Every week, the administration department of an average industrial SME processes between 80 and 300 supplier invoices. Seventy percent arrive as PDFs — scanned, generated from the supplier's ERP, or in Factura-e format — and the rest come on paper, as images, or by email. Entering that data manually into the ERP costs time, generates errors, and keeps skilled people busy with work below their potential. The good news: the technology to automate supplier invoice data extraction and inject it directly into the ERP is mature, available to SMEs, and recovers its investment in under twelve months in most cases.
This article explains which technologies are involved, how the complete workflow runs from invoice receipt to the accounting entry, what factors determine accuracy, and when it makes sense to commission a custom implementation rather than use a standard SaaS solution.
Why Manual Invoice Processing Is Still a Bottleneck
Ardent Partners (2025) estimates that the average cost of manually processing a supplier invoice in Europe ranges from 8 to 15 euros, including operator time, error correction, and discrepancy management against purchase orders. An SME processing 150 invoices a month may be spending between 1,200 and 2,250 euros monthly on that process alone, not counting late-payment penalties caused by approval delays.
Beyond direct costs, manual processing introduces errors at an average rate of 1–3% per field (according to IOFM, the Institute of Finance & Management), leading to duplicate payments, order discrepancies, and longer internal audits. With the rollout of mandatory B2B e-invoicing in Spain — scheduled for 2025 for large companies and 2026 for the rest under the Ley Crea y Crece and its implementing Royal Decree — the volume of structured invoices will grow, along with stricter requirements for traceability and matching against the original purchase order.
The Three Technologies That Make Automatic Extraction Possible
Intelligent OCR and Document Understanding Models
Traditional OCR (Optical Character Recognition) converts text images into digital characters but has no contextual understanding: it knows there is a number, but not whether it is a tax ID, a total, or an invoice number. Document understanding models — such as Azure Document Intelligence, AWS Textract, or Google Document AI — combine OCR with language models trained specifically on invoices. They identify semantic fields (issuer, recipient, issue date, due date, taxable base, VAT rate, VAT amount, total, invoice number, associated order number) with accuracy exceeding 95% on structured documents without any prior fine-tuning, reaching 99%+ with a model fine-tuned on invoices from a specific supplier.
For Factura-e (structured XML) or Peppol BIS 3.0 (the European B2B e-invoicing standard), no OCR is needed: the data already arrives as labelled fields and extraction is a direct XML read. As e-invoicing becomes the norm through 2026, the share of invoices in structured format will grow and extraction costs will fall.
Extraction via n8n or Custom Automation Workflows
Once the document understanding technology returns the extracted fields, an orchestrator is needed to collect the result, validate it against business rules (does the supplier's tax ID exist in the ERP master data? does the amount fall within the purchasing manager's approval threshold?) and send it to the ERP through its API. Tools such as n8n — which Summum IA regularly deploys for SMEs — allow this workflow to be built without code, using visual nodes for each step: email receipt, attachment extraction, call to the Document Intelligence API, validation, write to the ERP, and notification to the approver when required. To learn more about how this orchestrator works, you can read about n8n automation for SMEs.
Direct ERP Integration via API or Native Connector
The final step is writing the entry into the ERP. Most modern ERPs — Odoo, Sage 200, Microsoft Dynamics 365 Business Central, Holded, SAP Business One — expose a REST or SOAP API for creating supplier invoices. ERPs with less openness have third-party connectors or the option to import a CSV/XML in the system's bulk import format. The automated document processing service offered by Summum IA covers both extraction and integration with the target ERP, including configuration of the chart of accounts and cost centres.
The Complete Workflow: From Inbox to Accounting Entry
To achieve a process with no human intervention for the standard case, the workflow must cover these stages in sequence:
- Multichannel receipt: email (PDF or XML attachment), shared network folder, supplier portal, or EDI integration.
- Classification: distinguishing invoices from other documents (delivery notes, quotes, credit notes). A lightweight classifier — trained on 200–300 examples — achieves over 97% accuracy.
- Field extraction: intelligent OCR or XML reading. Critical fields are: issuer tax ID, recipient tax ID, invoice number, issue date, due date, line items (description, quantity, unit price, VAT rate), taxable base by type, VAT amount, and total.
- Validation against master data: confirming the supplier's tax ID exists in the ERP master, that the associated order number matches an open purchase order, and that the amount does not exceed the order by more than a configured margin (for example, 5%).
- Approval routing: if the invoice exceeds the automatic approval threshold (for example, 3,000 euros), notify the approver with a one-click approval link.
- Write to ERP: create the supplier invoice with status «pending payment», with the accounting allocation, cost centre, and due date already calculated.
- Digital archiving: store the original PDF linked to the ERP record for traceability and compliance with the Ley Crea y Crece.
Comparison of Approaches for Spanish SMEs
| Approach | Typical accuracy | Implementation timeline | Indicative market cost 2025–2026 | Best suited for |
|---|---|---|---|---|
| Capture SaaS (Rossum, Hyperscience, Klippa, Nanonets) | 93–97% without fine-tuning | 2–4 weeks | €150–500/month (volume-dependent) | SMEs with <500 invoices/month, varied suppliers |
| Azure Document Intelligence + custom n8n | 97–99%+ with training | 6–12 weeks | Project €4,000–15,000 + €100–300/month operational | SMEs with a specific ERP or custom approval workflows |
| Native ERP connector (Odoo OCR, Sage Capture, BC Document Capture) | 90–95% | 1–3 weeks | Included in licence or add-on module €50–150/month | Companies already on that ERP with no need for multichannel flow |
| RPA (UiPath, Power Automate Desktop) | 85–92% (sensitive to layout changes) | 4–8 weeks | Project €5,000–20,000 depending on complexity | ERPs with no accessible API, Windows screens as the only route |
| Factura-e / native Peppol (structured XML) | 99.9% | 1–2 weeks (if supplier already sends XML) | Very low (direct XML read) | Large suppliers already issuing in electronic format |
Indicative sources: market prices published by the vendors themselves (Rossum pricing 2025, Microsoft Azure pricing calculator, Nanonets website) and benchmarks from IOFM e-Invoicing Report 2025. Project costs vary according to ERP complexity and the number of distinct suppliers.
Factors That Determine Real-World Accuracy in Your Case
The 95–99% accuracy vendors advertise is measured on test datasets with well-structured invoices. In practice, several factors reduce it:
- Scan quality: paper invoices scanned at below 200 DPI, with crumpled paper or faded ink, produce degraded OCR. The recommendation is to require suppliers to send digitally generated PDFs from their software rather than scanned copies.
- Layout diversity: the more distinct suppliers with their own formats, the greater the variety of layouts. A model without fine-tuning takes longer to stabilise. With 50–100 sample invoices per regular supplier, fine-tuning raises accuracy to 99%+.
- Invoices with complex line items: professional-services invoices with income tax withholdings, surcharges, or disbursements require additional logic to correctly map fields to the chart of accounts.
- Language and number formatting: invoices from international suppliers may use a full stop as the thousands separator (1,234.56 vs 1.234,56). The system must be configured for each convention.
Integration with Verifactu and Mandatory B2B E-invoicing
With the rollout of Verifactu (the invoicing record verification system required by the Spanish Tax Agency from 2025–2026 under Royal Decree 1007/2023) and mandatory B2B e-invoicing under Law 18/2022 (Ley Crea y Crece), Spain's invoice ecosystem is moving towards structured formats. This directly affects automation: invoices received in validated XML format (Factura-e UBL or the format defined by the forthcoming B2B e-invoicing regulation) eliminate the OCR step and reduce the operational cost to simply reading and integrating the file. SMEs that implement a well-designed capture system today can prepare it to receive both PDFs (while they coexist) and structured XMLs as adoption becomes widespread.
When to Build Custom vs. Use a SaaS
The choice between a capture SaaS and a custom development depends on three variables: volume, integration, and control. A SaaS such as Rossum, Klippa, or Nanonets handles the standard case well: it receives the PDF, extracts the fields, and returns a JSON or CSV for import into the ERP. If the ERP accepts bulk import and the approval workflow is straightforward, configuration takes days. The problem arises when the ERP has its own API with OAuth2 authentication, the approval workflow is hierarchical with delegations, or automatic accounting allocation rules are needed based on supplier or cost centre. In those cases, a custom project with n8n or a proprietary middleware offers more control, a lower per-transaction cost at scale, and — crucially — no dependency on a SaaS vendor that could change its pricing or discontinue the product.
Frequently Asked Questions
Does it work with any ERP or only the most well-known ones?
Data extraction is ERP-agnostic: the capture layer works with any system. Integration is where differences appear. ERPs with documented REST APIs (Odoo, Holded, Dynamics 365 Business Central, SAP Business One) allow direct writes. More closed or legacy ERPs (some industrial verticals, older Sage versions) require file-based import or interface automation via RPA. In every case, integration is technically feasible; the variable is the project cost and timeline.
What happens with invoices the system does not extract correctly?
Every well-designed system includes a human review queue for invoices with a low confidence score (for example, below 85% on a critical field). The operator reviews and corrects only those invoices, which in a mature system typically represent less than 3–5% of the total. Each correction feeds back into the model, progressively reducing that percentage over time.
How long does it take to recover the investment?
It depends on volume and the current cost of the process. With 200 invoices per month and a manual cost of 10 euros per invoice (including review time and error correction), the monthly saving is around 1,600–1,800 euros, assuming the automation handles 80–90% of invoices without intervention. A custom implementation project costing 8,000–12,000 euros recovers its investment in 5–7 months in that scenario. A SaaS with a 300 euro/month subscription and 1–2 days of configuration recovers in the first month.
How does the GDPR affect processing invoices with AI?
Supplier invoices contain personal data (administrator name, sole trader tax ID, address). Processing them to automate the accounting workflow has a legal basis under Article 6(1)(b) of the GDPR (performance of a contract or pre-contractual measures) and 6(1)(c) (legal accounting obligation). If a third-party SaaS is used for extraction, that vendor acts as a data processor and a Data Processing Agreement (DPA) must be signed. If processing is on-premises or within the company's own cloud, there is no transfer to a third party. In any case, the data must not be used to train the vendor's models without explicit consent — a point to check carefully in the SaaS terms of use.
How to Get Started: The First Three Steps
If you are evaluating the implementation of automatic invoice extraction in your company, these are the initial steps we recommend:
- Measure the baseline: how many invoices you process per month, how many distinct suppliers generate 80% of the volume, what percentage arrives as digital PDFs versus scanned or paper, and how long the current end-to-end process takes (receipt to ERP entry).
- Choose the primary input channel: if 60% of invoices arrive by email to a specific address, that is the natural entry point for the first workflow. If they come through a supplier portal, EDI, or a network folder, each channel has its own specific connector.
- Assess the ERP API: before choosing a capture solution, confirm whether the ERP accepts writes via API and which fields it exposes for creating a supplier invoice. That information defines the scope of the integration project.
At Summum Marketing we have been implementing digital solutions for SMEs since 2007. Our Summum IA team has designed and deployed automated document processing workflows in industrial, distribution, and professional-services companies, integrating with Odoo, Sage 200, and Dynamics 365. If you would like to assess whether your case fits a standard solution or requires a custom development, we can carry out that initial evaluation with no commitment.