Digital PDFs give up their text directly. Scanned pages or phone photos go through OCR first.
- AI maps it to your columns. An OpenAI or Claude model receives the text plus a strict list of fields (vendor, invoice number, date, line items, total) and returns structured data.
- Checks run. Dates parse, totals add up, required fields are present, the invoice number is not already in the sheet.
- A row is written. Passing records go into Google Sheets with a link back to the original file. Anything that fails a check goes to a review tab.
The orchestration runs in n8n, Make or Zapier, the same way the lead follow-up demo build chains a trigger, an AI step and a destination, just with documents instead of form submissions.
Why AI instead of a template parser
Older PDF tools need a fixed template for every layout: the total is always in this box on this page. That breaks the moment a new vendor sends a slightly different invoice. A language model reads the document more like a person does, so one workflow can handle dozens of layouts. The tradeoff is that it can occasionally misread or guess, which is why the validation step matters.
Designing the output sheet
Spend ten minutes on this before anything gets built. It decides how useful the result is.
- One row per what? Per document, or per line item? Line items usually need a second tab linked by document ID.
- Exact column names and formats. Dates as YYYY-MM-DD, amounts without currency symbols, a separate currency column if you deal in more than one.
- Source link and timestamp on every row, so anyone can check the original.
- A status column: extracted, needs review, approved.
What makes extraction unreliable
- Low-quality scans, skewed photos or handwriting
- Password-protected or image-only PDFs with no text layer
- Fields that mean different things across senders ("date" could be issue date or due date)
- Very long documents where the key numbers are buried on page 30
Each of these has a fix: better capture guidance for whoever scans, explicit field definitions in the prompt, or splitting long files. Testing on a real batch of your own documents shows which apply to you.
A note on sensitive documents
Medical forms, tax documents and bank statements contain personal data. Keep files in accounts you control, send the model only the text it needs, and confirm your industry's rules before routing regulated records through any AI service. I set up API accounts in your name so data handling stays under your terms.
Getting started
Send me 10 to 20 sample PDFs (redacted is fine) and the sheet layout you want. For example, if someone spends five hours a week on data entry, the automation ROI calculator will show what that costs you in a year. ChatGPT or Claude integration builds start at $150, most finish in 3 to 7 days, and you own the workflow. Start a project with a description of your documents.