Document automation

It starts with capture: a digital form on a phone, an uploaded PDF or a photo of a delivery note. Structured input goes straight into the database. For scans and PDFs, a language model reads the fields and returns them in a fixed schema — invoice number, date, amount, line items.

Validation comes next. Rules catch whatever can be checked unambiguously: required fields, totals, duplicates, unknown suppliers. Anything that does not clearly pass goes to a review list for a person. Only then does the workflow route the item onward — to approval, into the ERP or into the archive. We usually build this part with n8n, because the flows can later be adjusted without a developer.

Good candidates are tasks that occur often and always look similar: incoming invoices, timesheets, inspection reports, applications, delivery notes. The higher the volume and the clearer the rules, the faster automation pays for itself.

Poor candidates are documents where every copy is different and the decision needs expert judgement — individual contracts, for example. A model can pre-sort and summarise there, but it cannot decide. We say so openly in the first call, before budget goes into a process that cannot be standardised.

The biggest lever is the number of document types. One type with one approval step is a manageable project; ten types with different routes are ten small projects. That is why we start with the type that costs the most time.

• Quality of the source: clean PDFs are easy, handwritten forms need more review steps
• Target systems: an ERP, DMS or accounting tool with an API connects quickly, systems without one do not
• Approval logic: deputies, amount thresholds and four-eyes checks
• Record keeping: who changed what and when, and how long it is retained

Documents almost always contain personal data. So we decide early which fields a language model sees at all, where the data is stored and who has access. A data processing agreement covers the processing.

Every step is logged: the original, the extracted values, corrections and approvals stay retrievable together. That helps in audits — and it shows where the process still gets stuck. More on our approach is on the AI automation page.