Stacks of PDFs, one workflow: how AI reads documents so you don't have to

Document processing is a recurring challenge in almost every sector. Think of a financial analyst in the middle of an acquisition, who has to wade through hundreds of pages of annual accounts and contracts. Often from dozens of parties, each with its own layout. Or a logistics coordinator who processes incoming freight documents every day: bills of lading, packing lists, customs forms.

The information in them is crucial, but it does not come out by itself. So it is read, retyped and checked. Manually, every single time.

This work is not complicated. But it takes a lot of time, scales poorly and the risk of errors is higher than you would like. Especially at moments when precision matters most.

An AI solution that automatically reads PDF documents, extracts the relevant data and prepares it for further processing can make the difference. In this blog I explain how such a solution works, which technical choices are involved and what you can expect from it in practice.

What is Intelligent Document Processing?

Intelligent Document Processing, or IDP, is not a new technology. Traditional IDP systems work with OCR (optical character recognition) and fixed templates. They recognize text at predetermined positions in a document. For high volumes with a fixed layout, that works fine.

But as soon as the layout varies, a supplier changes its invoice layout or documents arrive from different countries, it becomes harder.

What makes our solution different is the combination of IDP with a Large Language Model (LLM), the technology behind tools like ChatGPT. An LLM understands context. It does not matter where an amount appears on the page, which language a document is written in, or how the structure differs from year to year. The AI reads the document the way an experienced employee would and pulls out the right information.

That flexibility does require good instructions. The quality of the output depends on how clearly you specify which fields you want to extract, what the expected formats are and when something needs to be checked by a human. That preparation is the real work. The technology then carries it out reliably.

The diagram below shows a general IDP workflow, as described by Gartner:

How we built it

For the infrastructure we chose Microsoft Azure. The main reason is practical: all data stays on servers in Western Europe. For organizations with strict privacy or compliance requirements, that matters.

In addition, Azure offers everything you need in one place: storage, AI processing, a database and the logic that connects it all.

The workflow itself is simple. A user uploads a PDF. The AI receives the document, together with instructions about which information is needed. The model extracts the requested values and converts them to JSON, a format that is easy to read for both systems and people.

The system then automatically checks whether the values are correct and fall within expected limits. What is correct goes straight to the database. What is doubtful, the system presents to an employee, who approves or corrects it with a quick check.

What further increases accuracy is that after the first extraction, the model performs a reflection pass. It receives its own output back as input and assesses it again, field by field, to detect and correct inconsistencies and missed values. That may sound like a detail, but in practice it really makes a difference, especially with complex or irregularly formatted documents.

The results

We validated the solution in a pilot test with seven documents in four languages: English, Dutch, German and French. With seven documents you cannot draw statistically hard conclusions, and that is not the purpose of such a pilot. What you can establish is: does the approach work, are the results usable and where is the friction? On all three questions the answer was positive.

How accurately the model performs in production differs per situation. That depends on the quality of the documents, the complexity of the fields to be extracted and how well the instructions are tuned to the model. That tuning, configuring the solution for your specific document types, is therefore a fixed part of how we approach this kind of project.

To illustrate what the model can handle: we also tested documents in which fields were missing, amounts were written in unusual currency notations or tables ran across multiple pages. In most cases the model correctly indicated that a value was missing or doubtful, instead of making something up. That behavior is at least as valuable as a high score on the easy cases.

We also tested several LLMs side by side, from smarter to more compact and processing entirely within the EU. Depending on what an organization considers important, a suitable choice can be made.

What this means in practice

So what does this actually mean for an organization? It means employees no longer have to do a large part of the manual reading work. Documents come in, the workflow runs and the data is ready in the database. Only the doubtful cases are presented to an employee for a quick check.

That has a few practical benefits. The work scales with the volume: ten documents or a thousand, the workflow does the same. You save a lot of time and the chance of typos or missed values drops considerably.

Also helpful: the solution is not tied to one document type or one sector. As long as you can describe which information you want to pull from a document, the AI can get to work. That makes it applicable to many more situations than just the example we started with.