Make AI useful with documents and data

Three steps that make document work reliable

1. 11. Get the text outAI tools read text best. Convert the file, then look at what the conversion produced before you trust it.
2. 22. Ask for a fixed shapeSay exactly which fields you want, in what format, and what to do when one is missing. Free-form summaries are hard to check.
3. 33. Check with something better than AITotals should add up, dates should exist, every value should be in the original. Arithmetic and a short script check these perfectly.

Step 1: Get the text out

Many chat tools accept files directly, and for one-off questions that's fine. For repeatable work, converting first lets you see exactly what the AI will read. Microsoft's MarkItDown is one free option: "a lightweight Python utility for converting various files to Markdown for use with LLMs and related text analysis pipelines". It handles PDF, Word, PowerPoint, Excel, HTML, CSV and more.

For our worked example we made a short invoice in Word for a fictional print shop, then ran markitdown invoice.docx. Here's the real output:

INVOICE

Hollow Lane Print Co.
14 Mill Street, Exampletown
VAT no. GB000 0000 00

Invoice number: HL-2291
Invoice date: 3 November 2026
Bill to: Riverside Community Café

Payment terms: 30 days from invoice date

|  |  |  |  |
| --- | --- | --- | --- |
| Item | Qty | Unit price | Line total |
| A5 menus, 350gsm | 200 | £0.45 | £90.00 |
| Table tent cards | 40 | £1.20 | £48.00 |
| Design time (hours) | 2 | £35.00 | £70.00 |

Subtotal: £208.00
VAT at 20%: £41.60
Total due: £249.60

Thank you for your business. Please quote the invoice number with your payment.

Step 2: Write an extraction spec

An extraction spec says what to return, in a fixed shape, and what to do when something isn't there. The shape is often described with JSON Schema, "a declarative format for describing the structure of other data". Some APIs can enforce it: Anthropic's structured outputs "constrain Claude's responses to follow a specific schema", and OpenAI's Structured Outputs "ensures the model will always generate responses that adhere to your supplied JSON Schema". A schema guarantees the shape, not that the values are right.

Extract these fields from the invoice below. Reply with JSON only, no other text.

Rules:
- Use only what the invoice says. If a field isn't stated, use null. Never calculate or guess a value.
- Copy amounts as numbers in pounds, without the £ sign.
- Dates as written in the invoice.

Fields:
{
  "supplier": string,
  "invoice_number": string,
  "invoice_date": string,
  "due_date": string or null,
  "lines": [{"item": string, "qty": number, "unit_price": number, "line_total": number}],
  "subtotal": number,
  "vat": number,
  "total": number
}

{
  "supplier": "Hollow Lane Print Co.",
  "invoice_number": "HL-2291",
  "invoice_date": "3 November 2026",
  "due_date": null,
  "lines": [
    {"item": "A5 menus, 350gsm", "qty": 200, "unit_price": 0.45, "line_total": 90.0},
    {"item": "Table tent cards", "qty": 40, "unit_price": 1.2, "line_total": 48.0},
    {"item": "Design time (hours)", "qty": 2, "unit_price": 35.0, "line_total": 70.0}
  ],
  "subtotal": 208.0,
  "vat": 41.6,
  "total": 249.6
}

Step 3: Check with arithmetic, not another AI

An invoice checks itself: each line's quantity times price should equal its line total, the lines should add up to the subtotal, and subtotal plus VAT should equal the total. This short script tests all of that. You can ask an AI to write one for your own documents; then test the script, as the coding lesson shows.

import json, sys

inv = json.load(open(sys.argv[1]))
problems = []
for line in inv["lines"]:
    if round(line["qty"] * line["unit_price"], 2) != line["line_total"]:
        problems.append(f"{line['item']}: qty x price doesn't match line total")
if round(sum(l["line_total"] for l in inv["lines"]), 2) != inv["subtotal"]:
    problems.append("line totals don't add up to the subtotal")
if round(inv["subtotal"] + inv["vat"], 2) != inv["total"]:
    problems.append("subtotal + VAT doesn't equal the total")
for field in ["supplier", "invoice_number", "invoice_date"]:
    if not inv.get(field):
        problems.append(f"missing {field}")
print("\n".join(problems) if problems else "All checks passed.")

On the real extraction it printed "All checks passed." To be sure the check could fail, we changed the subtotal from 208 to 280, the kind of digit swap that happens in real data entry. It printed "line totals don't add up to the subtotal" and "subtotal + VAT doesn't equal the total". A check you've never seen fail isn't proven.

Before you upload anything

• Check your organisation allows that tool for that kind of document. Invoices, contracts and HR files often contain personal or confidential data.
• Free and paid plans can have different data terms. Read the current terms for the plan you're on.
• Converting with a local tool such as MarkItDown keeps the conversion on your computer. Sending the text to an AI service still sends it.

Try it

What to remember

• Look at the converted text before you trust it. Conversions can quietly rearrange tables.
• A fixed shape and a rule for missing values make AI output checkable.
• Check numbers with arithmetic and text against the original. Don't ask another AI to check the first.
• Our example is one simple document and one run. Scanned, handwritten or complex documents fail more often.