IDP Agents: extract exactly the data you need from your documents

Result first: Ten fields from a scanned invoice

Hello and welcome to the PaperOffice AI Academy. This video is about IDP Agents. With this, PaperOffice extracts exactly the data you need from your documents, consistently into the same fields for every document. Invoice number, purchase order number, amounts, IBAN: If you type such data document by document, you waste time and make mistakes. First the result. This invoice from Veldtmann came in as a scan. Our IDP Agent extracted ten fields from it: invoice number, invoice date, supplier, purchase order number, net amount, tax amount, total amount, currency, IBAN, and due date. Hover over a field, and the view on the left jumps to the location in the document from which the value originates. Here is the IBAN from the payment information. This way, you can check every value at a glance. And because every invoice gets the same fields, an export turns them into an Excel table, with each invoice in one row. Here is how you can do it yourself: create an agent from a sample document and a template, define the fields and test them, assign the agent to a document type, process existing documents afterwards, review the results and export them.

Chapter 1: Overview of IDP Agents

PaperOffice already reads the data out of each document automatically when you upload it, depending on the AI-DMS tier of the workspace. An IDP Agent is your own saved extraction for recurring documents: you decide which fields you need, what they are called and in which format the values come out. Your IDP Agents can be found under Document AI. The list is currently empty. An IDP Agent is always created from a sample document. There are three steps: select the document in the workspace, create a new IDP Agent and choose a template from the catalog. Afterwards, you adjust the fields and prompts in the agent editor.

Chapter 2: Create agent from sample document and template

In our Workspace IDP-Demo there are five supplier invoices. As an example document we take the invoice from Veldtmann: select it, then Actions, Agents and workflows, IDP, and Create new agent. The template catalog now opens, sorted by categories: Finance and Accounting, Logistics, Legal, IDs and more, also for e-invoices like XRechnung, ZUGFeRD and Factur-X. With Show recommended only you see only the templates that match this document. For incoming supplier invoices, we use Invoice Quick Capture Plus. The template comes with the most important fields: invoice number, invoice date, amounts, supplier and currency. It also comes with the recommended model. Create agent. PaperOffice reads the sample document completely now: text recognition, fields, completion. This takes a moment; I will fast-forward. This is the agent editor. On the left is the sample document, and on the right first comes the analysis: what kind of document it is and what the template found in it.

Chapter 3: Define the fields

With Continue and review detected fields, we move on to the fields. PaperOffice immediately starts a first test run using the sample document. Each field displays the found value and next to it a small snippet from the document so that you can immediately see where it comes from. Expand a field to see its prompt, which is a short instruction for the AI on where to find what. You can edit this text directly here. At the top of the editor, you see the model. The template recommends Basic Pro Max, text recognition without image analysis. Since we also receive mobile phone photos and scans with spots, we choose Premium Pro Max: it additionally examines the image of the page and thus reads difficult scans as well. For the payment, we still need fields. New field: Label IBAN, plus a technical Key, and for Field type there is a dedicated IBAN field type. In the Prompt, it says where to find the IBAN. In the same dialog, you specify which pages to search and how the value is validated, using a pattern or a list of allowed values. The IBAN is a mandatory field. If it is missing from a document, PaperOffice marks the field as missing, and you can find the document using the filter Manual review required. A review task for your team is created when you set up a HITL agent for this purpose. Next comes the due date, of type Date. Here, Soft HITL review is enough if it is ever missing. After each change, test the agent again; only then can you save it. And now the new fields are also displayed.

Chapter 4: Test reliability and save

Does the agent also reliably produce the same result? For this, there is the multiple test. We let it run five times. A run is considered successful only if every field has a value. After five successful runs, it will say: Production safe. Under Output and review you can set how the values are written and when a person needs to check them. In the Extracted data tab, you can view the test run results as a tree, JSON, or table and download them. And with Export as shareable template, you pass the agent on as a file. Another team then imports it in the template catalog. Now save. The agent is named “Supplier invoices” and is immediately active.

Chapter 5: Assign the agent to a document type

To ensure the agent runs automatically in the future, we connect it to a document type. Under Document Types, we open the Invoice type. On the Default IDP Agent tab, we choose Supplier invoices and save. From now on, the following applies: if the AI-DMS processing recognizes a new document as an invoice, PaperOffice automatically queues the agent, in every workspace with AI-DMS.

Chapter 6: A new invoice is read automatically

Let’s give it a try. A new invoice arrives, this time from Iberica in Bilbao, in Spanish. By the way: in the upload dialog, you can also directly select an agent for exactly these documents. Here we let PaperOffice decide for itself. First, the AI-DMS processing recognizes the invoice type, and then our agent runs. I am fast-forwarding. Here is the result: the same ten fields as for Veldtmann, even though the invoice is in Spanish. Pedido cliente becomes the purchase order number, Vencimiento becomes the due date, and the Spanish IBAN is in the IBAN field.

Chapter 7: Existing documents in one go

And what about the five invoices that were already in the workspace? That assignment does not apply to them retroactively. We process them in one go: select all of them, then Actions, Agents and workflows, IDP, Apply existing agent. Select an agent and start IDP Agent. PaperOffice queues all documents and processes them in the background. You can see the status under Processing. I am fast-forwarding.

Chapter 8: Review the results

In the document itself, the values are located under Data Fields, with each field accompanied by a small snippet from the document. Here is the invoice that someone photographed with their mobile phone. All fields are present here as well, and hovering over them jumps the view to the source location. If a value is incorrect, you can change it directly in the field, and PaperOffice saves immediately. For a single document, there is an Apply agent button here at the top. And the freight invoice with the coffee stain: invoice number, purchase order reference, amounts and the British IBAN are correct.

Chapter 9: Export to Excel

Finally, the list for accounting: select all the invoices, then Actions, Share and export, Export data. As columns, we choose the fields of our agent, in Excel format. Each invoice is listed on a separate row in the file, with the same columns. And now something stands out: three invoices from Veldtmann refer to the same German IBAN, while the fourth one lists an account in Lithuania. How PaperOffice automatically checks and stops this is shown in the episode Audit and Quarantine. In summary: an IDP Agent is your own extraction setup for recurring documents. You start with a sample document and a template, add the fields you need, and test until the agent delivers reliably. Linked to a document type, it runs automatically for every new document, while existing documents are processed in one go. You review each value at its source location, and with an export, everything ends up in Excel. Our n8n video shows you how to call an IDP Agent from your own software or from n8n. The next video is about Workflow Agents. Until next time at the PaperOffice AI Academy.