Turning PDFs into Structured Data with AI

Extraction is one of the highest-value AI tasks in a back office, and one of the easiest to get subtly wrong at scale.

Step 1: separate text from layout

Try plain text extraction first. If the document is a straightforward report, it is enough. Scanned pages and multi-column layouts need OCR plus layout detection, and tables need a dedicated extractor — they lose their structure otherwise.

Step 2: define the schema before prompting

Write the exact output shape you want: field names, types, formats. Include a field for anything uncertain so the model has somewhere to put ambiguity instead of guessing.

{
 "invoice_number": "string",
 "issue_date": "YYYY-MM-DD",
 "total_amount": "number",
 "currency": "string",
 "line_items": [{"description": "string", "amount": "number"}],
 "confidence_notes": "string"
}

Step 3: instruct on missing data

Add an explicit rule: if a field is not present in the document, return null and describe why. Without this, models invent plausible values, which is the failure mode that does the most damage downstream.

Step 4: validate in code, not with the model

Check that totals reconcile, dates parse and required fields are present. Arithmetic verification catches a large share of extraction errors and costs nothing to run.

Step 5: route the uncertain cases to humans

Send low-confidence extractions to a review queue. A pipeline that is 95 percent automatic with a clean exception path beats one that claims 100 percent and quietly corrupts records.