Bank statements to JSON, with balance-chain integrity

What we extract

Every field, with confidence and citations

Every field comes back with a confidence score in [0,1] and a normalized bounding box on the source page. Your agent decides on its own when to escalate to human review.

{
  "schema_version": "bank_statement.v2",
  "result": {
    "institution_name": "KBC Bank",
    "account_holder": "Pieter Janssens",
    "account_iban": "BE68539007547034",
    "currency": "EUR",
    "statement_period_start": "2026-04-01",
    "statement_period_end": "2026-04-30",
    "opening_balance":  { "amount": "2840.50", "currency": "EUR" },
    "closing_balance":  { "amount": "1956.30", "currency": "EUR" },
    "transactions": [
      {
        "posted_date": "2026-04-03",
        "value_date":  "2026-04-03",
        "description": "Maandelijkse loon Acme NV",
        "amount":      { "amount": "2950.00", "currency": "EUR" },
        "transaction_type": "credit",
        "balance_after":   { "amount": "5790.50", "currency": "EUR" },
        "counterparty_iban": "BE54000123456789",
        "confidence": 0.97,
        "bbox": { "page": 1, "x": 0.06, "y": 0.34, "w": 0.88, "h": 0.018 }
      }
    ],
    "validation": {
      "balance_chain": {
        "valid": true,
        "computed_closing":  "1956.30",
        "declared_closing":  "1956.30",
        "diff": "0.00",
        "tolerance": "0.01"
      }
    }
  },
  "field_metadata": {
    "closing_balance.amount": { "confidence": 0.98 },
    "account_iban":            { "confidence": 0.99 }
  }
}

Domain-specific validation

What makes this a specialist

Format coverage

Tested across 50+ EU bank formats

50+ EU banks · 10 US banks · 6 countries with multi-bank coverage

Code recipes

From document to JSON in five lines

curl -X POST https://eu-api.tryparsr.dev/v1/parse/bank_statement \
  -H "Authorization: Bearer $PARSR_API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
    "document_url": "https://example.com/april.pdf",
    "wait": 60
  }'

import os, httpx

resp = httpx.post(
    "https://eu-api.tryparsr.dev/v1/parse/bank_statement",
    headers={"Authorization": f"Bearer {os.environ['PARSR_API_KEY']}"},
    json={
        "document_url": "https://example.com/april.pdf",
        "wait": 60,
    },
    timeout=70,
)
result = resp.json()["result"]
chain = result["validation"]["balance_chain"]
if not chain["valid"]:
    raise ValueError(f"Balance chain broken — diff {chain['diff']}")
for tx in result["transactions"]:
    print(tx["posted_date"], tx["amount"]["amount"], tx["description"])

const resp = await fetch("https://eu-api.tryparsr.dev/v1/parse/bank_statement", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.PARSR_API_KEY}`,
    "Content-Type": "application/json",
    "Idempotency-Key": crypto.randomUUID(),
  },
  body: JSON.stringify({
    document_url: "https://example.com/april.pdf",
    wait: 60,
  }),
});
const { result } = await resp.json();
if (!result.validation.balance_chain.valid) {
  throw new Error(`balance chain broken — diff ${result.validation.balance_chain.diff}`);
}

from langchain_parsr import ParsrToolkit
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent

tools = ParsrToolkit.from_env().get_tools()
agent = create_react_agent(ChatOpenAI(model="gpt-4o"), tools)

result = await agent.ainvoke({
    "messages": [(
        "user",
        "Parse this April KBC statement and tell me total income vs spend: "
        "https://example.com/april.pdf"
    )]
})
print(result["messages"][-1].content)

What teams build with this

Three integration patterns

Compared

How parsr's bank-statement parsing compares

FAQ

Common questions

• Which bank formats are supported?

  50+ EU banks across BE / DE / FR / NL / UK plus 10 US banks. Each format has a dedicated detail page under /banks/{country}/{bank} with format quirks documented. If your bank isn't listed, email a sample to support@ — we ship new formats in 48 hours.

• How does balance-chain validation work?

  Every parsed statement returns validation.balance_chain with valid (bool), computed_closing, declared_closing, diff, and tolerance (1-cent default). Computed = opening + Σ(credits − debits). If diff > tolerance, valid=false — the strongest single signal that a statement has missing transactions, sign errors, or has been tampered with.

• Multi-currency statements?

  Yes. Each transaction returns its own currency in transaction.amount.currency. Statement-level currency is the dominant currency. Mixed-currency statements return validation.currency_consistency reporting any drift between statement-level and per-transaction currency.

• What about scanned / photo statements?

  Vision LLMs handle PDF, JPEG, PNG, and HEIC. Image quality matters — phone photos work down to ~150 DPI equivalent. Statements with deep skew or partial occlusion drop confidence; field_metadata.<field>.confidence < 0.85 is the recommended escalation threshold for review.

• Where is the data processed?

  Region-bound API keys (sk_eu_… vs sk_us_…) route to either eu-api.tryparsr.dev (Exoscale Frankfurt, EU operator) or us-api.tryparsr.dev (Hetzner Ashburn). EU keys never leave the EU. Object storage uses Cloudflare R2 with jurisdiction='eu' for the EU bucket — true residency, dedicated EU endpoint.

• Can I cache results to avoid re-billing on retries?

  Yes — file-hash caching is built in. Re-uploading the same bytes within 30 days hits the cache, returns the original parse, and is not billed (billable=FALSE in usage events). Idempotency-Key on the request makes retry-safety explicit at the request level.