AI PDF to JSON & Invoice OCR: Receipts, Bank Statements & Any Document

Convert PDFs, scanned documents, photos, DOCX and HTML to clean, validated JSON, plus CSV you can open in Excel or Google Sheets. Use it as an invoice OCR and parser, a receipt OCR API, a bank statement to CSV/Excel converter with balance reconciliation, or a general document parser with your own fields. Every document comes back with typed values, line items, per-field confidence scores and math checks: line items vs subtotal, subtotal + tax = total, and running balances.

No API keys needed. The AI runs through Apify's official OpenRouter proxy and is billed at cost to your Apify account. You can also bring your own OpenRouter key. You pay $0.01 per successfully extracted document (up to 5 pages included). Failed documents are free.

What does this PDF to JSON extractor do?

• Reads any document: text PDFs, scanned PDFs, JPG / PNG / WebP / TIFF / GIF photos, DOCX, HTML and TXT. Files can come from URLs (including Google Drive and Dropbox share links), file uploads, an Apify key-value store, or a dataset from another Actor.
• Uses the cheapest method that works: if a page has a text layer, that text is used, which is cheap and exact. The vision model only reads pages that are scans or photos.
• Extracts structured data: 9 ready-made presets with line-item tables, auto-detection of the document type, or your own fields and JSON Schema.
• Normalizes values: ISO 8601 dates, ISO 4217 currency codes and real numbers. It handles 1.234,56, 1,23,456.00, (120.00) and Rs. 450/-, and checks IBANs and container numbers.
• Validates the result: quantity × unit price, line items vs subtotal, subtotal + tax = total, tax breakdown, amount due, bank statement balance reconciliation, running balances, due date ≥ invoice date and required fields.
• Tells you what to trust: each field gets a confidence score. It is lowered automatically when a value cannot be found in the document text or when a check fails. Any document that needs a human look gets needsReview: true with reasons.
• Exports to spreadsheets: RESULTS.csv has one row per document, and TABLE-lineItems.csv (or transactions, etc.) has one row per line. Both open cleanly in Excel and Google Sheets.

Use cases: invoice OCR, receipt OCR, bank statements to CSV/Excel and more

• Invoice and receipt OCR for accounts payable and bookkeeping: batch-process supplier invoices and receipts into your ERP, Xero, QuickBooks, Tally or a spreadsheet.
• Bank statement to CSV / Excel for fintech, lending and bookkeeping: turn bank and card statements into transactions, with balance reconciliation so you can see missing or misread rows.
• Operations and logistics: purchase orders, bills of lading and shipping documents.
• HR and recruiting: parse resumes into structured profiles.
• Developers and automation builders: a single API call (or Make / Zapier / n8n step) that turns any file into JSON you define.

Supported documents and presets

Preset	Typical fields
Auto-detect	Picks the best preset below for each document
Invoice / bill	Number, dates, PO, vendor, customer, line items, subtotal, tax breakdown, total, amount due, bank details
Receipt	Merchant, date, items, tax, tip, total, payment method
Bank / card statement	Account, period, opening and closing balance, transactions (reconciled)
Purchase order	PO number, buyer, supplier, line items, totals, delivery details
Resume / CV	Contact details, experience, education, skills
Business card	Name, title, company, phones, emails, address
Generic form / any document	Key-value pairs and tables
Contract key terms	Parties, dates, term, renewal, payment terms, governing law
Bill of lading / shipping document	Shipper, consignee, ports, containers (number check), goods, weights
Custom schema only	Only the fields you define

You can add your own fields to any preset.

Why this Actor?

• Your schema, not ours. Add fields to a preset, or describe the whole output with a simple field list or a JSON Schema. Most extractors on the Store only return raw text.
• Validated, not just extracted. Math and format checks catch the classic AI mistakes, such as a dropped line, a misread total or a swapped day/month.
• Confidence you can act on. Use needsReview to send only uncertain documents to a human.
• Cheap by design. Text-first processing and a small default model (google/gemini-2.5-flash-lite) cost about $0.0002–$0.001 of AI per document. Switch to the accurate model for hard scans.
• Reliable. Retries, automatic fallback to a second model, chunking for long documents, and failed documents are free.
• Private by default. Nothing is stored outside your Apify storage, and the model is told to ignore any instructions written inside documents. Government ID documents (passports, ID cards, driver's licenses) are refused and not charged.

How to use it

1. Add documents under Document URLs, upload files, or point to a key-value store / dataset.
2. Choose a Document type: Auto-detect, a preset, or Custom fields only.
3. Optional: add Custom fields, for example vatNumber (string): Seller VAT ID or totalAmount* (number).
4. Click Start. Each document becomes one dataset row. Download it as JSON, CSV or Excel, or use the CSV files in the key-value store.

You can call it from the API, schedule it, or chain it after another Actor that collects file URLs (use Dataset ID + URL field).

Input example

{
  "documentUrls": [
    { "url": "https://api.apify.com/v2/key-value-stores/xBqlN4kRAyAIWZpIU/records/sample-invoice.pdf" },
    { "url": "https://api.apify.com/v2/key-value-stores/xBqlN4kRAyAIWZpIU/records/sample-receipt-scan.jpg" }
  ],
  "documentType": "auto",
  "customFields": ["costCenter (string): internal cost center if printed"],
  "model": "fast",
  "saveCsvFiles": true
}

Custom-only example (a utility bill, no preset):

{
  "documentUrls": [{ "url": "https://api.apify.com/v2/key-value-stores/xBqlN4kRAyAIWZpIU/records/sample-utility-bill.pdf" }],
  "documentType": "custom",
  "customFields": [
    "accountNumber*: utility account number",
    "provider*: utility company name",
    "billingPeriodStart (date)",
    "billingPeriodEnd (date)",
    "usageKwh (number): electricity used in kWh",
    "amountDue* (number)",
    "currency (currency)",
    "dueDate* (date)"
  ]
}

That run returned {"accountNumber": "7731-0042-19", "provider": "Lakeside Power & Light", "billingPeriodStart": "2026-09-01", "billingPeriodEnd": "2026-09-30", "usageKwh": 632, "amountDue": 125.78, "currency": "USD", "dueDate": "2026-10-25"} for $0.000166 of AI.

Field-list syntax: name* (type): description. * marks a required field. Types are string, number, integer, boolean, date, currency, email, phone, url, percent, array, and lists such as date[]. For nested objects and tables, use Custom JSON schema.

Output example

A real row from a cloud run (2-page text invoice, default model; long arrays shortened):

{
  "documentId": "e5746a62b850",
  "sourceUrl": "kvs://xBqlN4kRAyAIWZpIU/sample-invoice.pdf",
  "sourceType": "key-value-store",
  "documentType": "invoice",
  "documentTypeTitle": "Invoice",
  "detectedDocumentKind": "Invoice",
  "status": "ok",
  "needsReview": false,
  "reviewReasons": [],
  "overallConfidence": 1,
  "validationPassed": true,
  "data": {
    "invoiceNumber": "NWC/2026-27/0457",
    "invoiceDate": "2026-10-14",
    "dueDate": "2026-11-13",
    "purchaseOrderNumber": "PO-88213",
    "currency": "INR",
    "vendor": {
      "name": "Northwind Cloud Services Pvt. Ltd.",
      "address": "4th Floor, Example Tech Park, Outer Ring Road Bengaluru, Karnataka 560103, India",
      "taxId": "29AAACN1234F1Z5",
      "email": "billing@northwind-cloud.example",
      "phone": "+91 80 4000 1234"
    },
    "customer": {
      "name": "Acme Retail Solutions LLP",
      "address": "22 MG Road, Pune, Maharashtra 411001",
      "taxId": "27AAFFA9876K1Z2",
      "email": null,
      "phone": null
    },
    "shipTo": null,
    "lineItems": [
      {
        "description": "Cloud VM hosting - 8 vCPU / 32 GB (monthly)",
        "sku": "998315",
        "quantity": 3,
        "unit": "pcs",
        "unitPrice": 12500,
        "taxRate": null,
        "amount": 37500
      },
      {
        "description": "Managed PostgreSQL database (monthly)",
        "sku": "998315",
        "quantity": 1,
        "unit": "pcs",
        "unitPrice": 8750,
        "taxRate": null,
        "amount": 8750
      },
      "... 2 more line items"
    ],
    "subtotal": 63650,
    "discount": null,
    "taxAmount": 11457,
    "taxBreakdown": [
      {
        "name": "CGST",
        "rate": 9,
        "amount": 5728.5
      },
      {
        "name": "SGST",
        "rate": 9,
        "amount": 5728.5
      }
    ],
    "shipping": null,
    "total": 75107,
    "amountPaid": 0,
    "amountDue": 75107,
    "paymentTerms": "Net 30 days from invoice date.",
    "bankDetails": {
      "bankName": "Example Bank Ltd., Koramangala Branch",
      "accountName": "Northwind Cloud Services Pvt. Ltd.",
      "accountNumber": "50200012345678",
      "iban": null,
      "swiftBic": null,
      "ifsc": "EXMP0001234",
      "routingNumber": null
    },
    "notes": "Late payments attract interest at 1.5% per month.\nAll disputes are subject to Bengaluru jurisdiction.\nThis is a computer-generated invoice."
  },
  "confidence": {
    "invoiceNumber": 1,
    "invoiceDate": 1,
    "dueDate": 1,
    "purchaseOrderNumber": 1,
    "currency": 1,
    "...": "one score per field"
  },
  "lowConfidenceFields": [],
  "ungroundedFields": [],
  "validation": {
    "passed": true,
    "checksRun": 10,
    "checksFailed": 0,
    "checks": [
      {
        "check": "lineItems.lineMath",
        "passed": true,
        "message": "quantity x unit price = amount on 4/4 lines",
        "fields": [
          "lineItems"
        ]
      },
      {
        "check": "lineItems.sumMatchesSubtotal",
        "passed": true,
        "message": "Sum of 4 lineItems amount values (63,650.00) matches subtotal (63,650.00)",
        "fields": [
          "lineItems",
          "subtotal"
        ],
        "expected": 63650,
        "actual": 63650
      },
      {
        "check": "totalsAddUp",
        "passed": true,
        "message": "subtotal + taxAmount = 75,107.00; total is 75,107.00",
        "fields": [
          "subtotal",
          "total",
          "taxAmount"
        ],
        "expected": 75107,
        "actual": 75107
      },
      "... 7 more checks"
    ]
  },
  "coercionIssues": [],
  "warnings": [],
  "fileName": "sample-invoice.pdf",
  "fileType": "pdf",
  "fileSizeBytes": 4113,
  "fileSha256": "fb9cf7c42da1b10abd918cef369b1a7c8f6815c2d0d07a8102e89f8075638681",
  "pageCount": 2,
  "pagesProcessed": 2,
  "pagesTextLayer": 2,
  "pagesVision": 0,
  "extractionMethod": "text",
  "language": "en",
  "chunks": 1,
  "model": "google/gemini-2.5-flash-lite",
  "modelsUsed": [
    "google/gemini-2.5-flash-lite"
  ],
  "fallbackUsed": false,
  "route": "apify-openrouter-proxy",
  "classification": {
    "documentType": "invoice",
    "confidence": 1
  },
  "aiCostUsd": 0.000684,
  "promptTokens": 2355,
  "completionTokens": 1122,
  "llmCalls": 2,
  "attempts": 2,
  "durationMs": 3765,
  "error": null,
  "extractedAt": "2026-10-22T11:29:42Z"
}

Main fields

Field	Meaning
status	ok, error, refused (government ID) or skipped
documentType / detectedDocumentKind	Preset used, and what the model says the document is
data	The extracted JSON. Fields always exist (null when not present) and values are typed
confidence	0–1 per top-level field, lowered when ungrounded or failing validation
needsReview / reviewReasons	true if confidence < threshold, a check failed, or required fields are missing
validation	Every check with expected vs actual values
ungroundedFields	Values the model returned that could not be found in the document text
extractionMethod	text, vision or mixed (per page)
pageCount / pagesProcessed	Pages in the file and pages read
aiCostUsd	Real AI cost of this document as reported by OpenRouter, before any plan multiplier
model / modelsUsed / fallbackUsed	Which model produced the result
flatData	data flattened to vendor.name-style columns, which power the CSV / Excel view
error	{code, message} when the document failed

Dataset views

• Overview: one line per document with type, status, review flag, confidence and cost.
• Extracted data: the JSON.
• Flat (spreadsheet): one column per field. This is the best view for CSV / Excel export.
• Validation: checks and review reasons.
• Errors: failed and refused documents.

Key-value store

• RESULTS.csv: one row per extracted document (UTF-8 with BOM, opens correctly in Excel).
• TABLE-<name>.csv: one row per line item / transaction / table row, for example TABLE-lineItems.csv and TABLE-transactions.csv.
• OUTPUT: run summary with totals, review counts, AI cost, errors and CSV links.
• An optional webhook gets the summary as JSON when the run finishes.

Error codes

Code	Meaning
DOWNLOAD_FAILED / DOWNLOAD_TIMEOUT	The URL could not be downloaded (404, timeout, login required)
INVALID_URL / BLOCKED_URL	Not a valid public http(s) URL (private network addresses are blocked)
FILE_TOO_LARGE / EMPTY_FILE	Above Max file size, or 0 bytes
UNSUPPORTED_FORMAT	Not a PDF, image, DOCX, HTML or text file (for example XLSX, PPTX, old .doc, HEIC)
PDF_PASSWORD_PROTECTED / INVALID_PDF / INVALID_IMAGE	The file cannot be opened
NO_PAGES / NO_TEXT	Nothing readable in the document
PROXY_NOT_AVAILABLE	Apify's OpenRouter proxy refused the request. Add your own OpenRouter key
INSUFFICIENT_CREDITS / AUTH_ERROR	The AI account has no credit or the key is wrong
RATE_LIMITED / UPSTREAM_ERROR / MODEL_UNAVAILABLE / TIMEOUT	AI provider problems that remained after retries and the fallback model
BAD_JSON / EMPTY_ANSWER	The model returned unusable output after a retry

Government ID documents get status: "refused", error code GOVERNMENT_ID_NOT_SUPPORTED and no data. Failed documents are never charged.

Export to Excel / Google Sheets

• RESULTS.csv in the key-value store has one row per document, and TABLE-lineItems.csv / TABLE-transactions.csv have one row per line. They're UTF-8 with BOM, so they open cleanly in Excel.
• The dataset's Flat (spreadsheet) view can be downloaded as Excel (XLSX), CSV or JSON from the Apify Console or API.
• Use Apify's Google Sheets integration, Make, Zapier or n8n to push results into your sheet or accounting tool.

How much does document OCR cost?

You pay two things:

1. This Actor's fee: $0.01 per successfully extracted document ($10 per 1,000; document-extracted event), including up to 5 pages. Longer documents add $0.001 per extra page (extra-page event). Failed, refused and skipped documents are free, and Apify platform compute is included.
2. The real AI cost, at cost:
   • No key (default): billed to your Apify account through Apify's official OpenRouter proxy, at the OpenRouter price on paid plans and at 10× on the Apify FREE plan.
   • Your own OpenRouter key: billed by OpenRouter directly.

Example: 1,000 invoices of up to 5 pages = $10 in Actor fees, plus roughly $0.20–$1 of AI on the default model (paid plans).

Measured AI cost per document (real cloud runs, October 2026, before any plan multiplier):

Document	Model	AI cost
2-page text invoice, 4 line items	google/gemini-2.5-flash-lite (default)	$0.0007 – $0.0008
Scanned receipt (photo / image PDF)	google/gemini-2.5-flash-lite	$0.0003 – $0.0007
1-page utility bill, custom fields	google/gemini-2.5-flash-lite	$0.0002
Same documents, accurate model	google/gemini-2.5-flash	roughly 3–6× the above (estimate)

Auto-detect adds one tiny classification call. Choose a preset to skip it. Each document used about 60–80 MB of memory and a few seconds of compute at 256 MB.

Tip for the Apify FREE plan: add your own OpenRouter API key to avoid the 10× proxy multiplier.

Using your own API key (optional)

Add OpenRouter API key to send AI requests to OpenRouter with your key. You can then also use any OpenRouter model with Model → Custom. The key is a secret input: it is encrypted by Apify and never logged or written to output.

FAQ

How accurate is it? On clean text PDFs, the default model returned every field correctly in our tests, and all math checks passed. For poor scans, handwriting or dense tables, use Accurate and rely on needsReview plus the validation checks.

Does it OCR scanned PDFs? Yes. Pages without a usable text layer are rendered to images and read by a vision model. Use Image resolution to trade detail against cost.

Does it work with phone photos of receipts? Yes. JPG, PNG, WebP, TIFF and GIF photos are read by the vision model. Sharp, well-lit photos work best. Blurry ones get lower confidence and needsReview.

Does it work with bank statements from any bank? It uses a general AI model rather than per-bank templates, so it works with most layouts. Opening balance + transactions is reconciled against the closing balance, and running balances are checked. Statements that don't add up are flagged with needsReview so you can check them.

Long documents? Up to 200 pages (default 30). Long text is split into chunks and the results are merged. Tables are concatenated.

Can it make up values? It is told to return null for anything not on the page, and every value is compared against the document text. Values that cannot be found are listed in ungroundedFields and their confidence is capped.

Dates like 03/04/2026? Use Date format (DMY / MDY) or let the model decide from context such as the country or currency. Ambiguous dates are flagged.

Excel, PowerPoint or HEIC files? Not yet. XLSX, PPTX, old .doc and HEIC files are rejected as free UNSUPPORTED_FORMAT rows. Convert them to PDF or JPG first.

Passports or ID cards? No. Government ID documents are deliberately refused and not charged.

Can I call it from an AI agent or MCP client? Yes. Like every Apify Actor, it can be called through Apify's MCP server or the API.

Is my data stored? Only in your own Apify storages. Documents are sent to the AI model (via OpenRouter) for extraction. Don't process documents you are not allowed to share with an AI provider.

Limitations

• Handwriting and very low-resolution photos may be misread. Check needsReview.
• XLSX, PPTX, old .doc and HEIC files are not supported in v1. Convert them to PDF or JPG first.
• Password-protected PDFs cannot be opened.
• Government ID documents are deliberately not supported.
• On the Apify FREE plan, the OpenRouter proxy costs 10× the provider price.

Support

Found a document it gets wrong? Open an issue on the Actor's Issues tab with a sample (remove personal data first). We fix extraction problems quickly.