Skip to main content
Possible SetupFinance & Invoicing

Automate Invoice Processing: an End-to-End Demo

Incoming invoices from inbox to GoBD archive: four entry channels, one data model, posting to DATEV/SAP/Odoo. Demo reference architecture.

Industry
Bookkeeping · 100–500 invoices a month
Build, estimated
6 weeks

At a glance

It starts with
Four inbound channels
The machine handles
9 of 11 steps
A person keeps
Check uncertain fields · Approval
Bottom line
~440 %Year-1 ROI (example)

The problem

Real talk: nobody enjoys processing invoices. It's one of those jobs that exists because invoices keep coming — not because anyone signed up for it.

In a thirty-person company with three hundred invoices a month, it's one person's job. They open every PDF, type vendor names and amounts, hunt down purchase orders, chase managers for approval, post, file. Twelve minutes per invoice, sixty hours a month — time that produces nothing.

Then comes the day this exact person gets sick, quits, or goes on vacation for two weeks. Intake stalls. Cash discounts gone. Suppliers calling.

That's what we solve here — not in twelve months of consulting, but in six weeks of pipeline.

Example scenario: a business with about 30 employees, 35 €/h personnel cost, ~600,000 €/month purchase volume. A single person handles invoice intake — 300 invoices a month, 12 to 15 minutes manual work each.

Open PDF, type vendor, number, date, net, VAT, find purchase order, route for approval, wait, post, file. That's 60 to 75 hours a month producing nothing. Plus 5 to 8 % typing errors, occasional duplicate payments (avg. 1,500 € each), regularly missed cash discount deadlines.

The unaddressed question: what happens when this one person gets sick, quits, or takes two weeks off? Intake stalls, suppliers wait, cash discounts are gone. Bus factor of 1 — that's the real risk, not just the hours.

A business owner alone in the office in the evening, rubbing her temples in front of a laptop and stacks of paper.

Staged scene

Friday, 6:50 pm

One person knows how incoming invoices work. From Monday she is on leave for two weeks.

Everything has to be posted before then, or it waits. The workflow below no longer depends on one person.

Video of the process

How the automated invoice intake works

About 10 minutes from incoming email to finished posting. Including the question rarely asked: what happens when your bookkeeper is sick, on vacation, or quits?

Read the full guide

The process, step by step

Scroll through. The diagram stays put and highlights the step you are on.

The workflow01 / 11
  1. 01

    Four inbound channels

    Email / SFTP / Peppol / Portal

    Invoices arrive four ways: a dedicated mailbox (invoices@company.com), scans via SFTP, Peppol/AS4 for XRechnung and regular fetching from supplier portals. Everything runs into the same pipeline; there is no second path on the side.

  2. 02

    Detect format

    PDF / ZUGFeRD / XRechnung

    Format detection separates classic PDFs, ZUGFeRD PDFs with embedded XML and pure XRechnung files. That decides whether the invoice is read or recognised.

  3. 03

    Extraction

    OCR + LLM / XML parse

    For classic PDFs, OCR reads the text and a language model maps the fields, even when every supplier writes "invoice number" differently. ZUGFeRD and XRechnung are read directly as XML to EN 16931.

  4. 04

    One data model

    invoice.json with confidence

    All three paths end in the same data model. Every field carries a confidence score: how sure is the machine about the amount, the date, the IBAN?

  5. 05Branch

    All fields certain?

  6. 06Human decides

    Check uncertain fields

    only these, not the whole invoice

    Only the fields the machine is unsure about go to a person. They correct those instead of typing the whole invoice.

  7. 07

    Five checks

    Schema / VAT / Duplicate / Master data / Sanctions

    Five checks run in parallel: EN 16931 schema, VAT plausibility, duplicate hash, a match against ERP master data and the EU sanctions list. Duplicate invoices stop here.

  8. 08Branch

    Over €500?

  9. 09Human decides

    Approval

    1 click in Slack/Teams, four eyes from €5,000

    Up to €5,000 the department approves with one click in Slack or Teams; above that, four eyes with management. If someone is away, the approval goes to their deputy; after 24 hours it escalates.

  10. 10

    Posting

    DATEV / Lexware / SAP / Odoo

    Once approved, the invoice is posted: DATEV, Lexware Office, sevDesk, SAP FI or Odoo. The AI suggests the cost centre from earlier postings.

  11. 11

    GoBD archive

    Hash chain, 10 years

    The original is archived unchangeably, with a hash chain and ten years of retention under GoBD. The export for an audit is one click.

What changes

Possible setup, not a packaged product

The figures are modelled target values for a possible setup, derived from industry benchmarks and our own tests. They do not come from a customer project; real values depend on your business. We do not sell this setup as a finished product: we measure where your bottleneck is, then build with your team. For systems that require certification (e.g. HIS, gematik, DATEV-certified) we work with specialised partners.

Before vs. After

AspectBeforeAfter
Time per invoiceBefore: 12–15 min manualAfter: 1–2 min spot check
Hours / mo (300 invoices)Before: 60–75 hAfter: 8–10 h exceptions + sampling
Total cost / moBefore: ~2,350 €After: ~665 € (incl. software & maintenance)
Cash discount usageBefore: 30–50 %After: > 90 %
Error rateBefore: 5–8 %After: < 0.5 %
Throughput intake → bookingBefore: 8–12 daysAfter: < 2 days (median ~4 h)
Duplicate payments / yrBefore: 2–5 cases @ ~1,500 €After: 0 (hash duplicate lock)
Audit readiness (GoBD)Before: ad-hoc, 3–5 days prepAfter: permanent, one-click export
Bus factorBefore: 1 person holds the knowledgeAfter: Pipeline + caretaker (replaceable)

Technology

Technology Stack

n8nMake.comOpenAI GPT-4 VisionAnthropic ClaudeAzure Document IntelligenceGoogle Cloud VisionLexware Office APIDATEV ConnectSAP FIOdooPeppolPostgreSQLSlack

Integrations

Systems this setup can connect to

DATEV Rechnungswesen / Unternehmen OnlineBookkeeping
DATEVconnect online for direct receipt upload, EXTF-CSV as fallback for tax advisor handover.
Lexware Office / sevDeskBookkeeping (REST API)
Native integration via Lexware Office and sevDesk APIs. Voucher with original PDF, cost center and tax code in one step.
SAP S/4HANA / Business OneERP
OData and BAPI interfaces for PO matching (three-way-match) and posting to SAP FI.
Odoo / Microsoft Business CentralERP
REST/XML-RPC for Odoo, Microsoft Graph for Business Central. Full receipt flow, no manual re-entry.
Peppol Access PointE-invoicing
AS4 receipt for XRechnung — relevant for public sector and the German B2B mandate from 2027. Implemented via Phase4 or providers like ecosio.
Slack / Microsoft TeamsApproval
Block Kit (Slack) or Adaptive Cards (Teams) for approval buttons. Delegate routing and escalation built in.
Azure Document Intelligence / Google Cloud VisionOCR
Text recognition for classic PDF and scanned invoices.
OpenAI GPT-4 Vision / Anthropic ClaudeAI structuring
Function calling / tool use for field-level extraction with confidence scores. EU endpoints, zero data retention.

Security & Compliance

What the setup takes care of

GoBD-compliant archive
Original PDF hashed at intake (SHA-256) and archived tamper-evident for 10 years. Compliant with the German Ministry of Finance directive of 28.11.2019.
GDPR & EU hosting
Processing in EU data centers, DPA per GDPR Art. 28. LLM API calls (OpenAI/Anthropic) via EU endpoints with zero data retention.
Append-only audit log
Every step — intake, extraction, correction, approval, booking — logged in an immutable hash chain. One-click export for audits.
IBAN and sanctions check
Before each booking IBAN is validated against schema and check digit, vendor checked against the EU consolidated sanctions list.

Frequently Asked Questions

Sweet spot: 10 to 100 employees, ~100 to 500 incoming invoices per month. The example scenario shown (300 invoices/mo, 30 staff, 600k € purchase volume) is the typical mid-point. Smaller setups — solo with under 30 invoices — are better served by tools like GetMyInvoices at 29 €/month. Large enterprises with their own procure-to-pay setup usually have the lever already.
I do not quote a flat price here. What it costs depends on the accounting stack: a plain Lexware Office or DATEV integration with text recognition is much smaller than an SAP FI integration with three-way match against PO and goods receipt, or a Peppol access point. Before you decide, you get a concrete number, no day rate and no open-ended scope. There is no maintenance subscription; the solution belongs to you.
Conservatively. 300 invoices × 12 min = 60 h personnel time, ~50 h saved (×35 €/h ≈ 1,750 €). Cash discount: 3 % on additional usable purchase volume gives ~6,000 €/mo at 600k € volume. Avoided duplicate payments: ~440 €/mo on average. Software/cloud (n8n, LLM API, DMS): ~350 €/mo. Maintenance: ~140 €/mo. Net: ~8,050 €/mo. Even removing the cash discount effect entirely (only personnel + duplicate avoidance minus software/maintenance), you still net 2,000 €+/mo with payback under a year.
95 to 99 % on standard invoices with readable PDFs. Edge cases (poor scans, handwritten corrections, unusual layouts) go into a short spot-check — you see the AI suggestion with confidence score and confirm or correct in one click. The system learns from each correction for similar cases from the same vendor.
That's exactly the point. The pipeline runs 24/7 without anyone present. Incoming invoices are extracted, validated, and — where possible — auto-approved and booked. Approvals route via Slack or Teams to the delegate (Postgres table, optional calendar sync). Cash discount deadlines are monitored automatically. When the bookkeeper returns, everything is booked or in a clearly marked clarification stack.
Yes, as long as GoBD requirements are met. Original PDF is hashed (SHA-256) at intake and archived tamper-evident for 10 years (S3 with Object Lock or a DMS like d.velop, DocuWare, or Nextcloud with DSP). Each change goes through a hash-chained audit log. A GoBD process documentation is part of the deliverable — once written, reviewed by your tax advisor, usable for every audit.