Someone retypes invoices into accounting by hand
Try thisHover an extracted field — the spot it came from lights up.
| Description | Qty | Unit price | Amount |
|---|---|---|---|
| Server management — August 2026 | 1 ks | 18 500,00 | 18 500,00 |
| Data backup, 500 GB | 5 ks | 1 950,00 | 9 750,00 |
| Technical support | 20 h | 1 000,00 | 20 000,00 |
Select a field — its source region is highlighted in the document on the left.
- normalised from “28. 8. 2026”
- normalised from “11. 9. 2026”
- Needs reviewThe symbol is overwritten by hand in the document. The model offers two readings and leaves the choice to a human.
Fields below 0.90 confidence are never posted automatically — they wait for a human.
- 21fields from one invoice
- 0.90confidence threshold for posting without a person
- 1field the demo sends for review
The demo runs on made-up data. Figures describe the demo set-up, not client results.
I want thisField extraction from invoices and contracts, with a review path
An invoice arrives as paper. It leaves as a record.
Someone retypes supplier invoices into the accounting system every month. The model reads the document, pulls out the fields and shows where each value came from — hover a field and the spot on the page lights up. Anything it is not sure about stops and waits for a human instead of guessing; that is the part that decides whether a deployment survives its first month.
Who it is forFinance teams, accounting firms, shared invoice inboxes
Delivered as part ofAI implementation
What the supplier gets
Sends an invoice the way they always have — e-mail, PDF, a photo of paper.
Finance
Records land in the accounting system already filled in, with a link back to the source. Only the uncertain ones reach a person.
How it worksStep by step
The document arrives as it is
A PDF from e-mail, a scan, a phone photo. The model does not need a uniform template.
Reading the page
Text and layout together — where the supplier block is, where the line items are, where the VAT summary sits.
Pulling the fields
Company ID, document number, dates, variable symbol, amounts, rates. Each value keeps the page region it came from.
Uncertain goes to a person
A struck-through variable symbol, rewritten by hand. The model offers both readings and waits instead of picking one.
Posting and the trail
The record lands in accounting. Every field keeps a link back to the document, so any value can be traced later.
What it doesWhat it needs in production
Field extraction with provenance
Every value carries the region of the page it was read from, so a disputed number can be checked in one click instead of reopening the PDF.
Confidence threshold, not blind trust
Below the configured threshold nothing is posted automatically. The field goes to a queue with the candidate readings the model was choosing between.
Normalisation
Dates, amounts, VAT rates and account numbers are converted to one canonical form, so downstream systems do not each parse them differently.
Matching against orders
Extracted lines are matched to purchase orders and deliveries; differences are flagged rather than silently accepted.
Writes into the system you already use
Pohoda, Money, Helios, ABRA or a plain export. Extraction is not worth much if the result has to be retyped again.
Learning from corrections
What a person fixes in the queue becomes a correction rule for that supplier layout, so the same mistake does not return every month.
Who it is forWhere it makes sense
Accounting firms
Hundreds of invoices a month from dozens of clients, each in a different layout. Extraction handles the layouts; the review queue handles the exceptions.
Manufacturing and logistics
Delivery notes, customs documents, freight bills. Matching against orders is where the manual work actually is.
Municipalities and public bodies
Incoming mail, applications, contracts. Data stays in the EU and the audit trail shows which source each value came from.
Any company with a shared invoice inbox
The mailbox where invoices pile up and someone retypes them. That is the whole use case.
IntegrationsRuns on what you already have
- Pohoda
- Money S3/S4
- Helios
- ABRA
- Microsoft 365 mailbox
- Google Workspace
- SharePoint / OneDrive
- ISDOC / e-invoicing
- REST API
- CSV / XML export
The list is not exhaustive. We connect systems that are not here as long as they have an interface.
Want this in your business?
A no-obligation call with someone who builds these. We go through your brief and say what is realistic and what is not.
Book a consultation