AI Invoice Extraction with Claude API

Why use AI for invoice data extraction?

Traditional OCR + regex approaches break on layout variations. Every vendor formats their invoice differently. AI-based extraction reads invoices the way a human does: understanding context and structure rather than matching fixed patterns. Claude API is particularly well-suited because of its strong instruction following, structured JSON output, and ability to handle both digital and scanned PDFs.

Build an AI invoice extraction pipeline with Claude API

Step 1: Extract text from the PDF

import pdfplumber

def get_invoice_text(pdf_path):
    with pdfplumber.open(pdf_path) as pdf:
        return "\n".join(page.extract_text() or "" for page in pdf.pages)

Step 2: Extract structured data with Claude API

import anthropic, json

client = anthropic.Anthropic()

def extract_invoice(pdf_path):
    text = get_invoice_text(pdf_path)
    prompt = (
        "Extract invoice data. Return ONLY valid JSON, no markdown.\n"
        "Fields: invoice_number, vendor_name, vendor_address, invoice_date (YYYY-MM-DD), "
        "due_date, po_number, currency, subtotal, tax_amount, total_amount, "
        "line_items (array of description/quantity/unit_price/amount)\n\n"
        "Invoice text:\n" + text
    )
    response = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=1500,
        messages=[{"role": "user", "content": prompt}]
    )
    raw = response.content[0].text.strip()
    if raw.startswith("```"):
        raw = raw.split("\n", 1)[1].rsplit("```", 1)[0]
    return json.loads(raw)

Step 3: Validate the extraction

def validate_invoice(data):
    errors = []
    for field in ["invoice_number", "vendor_name", "invoice_date", "total_amount"]:
        if not data.get(field):
            errors.append(f"Missing: {field}")

    if data.get("line_items"):
        line_total = sum(i.get("amount", 0) for i in data["line_items"])
        if abs(line_total - data.get("subtotal", 0)) > 0.02:
            errors.append(f"Line sum {line_total:.2f} != subtotal {data['subtotal']:.2f}")

    calc = data.get("subtotal", 0) + data.get("tax_amount", 0)
    if abs(calc - data.get("total_amount", 0)) > 0.02:
        errors.append(f"Subtotal+tax {calc:.2f} != total {data['total_amount']:.2f}")
    return errors

Step 4: Batch processing for high-volume AP

from pathlib import Path
import pandas as pd

def process_folder(folder, output="invoices.xlsx"):
    results = []
    for f in Path(folder).glob("*.pdf"):
        try:
            data = extract_invoice(f)
            errs = validate_invoice(data)
            data["file"] = f.name
            data["status"] = "OK" if not errs else "; ".join(errs)
            results.append(data)
            print(f"OK  {f.name}: {data.get('total_amount')}")
        except Exception as e:
            print(f"ERR {f.name}: {e}")
    pd.DataFrame(results).to_excel(output, index=False)
    print(f"Saved {len(results)} invoices to {output}")

Accuracy benchmarks

• Digital PDFs: 97-99% field accuracy on headers. Line items 92-95% depending on table complexity.
• Scanned PDFs at 300 DPI+: 88-94% accuracy. OCR quality is the limiting factor.
• Traditional regex on consistent formats: 95%+ for known layouts, near 0% on new vendor formats.

The AI approach wins on vendor variety -- no retraining or new patterns needed for new invoice formats.

When to use Claude API vs Power Automate AI Builder

Use Claude API when: varied invoice formats, line-item extraction needed, or portability outside Microsoft required. Use Power Automate AI Builder when: already on Microsoft 365 and formats are consistent. The hybrid approach -- Python extraction feeding into a Power Automate approval flow -- gives the best of both.

We implement both. Get a free audit of your invoice workflow and we will recommend the right approach.