Home » Case Studies » Document Automation

Case study: Document automation for an SME

An anonymised example from our project portfolio, with a clear before-and-after comparison using real pilot figures. Industry details have deliberately been neutralised; where there is a specific need, we can provide genuine references during initial consultations, subject to customer approval.

Note: Anonymised at the customer’s request. Real figures, real setup.

Starting point

A medium-sized service provider with approximately 120 employees and around 1,500 incoming documents per month, including invoices, delivery notes and order confirmations. Two full-time accounting staff entered the data manually. The problems were:

Entering each document took an average of 4–6 minutes
The account-coding error rate was 8–12%
Early-payment discount deadlines were regularly missed
Covering holidays and sick leave in the accounting team was a bottleneck

Approach

Sample analysis (week 1): We tested 200 typical documents against our system and reported the recognition accuracy transparently.
Pilot with one supplier category (weeks 2–5): Incoming invoices from the three largest suppliers ran through the AI pipeline in production using confidence thresholds.
Rollout (weeks 6–12): We gradually expanded the system to all document types and suppliers and integrated it with DATEV.
Hypercare (months 4–6): We closely supported optimisation and built a KPI dashboard.

Solution

OCR and LLM-based extraction of the relevant fields for each document type
Automatic account-coding suggestions based on historical postings
A 95% confidence threshold, with anything below it sent to the validation interface for manual review
DATEV integration with the original PDF attached
Hosting in a German data centre with a data processing agreement in place

Results after 6 months

~70%less data-entry time per document

~50%fewer account-coding errors

0early-payment discount deadlines missed since rollout

~9 monthspayback period

Lessons learned

Analyse samples first. Without an honest accuracy measurement beforehand, excessive expectations will derail the rollout.
Start small. Beginning with three suppliers instead of all 200 proved invaluable.
Normalise human review. The team found it easier to accept 80% automation plus manual review than a promise of 100% automation.
Data control creates trust. Hosting in Germany was a prerequisite for approval by the data protection officer.

Do you have a similar setup?

We will analyse an initial sample of 50–100 documents free of charge and report the expected accuracy transparently.

Request a sample analysis Document processing service