AI document processing: from incoming pile to structured data
Invoices, contracts, forms, emails, receipts, reports: every Dutch business processes hundreds to thousands of documents every day. Many of them still end up manually in spreadsheets, ERP fields or mailboxes. We build AI pipelines that automatically read, classify and validate documents and pass them on to your existing systems, with human review where it matters.
Not an off-the-shelf SaaS that your documents have to fit into, but a solution that fits your document flows, your exceptions and your integrations. Whether it's Peppol invoices, scanned PDFs, photos from a field-service app or email attachments.
What AI document processing involves, beyond OCR
OCR is the oldest part of the chain: turning pixels into characters. But modern document processing is far more than that. A serious pipeline does at least five things in sequence, and each step has its own model choice, failure mode and guardrail. We lay them out below, not to impress, but to show where your existing tooling is likely to get stuck and where custom development pays off.
1. Document classification: before a single field is extracted, the system needs to know what type of document has arrived. A sales invoice, an incoming consignment note and a return form require completely different fields and validations. We train lightweight classifiers (often a fine-tuned DistilBERT or a prompt-based LLM router is enough) so the right downstream pipeline is triggered.
2. Layout analysis and OCR: for scans and photos we use, depending on the requirements, Tesseract (open source, free, fine for clean scans), AWS Textract or Google Document AI (stronger on tables and handwriting), or newer VLMs such as GPT-4o and Claude 3.5 Sonnet, which can produce structured JSON directly from an image. The choice depends on volume, languages, cost and compliance, not on what's trendy.
3. Field extraction and entity recognition: picking the right fields out of raw text. IBAN, KvK number, date, VAT number, invoice lines with VAT rates, signature fields. We do this with a combination of structured prompting (an LLM with a JSON schema) and classic regex/NER where that is cheaper and more reliable. Not everything needs to be an LLM: a good regex for postcodes is faster, more predictable and free.
4. Validation and business rules: checking extracted data against the real world. Does the VAT total match the lines? Does the supplier exist in the accounts payable ledger? Is the date within a reasonable window? Does the PO number match an open purchase order? This is where custom development almost always beats a SaaS, because your business rules aren't your neighbour's.
5. Human-in-the-loop for exceptions: when confidence scores are too low or validations fail, the document is sent to a review queue where a person corrects it. Those corrections then feed back into the model. A good pipeline takes over between 70% and 95% of the work without human intervention. Not 100%, and that isn't a failure but realism.
Document flows we automate
The flows below are the ones we encounter most often at medium-sized and large Dutch organisations. We build pipelines for them that connect directly to the ERP, CRM and DMS systems you already use.
Accounts payable invoices
PDF, UBL/Peppol, scanned paper invoice or a photo from a field-service engineer. We extract the invoice header, lines and VAT rates and match them to PO numbers. Output goes straight to Exact, AFAS, Twinfield, Unit4, Navision or a custom ERP via REST/SOAP.
Incoming contracts
NDAs, supplier agreements, leases. We identify clauses (term, notice period, indexation, liability), store key data in a contract register and set up alerts for expiry dates. Integration with DocuSign, PandaDoc or your own DMS.
Application and intake forms
Customer onboarding, damage claims, grant applications, job applications. Whether it's a digital form, a scanned paper form or a photo from a mobile app, we extract the fields, validate them against business rules and route them to the right department.
Email classification and routing
Generic inboxes (info@, support@) are classified by topic, urgency and audience. Attachments are taken into account in the assessment. Routing to Zendesk, Freshdesk, TOPdesk or your own ticketing system. Auto-replies for the simple cases.
ID verification and KYC
Passport, driving licence, identity card: extraction of the MRZ, checking of holographic security features and comparison with a selfie. Useful in fintech, mobility and the temp agency sector. With a fallback to specialist partners (Onfido, iDIN, Veriff) for the heavier compliance cases.
Report and spreadsheet analysis
Annual accounts, market reports, due diligence files, quarterly reports. We build retrieval pipelines that make documents searchable by semantic query, with proper source citations so an end user can click back to the page the answer came from.
Our typical technical stack, and when we choose what
We have no blind preference for any one vendor. We choose per project based on document type, volume, languages, latency requirements, cost and data residency. Below are the combinations we use most often.
OCR and layout
- Tesseract / OCRmyPDF: when the scans are clean and the volume is high. Open source, no API costs, and it can run on-premise when needed.
- AWS Textract: when tables and forms are central. Strong on structured layouts and question-answer pairs.
- Google Document AI: when there is a lot of handwritten text, or for specific pre-trained processors (invoice, passport, W-9).
- Azure AI Document Intelligence: for clients who are fully in the Microsoft 365 / Azure stack and want to keep their data within the EU.
LLMs for extraction and classification
- GPT-4o and GPT-4o-mini: excellent at structured JSON output, fast and relatively cheap in the case of mini. Can read images directly.
- Claude 3.5 Sonnet: strong on long documents and reasoning. Often our choice for contract analysis.
- Mistral models via a private endpoint: when data must not leave the company (defence, healthcare, financial sector).
- DistilBERT / spaCy NER: for classifiers and entity extraction where latency and cost really matter. A fine-tuned BERT model is often 10x cheaper and faster than an LLM call.
Workflow and orchestration
- Temporal / AWS Step Functions: for long-running, retry-safe document pipelines with human approval steps.
- n8n / Make: when the business wants to be able to adjust things itself and the logic isn't too complex.
- FastAPI + Celery + Redis: our go-to for custom pipelines that we manage end to end.
- Apache Kafka: for event-driven architectures where document events also need to trigger other systems.
Storage and retrieval
- S3 / Azure Blob for the raw documents, with versioning and lifecycle policies.
- PostgreSQL with pgvector when the number of documents is manageable: fewer moving parts.
- Pinecone / Weaviate / Qdrant when you're dealing with genuinely enterprise-scale volumes.
- Elasticsearch / OpenSearch for classic full-text search alongside semantic retrieval.
How we run a document processing project
We work in four phases. They look simple, but what each phase involves depends heavily on your document types and volumes. We don't bring a pre-packaged method, but we do bring a toolkit built from previous implementations.
Document audit
Two weeks. We collect 50–200 representative samples per document type, measure current turnaround times, map out exceptions and determine which fields really need to be extracted. Result: a scope document with use cases and success criteria.
Proof of concept
Three to four weeks. We build a working pipeline on your busiest document flow, with a simple review UI. The aim: to show that we can exceed an agreed accuracy threshold on real data, not on a demo set.
Production and integration
Six to twelve weeks, depending on scope. Integrations with ERP/CRM/DMS, monitoring (Datadog, Grafana or in your own stack), audit logging, user and role management, and a review UI that suits your end users.
Further development
Ongoing. Adding new document types, revising models, monitoring accuracy via dashboards, new fields, new language support. We keep models up to date — that's an ongoing responsibility, not a one-off.
Compliance, privacy and data residency
AI document processing almost always involves personal data, commercially sensitive information or compliance requirements. We factor that into our decisions from day one — not as an afterthought.
GDPR and purpose limitation
We build pipelines that capture only the data needed for the processing purpose. Personal data is pseudonymised or redacted where the use case allows. Retention periods can be configured per document type. Data processing agreements are arranged as standard.
EU-only data residency
If your documents must not leave Europe, we use models and infrastructure that stay within the EU: Azure OpenAI in West Europe, Mistral on private endpoints of your own, or self-hosted models on Dutch cloud infrastructure. No US SaaS data transfers when that's a no-go.
Sector-specific regulations
For healthcare (NEN 7510, ISO 27001), the financial sector (DORA, PCI-DSS) and government (BIO, NIS2), we have previous implementations behind us. Audit logging, role segregation, encryption at rest and in transit, and penetration testing are a standard part of the work.
EU AI Act classification
Together with you, we determine which risk category your use case falls under in the AI Act. For medium- and high-risk applications, we build transparency, monitoring and human oversight requirements into the design.
What it delivers — and what it doesn't
We can't make claims like "60% faster" or "40% cost savings" out of thin air — those figures depend entirely on your document volume, current turnaround time and the complexity of exceptions. What we can offer is an honest picture of what we've seen in comparable projects:
- With high-volume structured documents (such as incoming invoices in a known format), we often reach 85–95% straight-through processing without human intervention.
- With unstructured documents (handwritten forms, photos, contracts), that is typically between 50% and 75% — still a substantial saving on manual processing, but not the 95% that SaaS vendors claim in their marketing.
- The biggest gain rarely lies in pure FTE savings. More often it's in fewer errors (VAT mistakes, duplicate payments), faster turnaround (paying suppliers sooner improves the supplier relationship) and better data for reporting.
- Implementations almost always pay for themselves within 12–18 months at volumes of around 500 documents a week or more. Below that volume, a SaaS solution is sometimes cheaper — we're upfront about that.
How we differ from out-of-the-box vendors
There are good SaaS solutions for document AI: Klippa, Rossum, Hypatos, Docparser, Mendable. For companies with standard document flows and low requirements for customisation, they work very well — and we often recommend them instead of a custom build. We come into the picture when:
- Your documents don't fall into standard categories — industry-specific forms, architectural drawings, legal deeds, ship documents.
- Your business rules are complex and unique (such as your own pricing model for supplier invoices or specific approval routing).
- Data absolutely must not leave the EU or your own infrastructure.
- You integrate with legacy systems that have no modern API integrations.
- You want to own the model and the pipeline, rather than be locked in to a vendor.
- The volume is so high that SaaS per-document pricing eats into the business case.
Frequently asked questions
How accurate is AI document processing really?
For clean, structured documents such as digital invoices, modern pipelines often reach field accuracy above 98%. For handwritten forms or poor scans, that drops to around 80–90%. We test on your own data during the PoC phase — not on marketing examples — so the figures we put in the business case hold up.
Does this work on scanned PDFs and photos, not just digital documents?
Yes. We combine OCR (Tesseract, Textract, Document AI) with layout detection and LLM extraction. Photos from a mobile app are automatically straightened and optimised before they enter the pipeline. The difference in accuracy between a scanned paper document and a digital PDF is usually smaller than people think.
How do you handle documents in multiple languages?
Modern LLMs and OCR models support Dutch, English, German, French and Spanish natively. For languages with different scripts (Polish, Russian, Arabic, Chinese) we often add a separate language-detection step beforehand. We never limit a pipeline to a single language if your actual document flow is multilingual.
Can you integrate with our existing ERP / DMS?
In almost all cases, yes. We have previously built integrations with Exact Online, AFAS, Twinfield, Unit4, SAP, Navision/Business Central, Salesforce, HubSpot, TOPdesk, M-Files, SharePoint and on-premises SharePoint. For systems without a public API, we work with file-watching on SFTP exchange or database replication.
What if the model makes a mistake — who is liable then?
A document AI pipeline produces proposals, not decisions. The business logic determines which fields go into the accounting system without a human check and which need a review. We design this with you so that liability and risk fit your governance — not whatever is technically the maximum possible.
Can you also run it on-premises without the cloud?
Yes. For clients whose data may not leave the premises (defence, some healthcare organisations, traditional financial institutions), we run self-hosted models on your own GPU infrastructure or a Dutch private cloud. That is more expensive than a SaaS solution, so we go through the trade-offs with you.
Ready to automate your document processes?
An initial conversation is non-binding. We look at your document flows, give a realistic estimate of where AI pays off and where it doesn't, and tell you honestly when a SaaS solution is a better fit than a custom build.