AI document processing
Supplier invoices and delivery notes read automatically, not typed
We build a system that reads supplier invoices, delivery notes, receipts and contracts, extracts the data with OCR and AI, checks it and sends it to your accounting software or ERP. A person confirms only the uncertain fields. It suits companies typing hundreds of documents a month. It starts with a diagnosis on a sample of yours.
Today and after
One document, from the inbox to the books.
A typical flow in a construction company that buys materials from dozens of suppliers, built as an example for this page. It does not describe any client.
Illustrative example
Today
Documents arrive by many routes Manual
PDFs by email, phone photos, paper delivery notes brought in by drivers, invoices from foreign suppliers.
Someone gathers and prints them Manual
Paper goes into a folder; emails are hunted down by hand at the end of the week.
The data is typed into the software Manual
Supplier, number, date, lines, quantities, VAT. A few minutes per document, longer for long ones.
The delivery note is checked against the invoice by eye Manual
Delivered and invoiced quantities are compared on paper, when there is time left.
Mistakes surface at month end Manual
A swapped digit or an invoice entered twice is found only at reconciliation.
Every document passes through two people’s hands, and typing errors are found late.
After
Documents land in one place Automatic
A dedicated email address, a shared folder or a photo from an app. e-Factura invoices download on their own.
The data is extracted automatically Automatic
OCR for scans and photos, then a language model that finds the fields and lines, each with a confidence score.
It is checked against the order and delivery note Automatic
Totals, VAT, tax ID, quantities. What matches moves on; what does not gets a written reason.
Uncertain fields go to a person Human check
The operator sees the document next to the extracted data and confirms or corrects only what is flagged.
The data enters the accounting software Automatic
Through an API or an import, with the original document attached, ready to book.
People check the exceptions instead of typing everything. Corrections are logged, and the rules are adjusted from that log.
When this fits
When it is worth it, and when it is not.
It fits if
- You receive hundreds of documents a month that are not structured XML: invoices from foreign suppliers, receipts, delivery notes, handover reports.
- At least one person spends a large part of the day typing data from PDFs and scans.
- Quantities on delivery notes and invoices are rarely compared, or never, because there is no time.
- Supplier and customer contracts sit in folders, and their deadlines and prices are looked up by hand.
- Formats differ from one supplier to the next, so a fixed template would not work.
It does not fit if
- Almost all your invoices arrive through e-Factura. The XML is read with a parser, without AI, and the right answer is integration, not reading.
- You receive a few dozen documents a month. Typing them costs less than a system.
- Most documents are handwritten and hard to read even for a person.
- There is no accounting software or ERP for the data to go into yet. Then we start with the system, not the reading.
Where it is used
See how it works in your industry.
What we build
The components, one by one.
Each component can be delivered on its own. The diagnosis shows which ones matter for your documents and volume.
- 01
Document capture
A dedicated mailbox, a shared folder, a scanner or a phone upload. e-Factura invoices are downloaded automatically, before their 60-day window closes.
- 02
Data extraction
OCR for images and scans, then a language model that identifies the supplier, number, date, lines and totals. Every field gets a confidence score.
- 03
Validation
We check line sums against the total, VAT rates, the supplier’s tax ID and the match with the order, delivery note or contract. Simple rules run before any model.
- 04
Accounting or ERP integration
Confirmed data goes in through an API, a file import or directly into the database, whichever your software allows. The original document stays attached to the entry.
- 05
Exception queue and notifications
Documents with uncertain fields or mismatches land in a list with a written reason, with the person who can resolve them. Nothing gets lost in an inbox.
- 06
Reporting
How many documents came in, how many passed with no correction, which fields get corrected most and for which suppliers. That shows where the rules are worth adjusting.
Do the math
What does typing documents cost you today?
Enter your own numbers. The result is the time spent on manual data entry today, not a promised saving. How much of it can be taken over is what the diagnosis measures on your documents.
Estimate
- Manual time per month
- 40 hours
- Cost per month
- €1,200 per month
Per year: €14,400
Assumptions
- With the starting values: 600 documents × 4 minutes = 40 hours a month. At €30 an hour, that is €1,200 a month.
- The minutes cover opening the document, typing the fields, comparing with the delivery note or order and filing. An invoice with many lines takes longer than a receipt; use an average measured on your own documents.
- €30 an hour is a starting value for the full cost of an hour of administrative work, not an official average. Use the real cost in your company.
- The estimate leaves out the cost of typing errors found late, and of payments made for quantities that were never delivered.
- Automatic reading does not reach zero minutes either: documents flagged as uncertain still need a check. It is an indicative estimate, not a quote and not a guaranteed saving.
How we work
From a pile of documents to data you can rely on.
The first step is The Diagnosis: we work on a sample of your real documents, not on demo examples, and decide together what is worth reading automatically.
- 01
Diagnose
We collect a sample of real documents from recent months, count them by type and supplier, and measure how long each type takes today.
Output: document types, volumes, times and where errors appear.
- 02
Design
We set the fields to extract for each type, the validation rules and the confidence threshold below which a person confirms. We choose where the system runs.
Output: the field list, written rules and review thresholds.
- 03
Build
We build the flow and run it alongside manual entry, on real documents. We compare results field by field and measure accuracy on your sample.
Output: accuracy measured on your documents, not promised in advance.
- 04
Scale
We add further document types one at a time, such as contracts or handover reports, tune the thresholds on real corrections and hand over the documentation.
Output: a documented flow, a monthly report and an owner on your team.
Risks
What can go wrong, and how we handle it.
Personal data and GDPR
Receipts, contracts and some invoices hold names, addresses or personal ID numbers. We extract only the fields that are needed, limit access to the people who work with them and sign a data processing agreement for any data we can access.
Reading errors and human review
A model can misread a digit, especially on a poor scan. That is why every field has a confidence threshold, totals are checked by calculation and a person decides below the threshold. We also review a sample of the documents that passed automatically, on a regular schedule.
When a service is down
If the AI service, e-Factura or SPV does not answer, documents wait in a queue and are retried automatically. If the outage lasts, they move to the manual work queue, so nothing is stuck at month end.
Where the data runs
We use services that process data in the EU, do not use it to train their models and sign a data processing agreement. Original documents can stay on your servers or in a cloud with EU data centers.
Questions about automatic document reading
How accurate is the reading?
It depends on your documents: a PDF generated by software reads far more reliably than a photo of a crumpled receipt. That is why we do not promise a number in advance. We measure accuracy on a sample of your documents, and uncertain fields go to a person anyway.
Do we have to change our accounting software?
No. Extracted data goes into the software you already use, through an API or a file import. If it accepts neither, we find a route in the diagnosis or tell you honestly that it is not worth it.
Do we need AI for e-Factura invoices too?
No. e-Factura invoices are structured XML files, so a standard parser reads them without an AI model. AI makes sense for everything that arrives unstructured: invoices from foreign suppliers, receipts, delivery notes, contracts and scans.
e-Invoicing automationWhat about handwritten or badly scanned documents?
Legible ones are read, but with lower confidence, so they go to review. Illegible ones are flagged as such and go straight to the manual queue. The rule is simple: what cannot be read reliably goes to a person, not into the books.
Does the data leave the EU?
It does not have to. We choose services that process data in the EU and do not train on it, and the documents can stay on your servers. We decide together where each part runs, before anything is built.
How long until it works?
It depends on how many document types you have and how much they differ. We set the duration in the diagnosis and write it into the proposal. We start with the most frequent type, usually supplier invoices, and add the rest one at a time.
Can it read contracts too?
Yes, but for contracts we do not extract the whole text, only the fields that matter to you: the parties, the value, payment terms, expiry date, penalty clauses. The result is a list you can search and filter, with a pointer to the page in the contract.
What happens to the documents after they are read?
The original stays attached to the accounting entry and is kept in the company archive, searchable by supplier, number or date. We do not delete or change anything in the source documents.
Further reading
Read more on this.
Tell us which documents are still typed by hand in your company.
Which types, how many a month and which software they end up in. We reply honestly on whether automatic reading is worth it, and what a diagnosis would cover.