Skip to main content

Statement formats

Which bank statement formats Assetli reads (standards, national formats, Excel and CSV from any bank) and exactly how they are read.

Last updated: October 1, 2026

Assetli reads statements from banks around the world. An automated test currently checks 203 bank export formats from 31 countries. This page describes which formats those are and what exactly is read from them.

How the format is recognised

What decides is the content of the file, not its extension or the bank you pick in the dialog. Banks often mix up extensions (an .xls statement is frequently HTML, a .txt file can be MT940 or CFONB). So if you select your bank and upload a different export than we expect, the file is read according to what it actually contains.

Recognition order:

  1. Excel (XLSX, legacy XLS, HTML table) and PDF are converted to a table.
  2. Standards and national formats have their own parsers (tables below).
  3. Bank-specific parsers: Air Bank, Raiffeisenbank, UniCredit Bank (Czech Republic) and Wise.
  4. Smart CSV detection for any other tabular file.
  5. Manual or AI column mapping when detection is unsure.

International standards

FormatWhere it is usedExtensionWhat is read
camt.053 (ISO 20022, SEPA XML)euro area, Slovakia, Switzerland, Czech banks following the Czech Banking Association standard.xmlyour own IBAN, the counterparty and their account, reference, batch payments split into individual items
MT940 (SWIFT)Germany, the Netherlands, Austria, Switzerland, Poland, Scandinavia.sta, .940, .mt940, .txtyour own account, counterparty, reference; transaction details in all three common layouts (German ?20, Dutch /TRTP/, Polish ~20 and <20) as well as free text
OFX / QFX / QBOUnited States, Canada, United Kingdom.ofx, .qfx, .qboyour own account, the bank’s transaction ID (reliable duplicate detection), merchant
QIF (Quicken)United States, Canada, Australia, United Kingdom.qifdate, amount, payee, memo; bank accounts, credit cards and cash

For QIF, the order of day and month is determined from the whole file (a single day greater than 12 is enough). If it cannot be determined, US MM/DD is used and the preview says so.

National formats

FormatCountrySpecifics we handle
GPC / ABOCzech Republic, Slovakiafixed-width lines, amounts in hundredths, reversals flip the direction, full counterparty name and message for the payee from supplementary lines
Norma 43 (AEB, Cuaderno 43)Spainseveral accounts in one file, zero-amount movements are skipped, description from supplementary 23 records
CFONB 120Francesign and last digit of the amount packed into one character, the file may have no line breaks, currencies without decimals (XPF)
CBI (RH / EC)Italyvalue date comes before the booking date, payer of an incoming transfer from supplementary lines
CODABelgiuma summary movement with detail items is counted once, structured communication +++…+++ used as the payment reference
ElixirPoland (PKO BP and others)looks like CSV, but amounts are in grosze and the direction is carried by code 111/222

Excel

Upload an Excel statement exactly as your bank gave it to you. These are read:

  • XLSX (Excel 2007 and later),
  • XLS in the legacy binary format,
  • an HTML table with the .xls extension, which many Spanish and Italian banks send.

A date stored as a date in a cell is recognised from the cell format, so 2 March 2026 does not turn into the number 46083. If a workbook has several sheets, the one with the most rows is used. Header lines above the table (account name, period) are skipped just as with CSV.

Excel files are read by our own reader, which only reads cell contents. Formulas are not evaluated (the result Excel saved in the file is used), macros are never run, and a file that would grow beyond 50 MB when unpacked is refused. A password-protected workbook cannot be read; save it without a password, or download CSV from your bank. When you upload a file, Excel is read directly in your browser; with email import, it is read on the server.

PDF

Upload a PDF statement exactly as your bank gave it to you (up to 4 MB). A PDF has no notion of a table; it only holds pieces of text and their position on the page. So the table is rebuilt from them:

  • rows from the vertical position of the text,
  • columns from vertical gaps that are empty in almost every transaction row (the description is usually aligned left and the amounts right, but the gap between them is always in the same place),
  • a wrapped description (a second line with no date and no amount) is joined to the transaction above it,
  • a repeated header on later pages, the introductory details (period, account number) and summaries (closing balance) are skipped.

The resulting table then goes through the same column detection as CSV. When detection is unsure, the column mapping wizard opens.

This works for a PDF with a text layer, that is, one in which you can select the text with your mouse. Statements downloaded from online banking are like that.

Scanned statement

A scanned or photographed statement has no text layer. In that case the import offers to have it transcribed by AI, and before anything is sent it shows you:

  • that the whole statement goes to Anthropic, our AI provider (in every other part of the import, AI never gets the whole statement, only the header and a few rows),
  • the approximate price from your AI credit, based on the number of pages.

Nothing is sent without your confirmation. The AI only transcribes the statement into a table (date, description, counterparty, amount, currency). The columns, date and sign are then decided by the same checks as for CSV, and you see the result in the preview before anything is imported. You pay for actual usage. At most 10 pages can be transcribed at once. Statements sent by email are not read by AI, because consent cannot be given there. If you do not want a transcription, download CSV, Excel or XML from your bank.

CSV from any bank

Smart detection recognises columns by their content, not by the header names. That is why it works even for a bank we have never seen. It knows column names in dozens of languages, but uses them only as a hint.

  • Separator: comma, semicolon, tab and pipe.
  • Numbers: decimal comma and decimal point, thousands separated by a space, dot, comma or apostrophe (1'234.50 in Switzerland), negative amounts in parentheses or with a trailing sign.
  • Dates: 15.07.2026, 15/07/2026, 07/15/2026, 2026-07-15, 20260715, 15 Jul 2026, 02-Mar-26, Mar 02, 2026, 15. července 2026, also with a time. The order of day and month is determined from all rows; if that is not possible, the preview says so and lets you choose.
  • Payment direction:
    • a signed amount,
    • two columns (debit and credit, for example “Soll / Haben”, “Uscite / Entrate”, “Debe / Haber”),
    • a positive amount with a direction column next to it (S/H, D/C, DR/CR, Af/Bij, IN/OUT…).
  • Credit cards: some banks (American Express, Discover, BMO, Nubank) record a purchase as positive and a payment as negative. This is recognised from the export header, or from the fact that all negative rows are payments and refunds. The sign is then flipped so that a purchase is not saved as income.
  • Extra rows: header lines above the table, total rows (“Celkem”, “Total”, “Saldo”) and empty rows are skipped.

When detection is unsure, the dialog shows the columns and you assign them yourself, or let the AI propose the mapping. The AI receives only the header and a few sample rows, never the whole statement.

Character encoding

  • UTF-8 and UTF-16 (also with a BOM) are recognised automatically.
  • National formats have the encoding set by their standard: GPC and Elixir Central European (Windows-1250), CODA, Norma 43, CFONB and CBI Western European (Windows-1252). So a Belgian statement is read correctly even if you have the Czech market selected in Settings.
  • Other files in a legacy encoding are read according to the market in Settings: Czech Republic, Slovakia and Poland as Windows-1250, all other markets as Windows-1252. For markets outside Central Europe, Cyrillic, Greek, Japanese, Korean and Chinese are also recognised from the content.

What we cannot read yet

  • Scanned PDF sent by email. When you upload it, AI can transcribe it (see Scanned statement); by email it cannot.
  • Investment statements in OFX (a list of holdings instead of account movements). Broker trades are imported in the Investments module.
  • Password-protected files.
  • The account number in some GPC statements from Komerční banka. In one variant of the format, KB scrambles the digits of the account number. Transactions are read correctly; only email import cannot tell from the number which account the statement belongs to, and suggests an account based on weaker clues.

Limits

Limit
Statement size (text)8 MB
PDF4 MB
Scanned PDF via AI10 pages
Number of rows60,000
Excel when unpacked50 MB
Email attachment10 MB, at most 20 statements per message

Split a large statement by period. Duplicates from the overlap are skipped.

When the import does not recognise your statement

What is stored automatically. When the import cannot recognise a file with confidence (generic CSV, manual or AI column mapping, or nothing could be read at all), an anonymous record is stored: only the column names from the header, the separator, the number of columns, the selected bank and how the import ended. No data from the rows and no link to your account. If the header contains anything that looks like a value (a date, an amount, an account number, an email address), nothing from it is stored, because it was most likely not a header but the first row of data. These records tell us which bank exports people upload and where the import hesitates. A record is deleted two years after it last occurred.

What you can send us. In the column mapping wizard you will find the link “Help us learn this export”. The statement is anonymised right in your browser:

  • kept: the header, the separator, dates, the number format (decimal comma, thousands separator, sign) and codes that repeat within a column (DEBIT/CREDIT, payment type, currency);
  • replaced: all letters (x/X) and the digits in amounts, account numbers and references; an IBAN keeps only its country code;
  • at most 25 rows are sent.

Before sending, you see the full text that will leave your browser, and you have to confirm it. We store the sample without any link to your account (which is why it cannot be found or deleted on request), and it deletes itself after one year. You can send at most 5 samples a day. In the note, write only the name of the bank and the export, nothing personal.

How we verify coverage

We do not collect real customer statements; they are personal data. Instead, for each bank we keep a description of its export: columns, separator, date and number format, encoding, header lines. Each description cites its source (the bank’s documentation or a public importer) and a confidence level. From the description we generate a statement with eight typical transactions in the language and currency of that country. It goes through exactly the same import as your file, and the result is compared field by field. A row counts as read only if both the date and the signed amount match. A change that would read any format worse does not reach production.

A description can differ from reality, because banks change their formats from time to time. If your statement cannot be read, send us an anonymised sample from the wizard (see above). Please do not send us the full statement.

The import process, duplicate detection and undoing an import are described on the Importing data page. What to do when an import fails is covered in Import: what to do when it fails.