Document Automation — invoices, contracts, forms

Every company receives incoming documents: supplier invoices, purchase orders, signed contracts, expense reports, employee forms. Without automation, someone retypes it in an ERP. With AI: upload → structured extraction → automatic validation → feeds the system.

**Typical pipeline**: (1) **ingest** (email, scan, PDF upload), (2) **OCR if needed** (scanned image → text), (3) **classification** (is it an invoice, a PO, a contract?), (4) **extraction** (structured fields: supplier, amount, VAT, dates), (5) **validation** (totals matching, valid supplier, rules respected), (6) **routing** (auto-approved if < 500€ and everything matches; to human otherwise).

**Realistic ROI**: a 3-month document automation project at mid-size company usually produces: 60-80% of incoming documents fully automated, 20-40% to human review (the hard cases), 1 FTE saved on typical AP (accounts payable) team, payback < 6 months if volume > 2000 docs/month.

**Limits to know**: handwritten → unreliable, blurry scans → unreliable, unusual formats (Word tables, exotic languages) → needs examples. Don't promise 99% accuracy from day 1 — start at 85%, iterate on real failures.

**Tools / approaches**: (a) **specialized APIs** (Google Document AI, Azure Form Recognizer, AWS Textract, Amazon Bedrock Data Automation) — pre-trained on common document types; (b) **generic LLM with vision** (Claude Sonnet with PDF input, GPT-4o, Gemini 2.5) — more flexible, needs good prompts; (c) **dedicated startups** (Rossum, Klippa, Docugami, Nanonets).