Bank statement processing

This workflow reads a bank statement and returns the header (account, period, balances) as scalar fields and the transactions as rows of an object field. Create it once, then run every statement through it.

The statement

BANK STATEMENT

→

What comes back

account_holder99%

account_number99%

closing_balance99%

transactions

NodesParse → ExtractCredits60 per pageYou get

The graph

{
  "name": "Bank statement processor",
  "nodes": [
    { "id": "parse_1", "type": "parse" },
    {
      "id": "extract_1",
      "type": "extract",
      "extraction_schema": {
        "fields": [
          { "name": "account_holder",         "description": "Name of the account holder as shown on the statement", "data_type": "string" },
          { "name": "account_number",         "description": "Bank account number, may be partially masked",         "data_type": "string" },
          { "name": "statement_period_start", "description": "First day of the statement period",                    "data_type": "date" },
          { "name": "statement_period_end",   "description": "Last day of the statement period",                     "data_type": "date" },
          { "name": "opening_balance",        "description": "Account balance at the start of the period",           "data_type": "float" },
          { "name": "closing_balance",        "description": "Account balance at the end of the period",             "data_type": "float" },
          {
            "name": "transactions",
            "description": "Every transaction listed on the statement, one row each",
            "data_type": "object",
            "nested_fields": [
              { "name": "date",        "description": "Date of the transaction",                                              "data_type": "date" },
              { "name": "description", "description": "Transaction description or memo",                                      "data_type": "string" },
              { "name": "amount",      "description": "Transaction amount: positive for deposits, negative for withdrawals",   "data_type": "float" },
              {
                "name": "type",
                "description": "Type of transaction",
                "data_type": "enum",
                "enum_options": [
                  { "name": "deposit",    "description": "Incoming deposit" },
                  { "name": "withdrawal", "description": "Outgoing withdrawal" },
                  { "name": "fee",        "description": "Bank fee or charge" },
                  { "name": "transfer",   "description": "Transfer between accounts" },
                  { "name": "interest",   "description": "Interest earned or charged" }
                ]
              }
            ]
          }
        ]
      }
    }
  ],
  "edges": [{ "source": "parse_1", "target": "extract_1" }]
}

import os
from anyformat.sdk import Client
from anyformat.workflow import Schema

client = Client(api_key=os.environ["ANYFORMAT_API_KEY"])

workflow = (
    client.workflow("Bank statement processor")
    .parse()
    .extract([
        Schema.string("account_holder",       "Name of the account holder as shown on the statement"),
        Schema.string("account_number",       "Bank account number, may be partially masked"),
        Schema.date("statement_period_start", "First day of the statement period"),
        Schema.date("statement_period_end",   "Last day of the statement period"),
        Schema.float("opening_balance",       "Account balance at the start of the period"),
        Schema.float("closing_balance",       "Account balance at the end of the period"),
        Schema.object("transactions", "Every transaction listed on the statement, one row each", fields=[
            Schema.date("date",          "Date of the transaction"),
            Schema.string("description", "Transaction description or memo"),
            Schema.float("amount",       "Transaction amount: positive for deposits, negative for withdrawals"),
            Schema.enum("type", "Type of transaction", options=[
                Schema.option("deposit",    "Incoming deposit"),
                Schema.option("withdrawal", "Outgoing withdrawal"),
                Schema.option("fee",        "Bank fee or charge"),
                Schema.option("transfer",   "Transfer between accounts"),
                Schema.option("interest",   "Interest earned or charged"),
            ]),
        ]),
    ])
    .create()
)

import { Anyformat, Schema } from "@anyformat/sdk";

const af = new Anyformat({ apiKey: process.env.ANYFORMAT_API_KEY! });

const workflow = await af
  .workflow("Bank statement processor")
  .parse()
  .extract([
    Schema.string("account_holder",       "Name of the account holder as shown on the statement"),
    Schema.string("account_number",       "Bank account number, may be partially masked"),
    Schema.date("statement_period_start", "First day of the statement period"),
    Schema.date("statement_period_end",   "Last day of the statement period"),
    Schema.float("opening_balance",       "Account balance at the start of the period"),
    Schema.float("closing_balance",       "Account balance at the end of the period"),
    Schema.object("transactions", "Every transaction listed on the statement, one row each", [
      Schema.date("date",          "Date of the transaction"),
      Schema.string("description", "Transaction description or memo"),
      Schema.float("amount",       "Transaction amount: positive for deposits, negative for withdrawals"),
      Schema.enum("type", "Type of transaction", [
        Schema.option("deposit",    "Incoming deposit"),
        Schema.option("withdrawal", "Outgoing withdrawal"),
        Schema.option("fee",        "Bank fee or charge"),
        Schema.option("transfer",   "Transfer between accounts"),
        Schema.option("interest",   "Interest earned or charged"),
      ]),
    ]),
  ])
  .create();

Run and read

One statement is one run. Keep each file in its own call so each statement gets its own result; a call that carries several files makes one document packet and one run.

for f in statement-jan.pdf statement-feb.pdf statement-mar.xlsx; do
  curl -X POST "https://api.anyformat.ai/v3/workflows/$WORKFLOW_ID/upload/run/" \
    -H "Authorization: Bearer $ANYFORMAT_API_KEY" \
    -F "files=@$f"
  # -> 202 {"run_id": "...", "status": "queued"}; keep each run_id
done

# Read each run. Repeat until "status" is "processed".
curl "https://api.anyformat.ai/v3/runs/$RUN_ID/" \
  -H "Authorization: Bearer $ANYFORMAT_API_KEY"

for path in ["statement-jan.pdf", "statement-feb.pdf", "statement-mar.xlsx"]:
    result = workflow.run(path).wait()
    print(result.field("account_holder").value, result.field("closing_balance").value)

    # An object field is a list of rows; each row maps a column name to a field
    for row in result.fields["transactions"]:
        if row["type"].value == "fee":
            print("fee:", row["description"].value, float(row["amount"].value))

import { readFile } from "node:fs/promises";
import type { ExtractedField, ExtractedRows } from "@anyformat/sdk";

for (const path of ["statement-jan.pdf", "statement-feb.pdf", "statement-mar.xlsx"]) {
  const file = new File([await readFile(path)], path);
  const run = await workflow.run(file);
  const result = await run.wait();

  // field() returns a scalar or rows; narrow it to what the schema declares
  const scalar = (name: string) => result.field(name) as ExtractedField<string> | undefined;
  console.log(scalar("account_holder")?.value, scalar("closing_balance")?.value);

  // An object field is an array of rows; each row maps a column name to a field
  const rows = result.field("transactions") as ExtractedRows;
  for (const row of rows) {
    if (row.type?.value === "fee") console.log("fee:", row.description?.value, Number(row.amount?.value));
  }
}

What comes back

One statement with six transactions, trimmed to two header fields and two rows. The run took 21 seconds.

Response

{
  "results": {
    "extractions": [
      { "fields": {
        "account_holder":  { "value": "Jordan Reyes", "confidence": 99.0, "evidence": [{ "text": "Account holder: Jordan Reyes", "page_number": 1 }] },
        "closing_balance": { "value": "2983.35", "confidence": 99.0, "evidence": [] },
        "transactions": [
          { "date": { "value": "2026-03-02", "confidence": 99.0 }, "description": { "value": "Salary - Brightline Ltd",  "confidence": 99.0 }, "amount": { "value": "3200.0",  "confidence": 99.0 }, "type": { "value": "deposit",    "confidence": 95.0 } },
          { "date": { "value": "2026-03-04", "confidence": 99.0 }, "description": { "value": "Rent - Harbor Apartments", "confidence": 99.0 }, "amount": { "value": "-1650.0", "confidence": 99.0 }, "type": { "value": "withdrawal", "confidence": 99.0 } },
          "…"
        ] } }
    ] } }

Every field also carries verification_status and value_override. Runs and results has the full envelope, section by section. Read the fields from extractions[0].fields. A scalar field is one object with value, confidence and evidence. An object field is an array of rows, and every cell in a row is that same object. Every value is a string on the wire: numbers as "2983.35", dates as "2026-03-31", booleans as "True" or "False". The SDKs keep them as strings too, so convert before you add them up. A field the model could not find comes back as {"value": null, "confidence": null, "evidence": []}.

Getting the data out

In the app, open the workflow and export the results as CSV, Excel or JSON. Each statement is a row and transactions expands into its own sheet or nested array. See Outputs.
Over the API, read each run as shown above, or list a workflow’s runs for a date range and flatten them yourself. Daily CSV export is a complete script for that.

When it goes wrong

Rows are missing from a long transaction list. A statement runs to hundreds of rows across several pages, and a single standard pass can drop one in a dense table. Ask for the totals the statement prints (total_deposits, total_withdrawals) as extra fields and compare them with the sum of the rows. A mismatch is the cheapest signal that a row was missed. When the tables stay dense, move the Parse node to the agentic tier with "mode": "agentic" at 100 credits per page. Agentic parse to markdown shows the presets. Withdrawals come back positive. amount is a float, and nothing in the type says which direction the money went. Put the convention in the description: “positive for deposits, negative for withdrawals”. The sign then holds across banks that print a Debit column instead of a minus sign. A transaction type comes back empty. type is an enum, so it returns only the names you declared. A statement that prints “standing order” or “card payment” has nothing to match among deposit, withdrawal, fee, transfer and interest. Add the option, or use a string field when the bank’s vocabulary is open. Your balances do not add up. Every value arrives as a string: "2983.35", "-1650.0", "2026-03-31". Cast each one to the type you declared before you do arithmetic on it. A field the model could not find comes back as {"value": null, "confidence": null, "evidence": []}, so test for null before you cast. A batch of statements fails halfway through. Upload and run is limited to 60 requests per minute. Submit statements serially with a short pause between them, or spread them across workers. Feed the workflow the XLSX or CSV the bank already offers when you have the choice: the cells are structured and no OCR runs.

Next steps

Extract node

Object fields, modes and smart lookup

Outputs

CSV, Excel and JSON exports from the app

Daily CSV export

Flatten a day of runs into one CSV over the API

Runs and results

The model behind the envelope

Examples

Read the page

Messy piles

Money and compliance

At scale

Bank statement processing

The graph

Run and read

What comes back

Getting the data out

When it goes wrong

Next steps

Extract node

Outputs

Daily CSV export

Runs and results

​The graph

​Run and read

​What comes back

​Getting the data out

​When it goes wrong

​Next steps

Extract node

Outputs

Daily CSV export

Runs and results

The graph

Run and read

What comes back

Getting the data out

When it goes wrong

Next steps