DATEV-EXTF Automation
A German tax practice with around fifteen staff was spending several person-days every month preparing receipts for DATEV. We built an OCR and classification pipeline that processes incoming and outgoing receipts automatically, pre-assigns VAT treatment, and produces a clean DATEV-EXTF export. The bookkeeping team now only reviews exceptions.
Background
The practice serves mid-sized clients, including several e-commerce businesses with high receipt volumes — hundreds to thousands of receipts per client per month. The workflow was classic: receipts arrived by email or via the client portal, were visually checked, manually keyed in or semi-automatically captured, then manually mapped to account and contra-account, manually classified for VAT, and finally exported to DATEV. For each high-volume client, that meant one to two working days per month of pure data-entry. Across roughly twenty comparable clients, the total volume was significant. The partners did not want to add more bookkeeping headcount; they wanted the experienced bookkeepers to spend their time on review and exceptions, not on typing. They had already evaluated two off-the-shelf DATEV add-ons but neither matched the SKR03 logic they had built up over the years for their particular client mix.
Solution
We built a pipeline that ingests receipts via email or upload, runs OCR for structured fields (date, document number, vendor, gross, net, VAT rate, country of origin), classifies the transaction, and converts it into the DATEV-EXTF format. The classification logic was calibrated together with the lead bookkeeper over several iterations: which account / contra-account combinations belong to which receipt types, when the BU key 240 applies for intra-Community supplies, how third-country transactions are handled. OCR runs on Claude Vision, classification is orchestrated in Langflow, and data is persisted in PostgreSQL on our Hetzner infrastructure in Frankfurt — meaning all processing happens within the EU under a standard Article 28 data-processing agreement. An optional on-premise variant is available; some of the practice's clients have particularly strict data requirements, and the option to run the pipeline inside the client's own network was a precondition for a few of the engagements. The DATEV-EXTF output is validated against the official schema before being released to the bookkeeper for import.
Outcome
The majority of receipts now flow through the pipeline automatically. The bookkeeper only inspects flagged exceptions — new receipt types, unusual amounts, missing mandatory fields. Monthly preparation time per client has dropped substantially. Internally, the practice talks about roughly seventy percent time saved on the capture step; that figure refers to the direct data-entry activity, not the full monthly close. Two new e-commerce clients have been onboarded since the pipeline went live without adding any bookkeeping headcount. GDPR-compliant via German hosting, on-premise variant available for clients with elevated requirements.
Lessons Learned
Three honest takeaways from this project. First: DATEV-EXTF is more mechanical than it looks, once the specific practice's account logic is properly documented. We spent almost the entire first week sitting next to the bookkeeper, verbally extracting the in-house SKR03 conventions — those hours paid back a hundredfold later. Second: OCR is never one hundred percent. Our first attempt routed low-confidence captures straight back into a generic inbox, which frustrated the bookkeeper because she could not tell what the system was uncertain about. We rebuilt the review UI so that each flagged item shows exactly which field the model was unsure of, with the OCR crop next to the suggested value — that single change turned the review step from annoying to fast. Third: the practice's own clients react very differently to the phrase "AI in bookkeeping". The on-premise option turned out not to be just a technical feature but the thing that kept two of the practice's more conservative clients on board.
A similar problem?
Drop us a short note and we will reply within one business day.
Request a project