SCYTIAN
Menu

PLATFORM / EXTRACTING DATA

Files in. Data out. The wrong one stays out.

PDF, Word, Excel, CSV, XML, a scan or a photo taken in a car. Whatever the file, the values end up in your system, checked against what you already know on the way in, instead of typed across and hoped about.

Get your agent
Back office{{ status }}
CHAT
INV-2026-0418.pdf212 kB

{{ typed }}

{{ agentText }}

1It reads the file
INV-2026-0418.pdfa scan · 212 kB · crookedPDF
2…pulls out the values and checks each one
{{ f.k }}{{ f.v }}{{ f.tag }}
3…and writes only what passed
YOUR ACCOUNTINGincoming invoices
Orion Construction · 04163 180.00
Meridian Insurance · 0417640.00
ACME Industrial · 04181 240,00
Held. On your list, with the reason

No purchase order for this supplier this month. Nothing was written, not the ten good fields either. The file waits with this sentence next to it.

{{ dest.note }}
Message Back office…

{{ foot }}

WHAT YOU CAN THROW AT IT

Whatever shape the data is stuck in.

Nobody asks your suppliers or your own people to change how they send things. The tidy file and the awkward one are the same job.

PDFPDFGenerated, where the text is really in there, or a scan, which is a picture of a document and nothing more.
DOCXWordTables somebody stopped keeping consistent three years ago, and a heading that has moved twice since.
XLSXExcelMerged cells, the real header on row four, three sheets where two matter, formulas that have to keep working.
CSVCSVSemicolon or comma, Windows encoding, decimal comma. Simple until it is somebody else’s.
XMLXML and JSONThe easy case: the structure is already there. Until the other side changes the schema and tells nobody.
JPGScan or photoCrooked, shadowed, taken at an angle on a phone in a car. Still has to give the same numbers.

HOWEVER IT ALREADY MOVES

a mailboxa folder or a drivethe office scannera portal you log intoan API or a feed

WHAT PEOPLE PUT THROUGH IT

incoming invoicesprice listspurchase ordersdelivery notesbank statementsproduct feedsreceiptstimesheetscontractsand more

As e-invoicing becomes mandatory country by country, invoices arrive as structured files. That is the easy case, not a new project, just one more shape coming in.

WHY THAT ONE FIELD STOPPED IT

“Checked” means something specific.

Every tool claims it validates. Here is what it actually compares, against your data, not against itself. Which checks run is agreed before anything goes live.

CHECKS · INV-2026-0418.pdf3 checks · 2 passed · 1 held
The lines have to add up
net1 033.33
+ VAT 20 %206.67
= total on the file1 240,00 ✓

Items against the total, net and VAT against the sum, the rate against the date. A cent of rounding is fine, a hundred is not. That tolerance is a number you set.

It has to be somebody you know
on the fileACME Industrial
company ID36 235 112
= your supplier#A-1042 ✓

Matched by registration number to a record you already have. Never by name. Names are spelled four ways, and that is how one company becomes four suppliers.

And it has to be expected
purchase ordernone this month
seen before?no · first time
→ held for youwith reason

An order it came against, and no earlier copy of the same document, forwarded again, scanned again, sent by two people. The same invoice paid twice is a real thing.

a price list gets different checks than an invoice · decided in the plan, before anything goes live

HOW YOU KNOW IT IS RIGHT

Measured on your files before it is switched on.

Not a demo on three clean examples. You confirm the correct answer on a real batch of your own files, including the ugly ones, and the platform scores itself against that, field by field. Below the agreed number, it does not go live.

1 · YOU SAY WHAT IS RIGHT

On a real batch of your own files

file 12 of 48 · INV-0412.pdf: supplier Atlas Petrochem Ltd · total 1 249,90 · cost centre: not on the file → empty ✓. You confirm or fix a value.

real files · including the ugly ones
2 · IT IS SCORED AGAINST YOU
99.1%gate 98 %
523 of 528 fields right
supplier · ID · invoice no.100 %
dates · totals99.0 %
line items · 5 wrong97.4 %
the 5 misses are listed, with the file
3 · ONLY THEN THE SWITCH

Write into accounting, live since Tuesday 06:00

99.1 % ≥ 98 % gate · 48 files confirmed by you. Below the gate? The switch stays off. Nothing is written until the number is right. Not a promise, a lock in the platform.

the same gate every agent goes through

WHERE THE DATA ENDS UP

Somewhere you actually work, not in another export.

The result is written where the work happens, and the original file stays attached, so a year later you can still see what the number came from.

Into the system you already use

Your accounting package, your ERP, your CRM, written the way your own people write it. Same account, same cost centre, same wording.

through its API, not through a person

Onto the record it belongs to

The delivery note ends up on the order it came against, the price list on the supplier. Attached and findable, not filed in a folder nobody opens.

the file stays with the data

Into a table, if there is no system for it

Plenty of processes have no software behind them. Then the result is a table you can read, sort and hand over, and move into a real system later.

no system is not a blocker

THE LIMITS

What it will not do with a file it cannot read properly.

It does not invent who somebody is

If the company on the file is not one it can identify, it does not create a record and does not pick the closest name. Guessing identity is how duplicate customers are born.

Missing stays missing

A field the file simply does not carry comes back empty and says so. It is never filled with a plausible value so the record looks complete.

Not switched on before it is measured

No accuracy number, no live writing. That is a gate in the platform, not a promise made on a call.

Name the file somebody opens every day and retypes into something else.

Send us twenty of them. Your agent arrives set up for them and shows you its score on those files before it writes anything.

Get your agent