Bank statement data extraction

Who reads bank statements, and why

Loan processors and underwriters are the heaviest users: they want balances, deposit history and statement dates for income and asset checks, across packets that arrive in a handful of formats. Onboarding analysts use the header alone, because a statement names a client and an address. Bookkeepers want the rows for reconciliation when a bank feed isn't available.

Those jobs need different depths, so an account can switch on only the fields a team asks for. Banks issue statements from a small number of templates: the applicant changes, the layout doesn't. That is the condition docRefinery depends on, and lending and mortgage fits it well.

Bank statement fields docRefinery extracts

There are 41 definitions, including the transaction table and its 12 columns. Deposits and withdrawals have columns of their own, each with a date and description, as well as a general transaction set, which suits statements that split money in and money out and those that share one column.

Example bank statement JSON

A trimmed excerpt with invented values, showing what a short table looks like: this statement prints 48 rows and 46 were read, so the run waits for a person. Extract a document has the full response.

{
  "document_type": "bank_statement",
  "result": {
    "fields": [
      { "alias": "Client Name", "text": "Alex Sample", "field_confidence": 0.97,
        "location": { "page": 1, "bbox": [0.08, 0.14, 0.27, 0.16] } },
      { "alias": "Ending Balance", "text": "2,310.45", "field_confidence": 0.98,
        "location": { "page": 1, "bbox": [0.72, 0.31, 0.86, 0.33] } },
      { "alias": "Overdraft Limit", "text": null, "location": null },
      { "alias": "Table Item", "table": [[
        { "alias": "Transaction Date", "text": "2026-09-03", "field_confidence": 0.96,
          "location": { "page": 2, "bbox": [0.07, 0.21, 0.17, 0.23] } },
        { "alias": "Transaction Description", "text": "EXAMPLE PAYROLL LTD", "field_confidence": 0.94,
          "location": { "page": 2, "bbox": [0.19, 0.21, 0.52, 0.23] } },
        { "alias": "Running Balance", "text": "3,120.45", "field_confidence": 0.95,
          "location": { "page": 2, "bbox": [0.80, 0.21, 0.92, 0.23] } }
      ]] }
    ],
    "extraction_approval": "human-approval",
    "low_confidence_fields": [
      { "alias": "Table Item", "text": null, "threshold": null,
        "reason": "rows_missing", "extracted_rows": 46, "document_rows": 48 }
    ]
  }
}

Why statements suit docRefinery

Statements are long, sensitive and repetitive, and that decides which strengths count. Here is what each one does for a statement desk.

What makes bank statements hard to read

• Length. A busy business account runs to dozens of pages, and the table restarts on each one, sometimes with a balance carried forward that isn't a transaction.
• Sign conventions. Debits may be a separate column, a minus sign, brackets or a DR suffix. Misreading one flips the meaning of a row.
• Wrapped descriptions. One transaction often spans two or three printed lines, and those lines must not become extra rows.
• Two dates. Transaction and posting dates differ, and some rows print only one.
• Weak copies. Printed-and-rescanned statements lose thin rules and small type, exactly where the columns are.
• Formats by country. IBANs, routing numbers and decimal separators all vary. An IBAN check confirms form only, never that the account exists.

How low-confidence statement values are handled

A dropped row is the expensive failure here, because a clean-looking table hides it. Beyond per-cell scoring, docRefinery flags rows_missing with both counts, so an underwriter knows to look. Cells scored under their threshold go to the same review queue, marked with their table and row, and the reviewer can correct a value or add the missing row against the highlighted page. A balance not printed on the statement comes back null, never calculated. A workflow rule can state that starting balance plus deposits less withdrawals equals ending balance; a failure is flagged like any other.