Extracting data from invoices and PDFs with AI: what works and how to start

Supplier invoices, delivery notes, orders, contracts, scanned forms, certificates… In many SMEs someone spends hours every week copying data from documents into another system. It is tedious work, with inevitable errors, and one of the cases where AI delivers the fastest, most measurable results.

In this guide we explain what has changed with generative AI, what you can expect, how to avoid errors and how to get started.

Classic OCR versus AI: what has changed

Traditional OCR turns an image into text. To pull out specific data (“total”, “due date”, “supplier tax ID”), you had to define templates for each document format: where each field is. It works well with a few fixed formats, but it breaks when a supplier changes its invoice layout or a new one arrives.

Today’s AI models (language models with vision) understand the document the way a person would: they know that “Total amount”, “Total due” and “Importe total” mean the same thing, even when they are in a different place. That is why:

• they do not need a template per supplier;
• they also extract table data (invoice lines, items, quantities);
• they understand documents in several languages;
• they can classify the document (invoice, credit note, delivery note, order) before extracting the data.

The usual approach today is to combine both: OCR or text extraction when the PDF already contains text, and an AI model that interprets the content and returns the fields in a structured format (for example, JSON) ready for the ERP.

1. 01

   The document arrives

   Through a mailbox, a shared folder or the scanner.

   invoice_0457.pdf

2. 02

   It is classified

   Invoice, credit note, delivery note or order.

   Type: invoice

3. 03

   AI extracts

   Header and lines, in a structured format.

   { total: 254.10 }

4. 04

   Human review

   It is validated

   If something does not add up, a person reviews it with the doubtful field marked.

   Net + VAT = total ✓

5. 05

   Into the ERP

   With a link to the original document.

   Proposed entry

Validations are the key

An AI model can misread a number or mix up two dates. The difference between a reliable system and one that causes problems lies in what happens after extraction:

• Arithmetic

  Net amount + VAT = total; the lines add up to the net amount.

• Format

  Tax IDs with their check digit, valid IBAN, consistent dates (the due date cannot be before the issue date).

• Cross-checks with your data

  The supplier exists in the master data, the related order or delivery note exists and the amounts match.

• Duplicates

  The same invoice received twice.

What to expect

With reasonably legible documents and good validations, most documents go through without intervention and the rest need a quick review. The exact percentage depends a lot on:

• the quality of the documents (digital PDF versus a phone photo);
• the variety of formats and suppliers;
• the fields you want to extract (the header is easier than the line items);
• how strict the validations are.

That is why we recommend measuring it in a pilot with your real documents before committing to a bigger investment.

Beyond invoices

The same technique works for many other documents:

• Delivery notes and orders: automatic matching with invoices.
• Contracts: renewal dates, penalties, payment terms.
• Customer or onboarding forms, including scanned or legible handwritten ones.
• Supplier certificates and documentation (insurance, approvals), with expiry alerts.
• Job sheets and technical reports.
• CVs, with care: if the extraction is used to screen candidates, it may be a high-risk use under the AI Act.

Integration: getting the data where it belongs

Extraction only really saves time if the result arrives on its own where it needs to go:

• Input: a dedicated mailbox (invoices@…), a shared folder or the office scanner.
• Output: the ERP or accounting software, through its API, a file import or an automation tool.
• Traceability: each record links to the original document, with a log of who reviewed what.

Bear in mind that e-invoicing between businesses will increasingly arrive in a structured format and will not need extraction. Even so, PDFs, documents from foreign suppliers, delivery notes and other paperwork will coexist for years and will still need it.

Privacy and security

Invoices and contracts contain personal and financial data. Before choosing a solution:

• check where the documents are processed and that the vendor offers a data processing agreement;
• use services that do not use your documents to train their models;
• decide how long documents are kept in the extraction system.

How to run a pilot in a few weeks

1. 1

   Choose one flow

   For example, a month of supplier invoices.

2. 2

   Gather a representative sample

   100 to 200 documents, with your usual suppliers and a few difficult cases.

3. 3

   Define the fields and validations

   What is extracted and what is checked before it reaches the ERP.

4. 4

   Measure the baseline

   How much time is spent today and how many errors are caught later.

   Baseline

5. 5

   Run the pilot

   Alongside the current process, comparing results.

6. 6

   Decide with data

   Percentage of documents with no intervention, time saved and errors.

   Pilot report

The whole process, in one picture

Visual guide · Automation

From invoice
to ERP
without typing

What AI has changed, why validations are the key and how to test it with your own documents.

01 · What has changed

From templates to understanding the document

OCR with templates

Each supplier needs its own template. When the layout changes or a new one arrives, it breaks.

AI that understands the document

“Total amount”“Amount payable”“Importe total” total

No templates: it recognises the same field under any name, anywhere and in several languages, tables included.

02 · The key

The validation funnel

• PDF
• Photo
• Email
• Scanner

1. Net + VAT = total
2. Valid tax ID and IBAN
3. Supplier and order exist
4. No duplicates

To the ERP

Most, untouched

Human review

With the doubtful field marked

03 · Not just invoices

The same technique works for…

• Delivery notes and orders
• Contracts
• Forms
• Certificates
• Work reports
• CVs: careful, screening is high risk

04 · The pilot

100–200

real documents from one month, with your usual suppliers and some difficult ones, processed alongside the current process to compare time and errors.

Download the infographic (PNG)

Next step

You can see how we approach this case in data extraction from documents, PDFs and forms and, if you want to automate the whole flow, in AI process automation. If you would rather your team learned to build simple automations, the AI automation with low-code tools course is a good starting point.