Guide · AI
How to extract data from PDFs with AI and move it into your systems
Packing slips, purchase orders, order confirmations, inspection reports, completed forms: many businesses receive dozens of PDFs every week, and someone types their contents into a spreadsheet or a system by hand. It isn't hard work, but it takes time and creates typos. AI can prepare a large share of it. This guide shows how to extract data from PDFs with AI in five steps, check it and move it into your systems automatically, and which tools work for it. Invoices have their own guide: How to automate invoice processing with AI.
Which PDFs are a good fit
Good candidates are documents that arrive again and again in a similar form and from which you always need the same details. Typical examples:
• Packing slips: vendor, date, PO number, line items with quantities.
• Purchase orders and order confirmations: customer, items, quantities, prices, delivery date.
• Forms: applications, registrations, claims, inspection reports with fixed fields.
• Certificates of insurance: usually on the ACORD 25 form, with insured party, carrier, policy numbers, coverage limits and expiration dates, for example from subcontractors.
Less suitable are one-off documents that need judgment, such as a contract you want a lawyer to review. AI can prepare a summary there, but extracting fixed fields adds little.
Check first: is the content already data?
Before you use AI, look at the format. Some suppliers send the same information as structured data, for example through EDI or a supplier portal. Where that exists, read the data directly. You don't need AI for that part.
The remaining PDFs come in two kinds:
• Text-based PDFs are created directly by software. You can select and copy the text. Here the text is pulled out and the AI maps it to the right fields.
• Scans are images inside a PDF. Text recognition (OCR) or a model that can read images has to recognize the text first. The quality of the scan then has a big effect on the result.
Extract data from PDFs with AI in five steps
Step 1: Send all documents to one place
Decide where the PDFs arrive: a dedicated mailbox such as packingslips@yourcompany.com, or a folder in SharePoint or another storage service. Documents that land with individual employees get forwarded there. Only what arrives in one place can be processed reliably.
Step 2: Identify the document type and PDF type
The workflow first checks what kind of document it is, for example a packing slip or an order confirmation, and whether it's a text-based PDF or a scan. Then it picks the right path. Documents it can't classify go to a person.
Step 3: Extract fields against a fixed schema
For each document type, write down which fields you need and in what format: date as a date, quantities as numbers, item number as text. The AI gets this list and returns the values in exactly that form. Tell it to leave a field empty when it isn't in the document instead of guessing. For tables with line items, the AI extracts each row as its own entry.
Step 4: Run the checks automatically
Before any data moves on, the workflow checks what it can:
• Are all required fields filled in?
• Does the PO number or item number exist in your system?
• Do totals and line items add up?
• Does the delivered quantity differ from the quantity ordered?
Anything that fails a check goes to a person together with the original. Everything else moves on.
Step 5: Hand off and file
The checked data goes where it is needed: into an Excel sheet, a SharePoint list or, through an integration, into your inventory or ERP system. The original PDF is filed and linked to the record, so every value can be traced back.
Before you build
Collect 20 real PDFs per document type, including the bad scans and the unusual layouts. Use them to test the workflow before it goes live. With clean sample documents, every workflow looks better than it is.
Which tool to use
The right tool depends mostly on what you already use. For a detailed comparison, see n8n vs Make vs Power Automate.
Microsoft Power Automate. The automation service in Microsoft 365 reads PDFs with AI Builder. For your own document types, you train a model with a few sample documents. For invoices, receipts and ID documents, there are prebuilt models. A flow with AI Builder needs Power Automate Premium or a Process license for the flow. Since November 1, 2026, new or renewed licenses no longer include AI Builder credits. Credits that came with existing licenses remain available through the end of the contract term. After that, you pay with Copilot Credits. According to Microsoft, one page with a custom document model costs 8 Copilot Credits, which is $0.08 with pay-as-you-go billing.
n8n. n8n is a workflow tool that can also run on your own servers. The "Extract from File" node pulls the text out of a PDF. An AI step then returns the fields according to your schema. For scans, you also need text recognition or a model that can read images. n8n is a good fit if you want to decide yourself where documents go. If they shouldn't leave your infrastructure at all, self-host n8n and use a model you host yourself. Otherwise, choose an AI provider whose data processing terms fit your obligations.
Make. Make has its own AI modules that pull text and data out of documents. Make passes data to many cloud apps and works without code. According to Make, its document extraction modules run on Microsoft Azure in the EU, even if your organization uses a US data zone. If your data has to stay in the US, check this for the modules you use.
Ready-made services and your own system. There are specialized document extraction services, and some inventory and ERP systems already read certain documents on their own. If your system can do that for your document type, it is often the simplest route. A custom workflow pays off when several document types come together or the data goes into several systems.
What a person should decide
The AI extracts and prepares. A person should see everything that fails a check, every difference between order and delivery, and anything with legal or financial consequences, such as changed prices or delivery dates in an order confirmation, or an expired insurance certificate. Decide in advance who gets these cases and how quickly they are handled. Otherwise they pile up somewhere new.
Where this reaches its limits
Handwriting, crooked or faint scans and tables that run across several pages push the error rate up noticeably. It also gets harder when you receive many different layouts with only a few documents each. And AI can misread a number while sounding confident. That's why the checks in step 4 aren't an extra. They're the core of the workflow. If documents contain personal data, for example in forms, they belong only in tools with the right contractual terms. More on that in Can our team put client data into ChatGPT?.
How aenuma helps
aenuma builds this workflow in your environment, for example with Microsoft 365, n8n, or Make, and passes the data to your systems where they offer an integration or API. We test it with real documents from your day-to-day work. Together, we agree in advance which steps run automatically and where a person reviews. Then we introduce the workflow to your team and, if you want, keep supporting it.
Frequently asked questions
Can't I just upload PDFs to ChatGPT?
For individual documents, yes. For recurring volume, the important parts are missing: a fixed schema, the checks and the handoff to your system. Also, documents with client data belong only in an account with the right business terms, not in a personal one.
How accurate is the extraction?
That depends heavily on the documents. Clean text-based PDFs with a fixed layout can usually be read reliably, poor scans much less so. Measure it with your own examples before you rely on a number.
What does extraction cost?
With Power Automate, you pay per page in Copilot Credits as described above, plus the license for the flow. With n8n, you pay the AI provider based on usage, plus either your own server or an n8n Cloud subscription billed by execution. With Make, the AI modules consume credits. How much that adds up to each month depends on the number of pages.
Do we need a new document management system?
Usually not. The workflow works with the mailbox and storage you already use and files the originals there.
Sources
Overview of licensing, AI Builder (Microsoft Learn)
End of AI Builder credits (Microsoft Learn)
Extract from File (n8n Docs)
AI providers and data zones in Make (Make Help Center)
By Joshua Schütt, founder of aenuma. Last updated: November 2026.
Which PDFs does your team still type in by hand?
Tell us which documents arrive at your company and where the data needs to go. We will tell you what can be extracted automatically and where a person should review. Not sure PDFs are what takes your team the most time? The Time Sink Check shows which task makes a good first workflow.
Book a call