Stop Typing Numbers Off Invoices. Let AI Read Them For You

Every business has a drawer of paper it has not looked at yet.

It is nine at night in Mexico City. Diego runs a three-person import business, and forty supplier invoices are sitting in his inbox. Some are clean PDFs. Some are photos taken on a warehouse floor, slightly blurred, at an angle. Three are handwritten delivery notes from a supplier who still does not own a printer.

Diego does what every small business owner does. He opens a spreadsheet. He opens each file. He types.

Vendor name. Invoice number. Line items. Total. Tax. Again. And again. And again.

By midnight he has typed the numbers off eleven invoices. Twenty-nine are still sitting there, waiting for tomorrow night.

You do not have a discipline problem. You have a format problem.

Diego is not behind because he is disorganized. He is behind because every supplier sends a different shape of document, and no software he owns can read all of them.

A scanner reads pixels. A template reads one layout. Neither reads meaning.

That gap, between a document and the data inside it, is where an entire category of paid software has lived for twenty years: optical character recognition, template-based extraction, outsourced data entry teams typing the same fields off the same documents, forever.

This week that gap got a lot smaller, and it did not cost anyone a subscription to notice. A thread reacting to it went out on X on 23 July, and inside hours the author, an AI educator with over 400,000 followers, wrote: "I'm deleting every paid document extraction tool because of this."

How do you actually automate invoice data entry with AI?

Every version of this workflow, whichever tool you use, breaks into the same four moves: capture the document, extract the fields with an AI model, map them into a structured shape, and sync that shape into wherever your numbers live, a spreadsheet, QuickBooks, Xero, or your own database.

The tools that have existed until now are strong at capture and weak at extract. They read the pixels fine. They fall over the moment a document does not match the template they were built for: a new bank's statement layout, a handwritten form, a scanned receipt taken at an angle in bad light.

That is the piece an open source project called Unstract just made dramatically easier, and it is why that thread picked up traction so fast.

Meet the tool: it builds a schema, not a template

Unstract is free and open source, built by a company called Zipstack, and it has quietly become one of the most-starred document tools on GitHub.

The headline feature is called Prompt Studio. You describe, in plain English, the fields you want: vendor name, invoice date, line items, total. Unstract writes the extraction prompt and the output schema for you, then returns clean JSON. Feed it a bank statement from one bank and it builds a schema. Feed it statements from two hundred different banks and the same schema keeps working, because it is reading for meaning, not matching a layout.

That is why the reaction on X was not limited to bookkeeping accounts. The same pattern, a document that varies wildly by vendor, region, or handwriting, shows up everywhere: a clinic reading insurance claim forms, a logistics firm reading bills of lading in three languages, a lettings agency reading lease agreements, a hiring team reading CVs submitted as scanned photos. Any of it is the same four-step problem: capture, extract, map, sync. Only the document changes.

Isn't this just OCR with extra steps?

No, and this is the part worth understanding before you trust it with real numbers.

Plain OCR reads pixels and hands you a wall of text: correct characters, no structure. Unstract runs the document through a large language model, the same category behind Claude, ChatGPT, or Gemini, and asks it to understand the document the way a person would: which number is the total, which line is a discount, which field is a date versus an invoice number.

Then it does something OCR tools never do. A feature called LLMChallenge runs the extraction twice, through two separate models, and only accepts a value both models agree on. If they disagree, the field comes back empty rather than a wrong number silently sitting in your ledger. Zipstack's own framing of it is blunt: null is better than wrong.

That single design choice is what makes this usable for money, not just a demo.

What it looks like to actually run this, in under ten minutes

In Unstract's own walkthrough of the product, a reviewer with a decade running a small business uploaded a stack of mixed documents, including a handwritten air waybill, and worked through this sequence live:

1. Upload a handful of the same document type. Not a mixed pile: variations of one kind, invoices, delivery notes, or tax forms, whatever you are drowning in.
2. Point it at an LLM you already have access to. Unstract plugs into GPT-4o, Claude, or Gemini, or a local model if you would rather keep documents off the internet entirely.
3. Describe the fields you want in plain English. Prompt Studio writes the extraction prompt and builds an output schema, close to two hundred lines of structure, in under a minute in the product's own demo.
4. Run the extraction and read the accuracy score. In that demo it came back at 100 percent before anything was trusted.
5. Promote the verified result. Every extraction after that checks itself against your approved example, so accuracy compounds instead of drifting.
6. Deploy it as an API or a folder watcher. New invoices land in a folder or hit an endpoint, and structured data comes out the other side, ready for Xero, QuickBooks, Google Sheets, or your own database.

You can try the whole flow for free, no account and no card, at Unstract's own playground, or self-host the entire platform from its GitHub repository if you would rather nothing leaves your servers. It also ships an MCP server and an n8n node, so an AI agent like Claude, or a no-code automation you already run, can call it directly as a tool.

What changes once your documents stop being the bottleneck

A week in, you have one document type running clean: your most repetitive invoice, the one you dread most, out of your hands.

A month in, you have added the second and third document types; delivery notes, receipts, maybe the tax forms your accountant asks for every quarter. The stack that used to sit until midnight now clears itself before you have finished your coffee.

Six months in, the real shift is not the time saved. It is what you do with it instead: pricing conversations, supplier negotiations, the parts of running a business a spreadsheet was never going to help with anyway.

Same business. Same invoices. A completely different relationship with your own paperwork.

If chasing the invoices themselves, not just reading them, is the bigger headache in your business, we covered the agent that chases your late invoices in an earlier guide.