Process documents automatically, check them against your own systems, and let an employee decide when there’s doubt.
Trusted by leading organisations
AI document processing, also known as intelligent document processing, identifies incoming documents, extracts the data, checks it against existing records and sends it to the right system. It is designed for organisations where staff retype and check invoices, forms or contracts every day. Data Science Lab builds this end-to-end workflow and can manage it after go-live.
On Monday morning, the accounts payable team’s shared inbox is full again. Every supplier uses a different format, some invoices are scans, and others are missing a purchase order number. An employee looks up the order in the ERP system, checks prices and quantities against the goods receipt, and enters the line items manually. If something does not match, the invoice waits until the purchasing team responds.
For a municipality, it might be documents attached to applications; for an insurer, claim forms. The manual entry is easy to spot, but the time goes into answering three questions: what kind of document is this, does it match what we already know, and who decides when it does not?
An OCR tool that extracts text from a PDF does not answer those questions. You need a workflow that classifies documents, extracts data, validates it, routes uncertain cases for review and writes the results to your own system. We build that workflow as part of our process automation and AI agents service.
Trained extraction models using OCR work well with standardised forms and provide a confidence score for each field. A language model based on generative AI can handle varying layouts and free text, such as contracts, more effectively. We choose the approach for each document type.
A classification model first identifies what has come in, such as an invoice, credit note or delivery note, and splits a PDF containing multiple documents. Because it assigns each document to a known class, you need a class for other documents or a threshold.
An extracted invoice is only useful if it matches the purchase order and goods receipt. This comparison, known in Dynamics 365 as two-way or three-way matching, flags differences in price or quantity before anything is posted.
For each field, the model estimates how likely the value is to be correct. A threshold for each field determines what goes through automatically: a total amount or IBAN needs a stricter threshold than a reference field. You test whether the scores are accurate on your own test set.
If a field falls below the threshold or a check fails, a team member sees the document alongside the proposed values. Corrections are recorded so you can see where the model consistently goes wrong.
The data needs to reach the system where the work happens: the ERP, case management system or policy administration system. We write the data through that system’s integrations and record which document led to which posting.
Tell us which document workflow takes the most time. We’ll see whether AI can help or whether a simpler change will do.
Together, we choose a clearly defined workflow, such as purchase invoices from one entity, and agree which step needs to be faster or more accurate.
We collect representative documents, including poor scans and exceptions, and record the correct value for each field. We use this to measure every version.
We compare a trained extraction model with a language model and choose based on the test set, not a successful demo.
We connect the validation checks to your ERP or case management system and work with the process owner to decide which fields and exceptions must always go to a team member for review.
At first, a team member reviews every proposed result. Only when the measurements justify it do we allow more results to pass through automatically, field by field or supplier by supplier.
We track corrections and scores by document type. We test new layouts, suppliers and model versions against the test set before they go live.
The NFZ Virtual Assistant has significantly improved our clients’ workflow. Thanks to the accurate information and ease of use it provides, made possible in part by our close collaboration with DSL’s developers, it has become an indispensable part of our clients’ daily work.
Documents often contain personal data, from names and addresses to citizen service numbers. The GDPR requires you to process no more data, and keep it no longer, than necessary. At the same time, statutory retention periods apply: you generally need to keep invoices for seven years for the Dutch Tax Administration, or ten years if they relate to immovable property. A scan replaces the paper original only if it is an accurate and complete reproduction, including features that establish authenticity. The extracted fields therefore do not replace the document; the original remains linked to the accounting entry.
This is particularly worthwhile for documents you cannot standardise yourself, such as invoices from many suppliers or forms submitted by citizens. If you use a language model in Azure, the deployment type determines where processing takes place: with Global, this may be in any region where the model runs; with Data Zone in the EU, it remains within EU member states. If your data platform runs on Databricks, processing can also take place there, using functions such as ai_parse_document and ai_extract.
Invoice and delivery note processing is not on the AI Act’s list of high-risk applications. The situation is different when the outcome helps determine decisions about people, such as screening job applications or assessing an application for benefits. Following the postponement under the AI omnibus, the obligations for those applications apply from 2 December 2027. We help you identify the applications, prepare their risk classification and establish arrangements for human oversight, including within AI governance. Your legal counsel or compliance team decides on the legal interpretation.
Any organisation can start an AI pilot. The challenge is what comes next: making AI part of the organisation. That’s where Data Science Lab makes a difference.
Read more about us
10+
8.7
200+
AI document processing automatically identifies, reads and checks documents, then sends the data to the right system. OCR, which converts an image into text, is one part of this. What sets it apart from OCR alone is the validation against existing data and the integration with your process.
For standard forms and common documents such as invoices, trained extraction models are accurate and predictable, with a score for each field. A language model is better suited to free text and varying layouts, such as contracts. If you build a solution yourself using a language model, Microsoft says you do not get confidence scores: you either accept everything, check everything or build your own scoring model.
That depends on scan quality, variation in layout and the field itself. That is why we measure performance on a test set from your own document flow, by field and document type. A confidence score is the model’s estimate, not a guarantee. You should only let fields pass through automatically once the scores prove reliable on your own documents.
The starting point is to keep your existing systems; the integrations needed depend on your environment. The data in those systems does need to be accurate: checking against purchase orders only works if orders and receipts are recorded consistently. For Fokker Services Group, we built a model that predicts the expected invoice amount based on the work delivered and highlights large discrepancies in a dashboard.
We set up processing so that only the necessary fields are retained, restrict who can view documents and define a retention period for each document type. If you use a cloud service, you choose an EU region or EU data zone and record the arrangements in a data processing agreement. You determine with your privacy officer whether a DPIA is needed.
Yes, although the focus then shifts from reading to interpreting. For WUA, we built a topic classification model that sorts respondents’ open-ended answers by subject. For NFZ, we brought together information from contracts with health insurers in a virtual assistant for pharmacists, replacing an annual pharmacy booklet. The key test is whether an answer can be traced back to its source.
If you have few documents, or a single form that you manage yourself, a web form or a fixed rule in your system is often a better choice. If invoices already arrive as structured e-invoices, there is nothing to extract; under the European definition, a pdf or scan is not an e-invoice. And if purchase order numbers are routinely missing, address that in the purchasing process first.
E-mail info@datasciencelab.nl
Phone +31 (0)20 3636 724
Tell us which documents come in and which system they need to go into. We’ll look at whether AI can help or whether a simpler change will do.
Message sent
Thanks for your message. We’ll be in touch as soon as possible.
