From Invoice to Ledger: A Semi-Automated Finance Workflow

Receive, check, code, approve, post. Few processes are as rule-based as accounts payable, and few are already so often partly automated. So the question is what generative AI adds.

What is changing right now

Since 1 January 2025, companies in Germany must be able to receive e-invoices in B2B transactions. From 2027, companies with more than EUR 800,000 in prior-year turnover must issue e-invoices themselves, and from 2028 everyone must. An e-invoice in XRechnung or ZUGFeRD format already carries its data in machine-readable form. Those documents need no AI extraction.

Generative AI remains useful where documents are unstructured: foreign invoices, small receipts, scans, credit notes with free text. It also helps with coding suggestions when booking rules are ambiguous.

What the numbers say

We have not found an independent German case study with before-and-after figures for AI-assisted invoice processing. McKinsey describes the bigger picture. In its August 2026 survey, 37 % of respondentsU see at least some impact of AI on EBIT, about as many as a year earlier. Only around 6 %U attribute 5 % or more of their EBIT to AI. For finance this means time savings alone will not convince the controller. Translate them into money.

The workflow

The most important branch comes first. Structured e-invoices are taken over directly. Only the rest goes through AI extraction.

Extract from the attached invoice: supplier, invoice number, invoice date, net amount, VAT rate, VAT amount, gross amount, purchase order number. Return the result as JSON. Write "UNCLEAR" if a value cannot be read with certainty. Do not guess values.

Limits and risks

• German bookkeeping rules (GoBD): Every posting must be traceable and documented in a tamper-proof way. Store the AI suggestion and the human approval.
• Four-eyes principle: AI suggests, people approve. Direct postings by a model are off limits.
• Extraction errors: Transposed digits are the biggest risk. Cross-check net, tax and gross automatically.
• Supplier fraud: Always confirm changed bank details by phone using a known number.

How to start

1. Count how many incoming invoices already arrive as XRechnung or ZUGFeRD.
2. Automate those first, without any AI.
3. Test AI extraction on 50 unstructured documents and measure the error rate per field.
4. Convert the time saved into euros and set it against licence and review costs.

Sources

1. UMcKinsey: The state of AI in 2026, On the road to ROI, 25 Aug 2026
2. RIHK Region Stuttgart: E-invoices, end of transition period for companies above EUR 800,000 turnover
3. RIHK Rhein-Neckar: E-invoicing mandate for B2B transactions