Service · Web development

Automate invoice recognition with AI and OCR.

Automatically read, classify and post supplier invoices to your accounting or ERP system. We build custom invoice processing where standard tools fall short: industry-specific, with deep integrations, and ready for Peppol and the EU e-invoicing mandate in 2030.

OCR & LLMPeppol / UBLERP integration3-way matchGDPR & audit

For when standard scanners don't go far enough.

Klippa, Basecone, AutoInvoice, FactuurSturen and Visma Scanner do a good job on standard PDF invoices, so we don't build a copy of those. Our clients come to us when those tools hit a wall: an industry-specific invoice format the standard parser can't read, an ERP integration that needs to go a little deeper, or an approval flow with multiple cost centres and budget categories that no off-the-shelf SaaS form accommodates. Often it's a combination: an organisation that wants not only to recognise invoices, but to match them straight away against purchase orders, contracts and framework agreements held in its own systems.

We build bespoke invoice recognition using modern OCR (Tesseract, PaddleOCR, Google Vision) and multimodal models such as Anthropic Vision and GPT-4o Vision that genuinely understand the nuance of an invoice. The result is a processing pipeline that fits into your ERP, not the other way around, and that grows with new supplier formats, new budget structures or a future move to Peppol-only e-invoicing. Get in touch if you run an accountancy practice, manage procurement flows in a large organisation, or need this as part of wider document workflow automation.

Our clients are SMEs processing fifty to five hundred purchase invoices a month, accountancy and bookkeeping firms handling client ledgers, and large organisations with formal procurement processes. What they share is a breaking point: the volume or complexity is too high for manual entry, but too specific for an out-of-the-box scanner.

Three scenarios where custom development makes a difference.

Not every organisation needs the same thing. In the first conversation we advise which approach suits your invoice volume and systems landscape.

Compact project · fixed sprint budget

Invoice processing for SMEs with deep accounting integration

For SMEs manually booking fifty to five hundred purchase invoices each month. We build an ingestion flow in which invoices arrive by email, drop folder or Peppol, are read automatically by a combination of OCR and a language model, and then post directly as journal entries in Exact, Twinfield or AFAS. The approver receives a notification, approves with one click, and the invoice is ready for payment.

OCR + LLM extractionExact / Twinfield / AFASApproval workflowMailbox ingestion
Mid-sized project · fixed sprint budget

Accounting firm: batch processing of client invoices

Accounting and bookkeeping firms often work for dozens of clients, each with their own bookkeeping package and their own records. We build a central processing pipeline that knows the correct chart of accounts for each client, automatically suggests the right general ledger account and cost centre, and only brings in a person when there is doubt. The LLM learns from previous postings — the same supplier on the same account, automatically.

Multi-tenantLedger suggestionHuman-in-the-loopAudit trail
Larger project · fixed sprint budget

Procurement flow with 3-way match and Peppol

For large organisations with a formal procurement process: incoming invoices are automatically matched against the purchase order and goods receipt (3-way match), with escalation to the right budget holder when discrepancies arise. Full Peppol BIS support for e-invoices, native UBL processing, and integration with SAP, Oracle or Microsoft Dynamics. An audit trail at every step, ready for your auditor's review.

3-way matchPeppol BISNative UBLSAP / Dynamics

What you get at the end.

A working processing pipeline in production, plus the tools to adjust and expand it yourself.

  • Live processing pipelineProduction and staging environments, hosted in your own cloud (GCP / AWS / Azure) or with us. Inbound channels for email, upload, Peppol and API ingestion.
  • Management dashboardA single view for your finance team: invoices in progress, escalations, processing times per supplier, and patterns that need attention.
  • Full codebase + documentationSource code in your repository, with architecture description, deployment instructions and an incident runbook. No vendor lock-in.
  • GDPR and audit reportingDPIA template, data catalogue, retention periods set per document type, and an audit log showing who changed or approved each invoice and when.
  • Training for finance and ITTwo sessions for key users (how to handle escalations, how to adjust posting rules) and a technical session for your IT team on administration.
  • Maintenance contract (optional)Ongoing maintenance, model updates, new supplier formats and integration extensions. Fixed monthly price, four response-time levels.

The technical stack behind invoice recognition.

We are not committed to a single vendor or model. For each project we choose the combination that suits your invoice mix, your cloud environment and your budget, and we explain why.

Layer 1 · Document extraction

OCR and multimodal vision models

For clean PDFs and e-invoices we work directly with the structured data. For scans and photos we combine classic OCR (Tesseract and PaddleOCR for local processing, Google Vision where cloud use is acceptable) with multimodal models such as Anthropic Vision and GPT-4o Vision. These models read an invoice as a whole: they understand that a number in the top right is the invoice number, even if the label reads "Invoice #" or "Faktura Nr.". For sensitive documents we run local models in your own environment, so invoice data never leaves your data centre.

TesseractPaddleOCRGoogle VisionAnthropic VisionGPT-4o Vision
Layer 2 · Data validation

LLM-based extraction and validation

The raw text is passed to a language model that extracts the fields you need: invoice number, invoice date, due date, VAT amounts per rate, cost centre, project code, supplier details, IBAN. We build business-rule validation around it: does the VAT total add up, does the invoice date fall within the financial year, do we know this supplier? If in doubt, the invoice goes to a human reviewer with the points of concern already flagged. This human-in-the-loop approach moves quickly to full automation once the model has learned the patterns.

Field extractionBusiness rulesVAT validationIBAN check
Layer 3 · Integration

Accounting, ERP and Peppol flow

The extracted data is then pushed into your system. For SMEs this is usually Exact Online, Twinfield, AFAS or Visma. For accountants, often Yuki or Snelstart per client. For enterprises, SAP, Oracle NetSuite, Microsoft Dynamics or Unit4. We process incoming Peppol invoices natively as UBL, and for outgoing invoices we connect to a Peppol Access Point. For specific integrations not covered here, we look at the existing API and build an adapter. Read more on our page about smart API integrations.

Exact / Twinfield / AFASVisma / YukiSAP / DynamicsPeppol Access Point

When bespoke development is the right choice.

Four signs we see in organisations that have outgrown their standard scanning tool. If you recognise one of them, we would be happy to talk further.

Industry formats

Your suppliers use a different invoice format

Construction, healthcare or agricultural suppliers issue invoices with line items and codes that standard scanning tools cannot separate properly. A custom extraction model that understands your invoice patterns delivers clean data straight away.

Deep ERP integration

The accounting integration needs to do a little more

Standard tools write to a single field in Exact or Twinfield. You have cost centre, project and VAT code logic that differs from invoice to invoice, and you want that filled in automatically before posting.

Volume & scale

The volume doesn't justify manual work

At fifty invoices a month, manual entry still works. At five hundred or more, it becomes a full-time job. A custom processing pipeline pays for itself quickly, especially if it can intervene more deeply in your processes.

Peppol & 2030

You are preparing for mandatory e-invoicing

The EU will require business-to-business e-invoicing via Peppol-style standards from 2030. If you build a platform now that handles both PDF and UBL, you won't have to scramble to switch later.

How an invoice recognition project works.

1

Introduction and data sample

We review a representative set of your invoices (typically one hundred to two hundred documents across your main suppliers) and your current booking process. From this we produce an initial assessment of extraction quality per field, the integrations we need, and where the biggest time savings lie. At this stage we're candid about what we can and can't solve for you. Sometimes a standard tool supplemented with a small custom connector is more than enough, and our clients prefer to hear that up front rather than halfway through a project.

2

Architecture and proof of concept

A short design sprint with your finance and IT teams. We choose the right OCR stack (Tesseract, PaddleOCR or Google Vision), decide which LLM handles the extraction work, and set up a proof of concept with which your team tests the first batch.

3

Building in sprints

A working build every two weeks: more supplier formats, integration with your accounting system, the approval flow, and an audit log. You test continuously, and finance joins in on the UX decisions.

4

Pilot and phased rollout

We start with one department or one set of clients, run in parallel with the old process for a period, and expand once extraction quality and approval speed are where they need to be.

5

Further development and maintenance

New suppliers, new integrations, model updates and expansion into areas such as computer vision applications for receipts or contract analysis. Many clients extend into full procurement automation in year two.

Frequently asked questions.

What clients usually want to know before we start.

Will you replace tools like Klippa, Basecone or AutoInvoice?
No. For standard PDF invoice processing, those tools do the job well and are far cheaper than custom software. We come in when you hit their limits: an industry-specific invoice format they can't read, an ERP integration that needs to go a bit deeper, or an approval flow that doesn't fit any SaaS form. Often our solutions run alongside an existing tool rather than replacing it.
What is the difference between classic OCR and AI-based extraction?
Classic OCR (Tesseract, PaddleOCR) converts pixels into text. That works well for clean PDFs but struggles with scans containing smudges, unusual layouts or handwritten notes. Multimodal models such as Anthropic Vision and GPT-4o Vision read an invoice as a whole in one go: they understand that a number in the top right is the invoice number and the amount at the bottom is the total, even with unusual formats. We combine both: OCR for the raw work, and a language model for interpretation and validation.
How do you support Peppol and UBL?
Peppol is a European network for e-invoicing and UBL is the underlying data format. We build native UBL processing: invoices arriving via Peppol are read directly as structured data, without the OCR detour. For sending, we connect to a Peppol Access Point. This matters in light of the EU e-invoicing obligation coming into force in 2030. A platform that already handles both PDF and UBL won't need a rushed migration later.
Which accounting and ERP systems do you integrate with?
We have experience with Exact Online, Twinfield, AFAS, Visma, Yuki and Snelstart, and on the ERP side SAP, Microsoft Dynamics, Oracle NetSuite and Unit4. For each integration we look at both the official API offering and what the system can actually handle in practice: some accounting packages have tight APIs, others only accept CSV imports. For topics beyond direct accounting integration, see also our smart API integrations.
What about GDPR, data location and audit trails?
Invoices contain personal data (names, bank details, sometimes a BSN for sole traders). We carry out a DPIA for every project, encrypt data in transit and at rest, and host in European data centres with providers under a data processing agreement. An audit log is standard: for each invoice you can see who changed, viewed or approved what. Retention periods are configurable per document type, in line with the seven-year accounting retention obligation and the GDPR data minimisation requirement.
What determines the cost of such a project?
Three factors: how many different invoice formats you process (a set of ten suppliers is much quicker to model than a hundred), how many integrations we need (one accounting package or four ERPs plus an approval platform), and the complexity of your approval workflow. After the introductory conversation we provide a reasoned estimate per sprint: not a fixed total price based on gut feeling, but a transparent budget per phase.
Can we start with a proof of concept?
Yes, and we often recommend it. In a short initial phase we train the extraction model on a sample of your own invoices, connect one accounting package, and deliver a report showing exactly what the extraction quality is per field and per supplier. Only then do you decide whether to proceed with the full build, so there are no surprises down the line. Many clients also use this proof of concept to secure internal budget for the full rollout. Read also about our approach to the custom LLM integrations that sit behind it.

Talk to us about your invoice processing.

A free 30-minute introductory call. Send over a few sample invoices and, during that conversation, we can show you concretely where standard tools reach their limits, and where custom development makes the difference.

Edit content