Back to blog

AI for SMBs · Document automation · Business

Invoice data extraction with AI for small businesses

7 min read · October 1, 2026

Your team opens every invoice and delivery note and types the data into Excel by hand. What the DIAN XML already solves, what AI can handle and what a person must always check.

Is your team typing invoices and delivery notes into Excel by hand?

It is Tuesday and month-end close is coming. Your accounting assistant has an inbox full of zipped files from suppliers, a folder of delivery-note (remisión) photos the warehouse sent over WhatsApp, and an open spreadsheet. She opens each PDF and copies the tax ID (NIT), invoice number, date, subtotal, VAT and total. Then she looks for the delivery note to confirm that what arrived matches what was billed.

That is a few minutes per document, sometimes more. An extra zero or a NIT with one wrong digit goes unnoticed until the close, when the numbers do not add up and everything has to be checked again.

If this sounds familiar, it is not a discipline problem. It is a process designed when documents arrived on paper, and nobody has rethought it since.

Why is data still typed by hand if Colombia already has e-invoicing?

When your supplier issues electronic invoices, the invoice is not the PDF. Under the DIAN technical annex adopted by Resolution 000165 of 2023, the invoice is a digitally signed XML file. The PDF is its graphic representation, and the information shown in that PDF must be contained in the XML.

In other words, much of the data your team types already arrives structured in the same email. It gets lost for other reasons:

  • People work from the wrong file. The email carries both the XML and the PDF, and the team uses the PDF because it is the one they can read.
  • Not everything is an e-invoice. Delivery notes, bills from individuals who are not required to invoice, and petty-cash receipts arrive on paper, as photos or as scanned PDFs, each in its own format.
  • Documents come in through every channel: email, WhatsApp, the supplier's portal or the courier's hand.
  • Nobody wrote down the rules. Which field goes in which column, how each expense is classified, what to do when the delivery note does not match. Everyone handles it their own way.

AI helps with the second point. For the first one you do not need AI: you need to read the XML you already have.

What can AI extract from a PDF invoice or a photo?

As of October 2026 there are two kinds of tools for reading documents that do not arrive structured:

  • General models with vision, such as Claude, ChatGPT or Gemini. They read the PDF or photo and return the fields you ask for. Claude, for example, processes each PDF page as both text and image, so it also handles tables and scanned documents.
  • Specialized document services, such as the Invoice Parser in Google Document AI or the invoice model in Azure Document Intelligence. They are trained for invoices and support Spanish documents. Google's also returns a confidence score for each field.

With either one, the key is to ask for structured output: you define the columns (NIT, number, date, subtotal, VAT, total, delivery-note number) and the tool fills exactly those, always in the same format, ready for your spreadsheet or for import into your accounting software.

It works well on headers, totals and clean product tables in digital documents. It struggles more with:

  • Handwritten delivery notes or notes with crossed-out values.
  • Blurry, dark or angled photos.
  • Tables that continue onto the next page.
  • Different number formats: one supplier writes 1.250.000 and another 1,250,000.
  • Stamps or signatures on top of the data.

How can you start automating data entry this week?

The first steps do not require buying anything:

  • Day 1: count the documents. Log how many arrive by type (e-invoice, delivery note, bill from an individual, other), through which channel, and how long each one takes to land in the spreadsheet. If you can, keep logging for the whole week.
  • Day 2: define the output template. The exact columns you need and their format: dates in a single format, amounts without symbols, the NIT with or without its check digit, but always the same way.
  • Day 3: split off the e-invoice path. Ask whether your accounting software or e-invoicing provider already imports received XML files. DIAN's "Facturando electrónicamente" portal also lets you download lists of documents received within a date range, which helps with reconciliation.
  • Day 4: test AI on 30 real documents. Pick delivery notes and receipts from several suppliers, including the messiest ones. Run them through the tool with your template and compare the output with what your team typed.
  • Day 5: write the validation rules. Subtotal plus VAT equals the total, the NIT exists in your supplier list, the invoice number is not a duplicate, and the delivery-note quantities match the invoice.

By the end of the week you know how big the problem is, what the XML solves and what needs AI. To connect the pieces you can use no-code tools such as n8n, Make or Zapier together with a model's API, or the document-intake module in your accounting software if it has one. The right choice depends on volume, your systems and which data it will touch.

How does this fit with Colombian e-invoicing deadlines?

This is where automation stops being only about saving time. For electronic invoices on credit, the buyer registers events with DIAN: acknowledgment of receipt, receipt of the goods or service, and acceptance. Under Article 773 of the Colombian Commercial Code, if you do not dispute the invoice within three business days after receiving the goods or service, it is tacitly accepted.

If delivery notes and invoices are compared at month-end, that window has already closed. A flow that matches both documents the day they arrive warns you in time when you were billed for 100 units and the delivery note says 90. AI flags the difference; a person decides whether to dispute it.

Something similar applies to purchases from individuals who are not required to invoice: there, you are the one who issues the electronic support document (documento soporte), under DIAN Resolution 000167 of 2021. AI can prepare the data from the individual's bill, but someone checks it before it is transmitted.

This is general information, not tax or legal advice. Before changing how you receive and accept invoices, review it with your accountant.

What are the risks, and what should AI not do?

  • Misreading a number. A model can mistake a 7 for a 1 in a blurry photo. Structured output guarantees the format, not that the value is correct. That is what the day-5 rules are for: anything that fails validation goes to review.
  • Deciding for you. AI extracts and proposes. It does not approve payments, accept or reject invoices, or decide the tax treatment of an expense.
  • Sending data where it should not go. Invoices and bills from individuals include names, ID numbers and sometimes bank accounts: that is personal data under Colombia's Law 1581 of 2012. Use business accounts or an API with clear terms on how your data is used, never a free chat.
  • Costs that grow with volume. These tools charge per page or by the amount of content processed, and rates change often. Estimate with your real volume from day 1 and check the provider's current pricing.
  • Throwing away the original. The spreadsheet does not replace the source document. Keep the XML files, the PDFs and the delivery-note photos.

During the first month, review everything the tool extracts. Once you know where it gets things wrong, limit review to documents that fail validation or come back with low confidence.

Is it worth it for your company?

If you receive only a few documents a week, importing the XML files and organizing the folder may be enough. If you get hundreds a month from many suppliers, with paper delivery notes, combining XML, AI and validation rules frees up hours and shows you in time the errors you now find too late.

It is the same logic we use at 77Rentals for contracts: structured data and clear rules first, automation second. AI does the repetitive part and a person reviews what matters.

More articles