Scan street invoices

Netherlands Institute for Public Safety

automated workflow in which incoming invoices, receipts or other documents are first read (scanned) and then run through a series of checks. OCR recognition, validation of supplier and payment data, checking for duplicate entries and any discrepancies.

Last change on 7th of November 2025, at 12:43 (CET) | Publication Standard 1.0

• Publication category: Impactful algorithms
• Impact assessment: Field not filled in.
• Status: In use

General information

Theme

Public finance

Begin date

Field not filled in.

Contact information

[e-mail]

Link to publication website

https://nipv.nl

Responsible use

Goal and impact

Automatic pre-processing of invoices. Invoices are still handled by humans afterwards.

Considerations

Advantages

• The AI automatically reads all the details of an invoice (amount, date, supplier) and enters them directly into the system.
• Fast recognition makes processing much faster; invoices can be approved immediately by a human without manual typing.
• Errors are reduced because the AI checks that the figures make sense (e.g. that the VAT percentage is correct).
• It saves staff time, allowing employees to focus on more important tasks such as checking exceptions.

Disadvantages

• The AI only works well if invoices are clear and legible; poor scans or handwritten invoices can give errors.
• Complex invoices (e.g. multiple lines or special discounts) still require human intervention.
• Implementation costs money and requires training the system on its own invoice formats.
• 
  

Why is it reasonably justified?

• When you receive a lot of invoices every day and processing time is a bottleneck.
• When you want to reduce the chance of error.
• If you already have a digital invoice system and can easily link the AI scan to your existing workflow.
• 
  

Alternatives

1. Manual input - safe, but slow and error-prone.
2. Standard OCR software without AI - reads text, but does not understand context; you still need to do a lot of checking.

Human intervention

Yes, there is always human intervention in the AI scanning line.

• Training the model - Humans provide sample invoices and correct the results so that the AI learns which fields (amount, date, VAT percentage, etc.) should be recognised correctly.
• Exceptions and complex invoices - When an invoice deviates from the standard layout (e.g. multiple lines, handwritten notes or special discounts), it is flagged and an employee must manually check the data and adjust it if necessary.
• Quality control - After automatic capture, the algorithm introduces a check step where an employee verifies the data entered before the invoice is finally processed. This prevents errors that the AI might make due to a bad scan or an unknown format.

So although the AI automates most of the data extraction, human supervision remains essential to ensure accuracy and handle exceptional cases.

Risk management

• The AI scanning engine creates invoices faster and with fewer errors, but it can sometimes misread something.
• That's why the system itself checks if the data makes sense; if it doesn't, a human has to check it.
• Everything the AI does is neatly recorded, so you can later see who did what - which helps with audits.
• Invoices remain securely encrypted, and only people with the right permissions can view them.
• Proactis provides backup options and a service guarantee, so you won't be completely without invoice processing.
• At the end of the day, people still have to approve invoice processing.

These measures allow us to leverage the benefits of the AI scanning engine while keeping key risks under control.

Legal basis

Legal obligation: e-Invoicing Directive (2014/55/EU)

Legitimate interest: The organisation has a legitimate interest in efficient, error-free and secure invoice processing

Links to legal bases

• Directive (2014/55/EU): https://wetten.overheid.nl/EUR20140055
• Act of 20 December 2017, amending the Procurement Act 2012 and the Defence and Security Procurement Act in connection with the implementation of Directive 2014/55/EU of the European Parliament and of the Council of 16 April 2014 on electronic invoicing in public procurement: https://www.eerstekamer.nl/behandeling/20180112/publicatie_wet_3/document3/f=/vkkx3ifxcfxq.pdf

Operations

Data

The AI scanning engine reads an invoice and automatically extracts the most important pieces of information: who sent the invoice, the invoice number and date, what exactly was purchased (items, quantities, prices), how much tax has to be paid and what the total amount is. In addition, the system checks that the data matches existing purchase orders and that there are no duplicate invoices. This data is then fed directly into the accounting system, reducing your manual work and errors.

Links to data sources

explanation of proactis: https://www.proactis.com/us/solutions/ai/invoice-capture-ai/

Technical design

1. The invoice is read

• Someone uploads a photo, a PDF file or a scanned paper invoice into the system.

1. OCR (Optical Character Recognition)

• A special programme looks at the image and converts all letters, numbers and symbols into plain text.
• This creates a "digital version" of the invoice that the computer can read.

1. AI and NLP algorithms

• The text is then run through artificial-intelligence (AI) and Natural-Language-Processing (NLP) algorithms.
• These algorithms recognise patterns such as "Invoice number", "Date", "Amount", "VAT percentage", "Supplier" and so on.
• They link each found word or number to a specific field (e.g. the number next to "Invoice no" becomes the invoice number).

1. Verification and validation

• The AI compares the found data with rules we have set:
• The subtotal plus VAT must equal the total amount.
• The invoice number must not already be in the system (duplicate invoice detection).
• The date must be a real calendar date.
• If something is incorrect, the system marks the invoice as "to be checked" and asks a human to look at it.

1. Link with ERP/accounting system

• Once the data is correct, the scan line automatically sends it to the financial system (e.g. SAP, Microsoft Dynamics, Oracle).
• There, the invoice lines immediately appear in the accounting system, ready to be approved and paid. This is done by a human being.

1. Storage and audit trail

• The original scan and extracted data are stored securely encrypted.
• The system keeps a log: who uploaded the invoice, who approved the AI extraction and when the payment took place.

External provider

proactis