Data extraction API: how it works, how to choose one and how to use it

A data extraction API turns a PDF, a scanned form or a web page into structured data. The data usually comes as JSON, a format other software can read. A document extraction API is the kind built for business documents. It first works out the document type. It then returns named fields and tables instead of raw text. Each value comes with a confidence score, which shows how sure the API is, and its place on the page.

First we compare four document extraction APIs, and what each one means if you build it into your own software. Then we cover the four kinds of API and what happens when your system sends a document to one. Last comes how to go live.

Document extraction APIs compared

Docsumo is our product. The other three are the document AI services of the big cloud providers. We took details from each vendor's own website in September 2026. The links are in the sources at the end.

For more OCR-only options, see the best OCR APIs.

Embedding an extraction API in your own product

Software companies that add document extraction to their product need more than JSON. Their users also need a way to fix the values the API is unsure about. Docsumo's document extraction API returns checked fields, and its review screen can run inside your product without Docsumo branding. Your product can create a Docsumo account for each of your customers through a signup API. Uploads can be up to 500 pages per file. With the cloud APIs, your team builds the checks and the review screen itself.

Types of data extraction API

Which kind you need depends on where the data lives.

• OCR APIs

  Return the text on a page image, with each word's position and a confidence score. You write the rules that turn text into fields. See the best OCR APIs.

• Document extraction APIs

  Return named fields and tables for a document type. Examples are an invoice's number and line items, the boxes on a tax or insurance form, and the tables in a report or research paper. Smart, intelligent and form data extraction APIs are all this kind, and it's the one for business documents.

• Web data extraction APIs

  Fetch web pages and return their content as clean data. Examples are an article's main text without menus and ads, or a product's name, price and reviews. Content extraction and scraping APIs usually mean this kind. See data extraction vs data scraping.

• Database and application APIs

  Return records a system already holds, from a database or a software product's REST API. Nothing is read, because the data is already structured.

How a document extraction API call works

You send a file, the service reads and checks it, and the result comes back as JSON. It often comes by webhook, because a long document can take a while.

Field names differ by vendor, but the response usually carries the same things. It holds the document type, each value, a confidence score and the page the value came from, often with its coordinates.

{
  "id": "doc_7Q2K9",
  "type": "invoice",
  "fields": {
    "invoice_number": { "value": "INV-20418", "confidence": 0.99, "page": 1 },
    "invoice_date": { "value": "2026-09-14", "confidence": 0.98, "page": 1 },
    "po_number": { "value": "PO-7731", "confidence": 0.64, "page": 1 },
    "total": { "value": 1260.00, "confidence": 0.99, "page": 1 }
  },
  "line_items": [
    { "description": "Work table", "qty": 2, "unit_price": 385.00, "amount": 770.00,
      "confidence": 0.97, "page": 1 },
    { "description": "Sheet pan rack", "qty": 2, "unit_price": 245.00, "amount": 490.00,
      "confidence": 0.96, "page": 1 }
  ]
}

The PO number at 0.64 is the value to send to a person. What you do with low-confidence values decides how much you can trust the rest.

The best data extraction APIs for documents

Docsumo is our product, so we've put it first and said plainly what it doesn't do. The other three are the document AI services of the big cloud providers.

• 99%99%field-level accuracy across 250+ document types
• 95%+95%+of documents processed straight through, without manual review
• 3,000+3,000+hours a month saved at Arbor

How to choose a data extraction API

• Models for your documentsCheck it has pre-trained models for the documents you actually receive, and what happens with the ones it doesn't know.
• Confidence and location on every valueWithout them you can't route doubtful values to a person or show a reviewer where a number came from.
• A review stepDecide who fixes low-confidence values. It can be a review queue in the product, or screens your team builds and maintains.
• WebhooksResults are pushed to you when a document is done, so you don't have to keep asking.
• SecurityCertifications such as SOC 2 Type 2, how long files are kept, and whether your documents train shared models.
• SDKs and a test environmentReady-made code (client libraries) for the programming language your team uses, and a place to try changes before they go live.
• Pricing modelPer page, per document or per plan. Compare the total cost per document, including review time.

Our take. Judge an extraction API by the values a person had to fix on your own documents, not by a vendor's accuracy figure. If you build the API into your product, your users will judge it the same way. So we'd test each API on a month of your real files and count the fixes before choosing.

Price per page is only part of the cost. Put in your own volumes.

How to use a data extraction API in 5 steps

Write down what you need first. List the document types, the fields, your monthly volume and how fast results must come back. Then follow these steps.

1. Get an API keySign up for a trial or a cloud account, create a key, and read the API docs for file size limits, rate limits and the current version. Keep the key on your server or in a secrets manager, never in browser code. Most APIs take it, or an OAuth 2.0 bearer token, in a request header over HTTPS.
2. Send a test documentPost one real file with cURL or Postman before writing code. Send the file or a link to it, and the document type if you know it. Add your own reference ID, so the result maps back to your record.
3. Take results by webhookExtraction can take seconds or minutes, so many document APIs work asynchronously. They call a webhook URL you give them when a document is done, or you check a status endpoint until it is. Reply to the webhook at once with a 2xx code, and process the payload in the background.
4. Map the JSON to your systemMap the vendor's field names to yours, and put dates and amounts in one format. Keep the page and position of each value, so anyone can trace it to the source. See schema mapping.
5. Handle errors and low confidenceRetry 429 (too many requests) and 5xx responses with growing waits, and honor any Retry-After header. Fix other 4xx errors, such as a bad file or an expired key, instead of retrying them. Send low-confidence values to a person, not into your system.

In the first weeks, compare the number of documents sent with the number of results received. Also check a sample of values against the documents themselves.

“We love Docsumo for its ease of integration and the flexibility it provides with API and webhook callbacks.”

Howard Leiner, CTO, Arbor

Building blocks or a platform

Choose the kind of API by where your data comes from. OCR APIs return text, web data extraction APIs read web pages, and document extraction APIs return named, checked fields from business documents. The big cloud services give developers basic parts to build their own tool. An IDP platform such as Docsumo adds the checks and the review step, so your team only handles the values the platform is unsure about.

Book a demo with a few of your own documents, or start a free trial and call the API on them.

Frequently asked questions