Skip to content
DocWise

Document AI · classify · extract · review

Document extraction that shows its work.

Teach DocWise the documents you receive from three samples. Every value comes back typed, checked and pinned to the text it was read from — and when DocWise isn’t sure, a person sees it before your systems do.

Start freeAPI quickstart

Free plan: 100 pages a month. No card required.

Try it: an invoice, its extracted fields and how it is routed

northwind-inv-20931.pdf · 1 page

  1. Scanned
  2. Text read
  3. Classified · Invoice
  4. 12 fields extracted
  5. 2 checks run
  6. Routed
NEEDS_REVIEW

Illustrative document. Hover or tap a field to see where it was read.

0.85

Set per document type. Product default 0.85.

Sent to review. PO number is below 0.85. A reviewer sees the page beside the fields, lowest confidence first.

low_confidencedocument.needs_review

01The problem

Every sender has their own layout. Your systems need one shape.

Retyping
Accurate when people are fresh. Cost grows with every document, and errors land in payments and filings.
A template per layout
Works until a supplier redesigns their invoice. Every new sender is another template to build and maintain.
Raw OCR or an LLM call
You get text or a best guess. No confidence, no evidence, no checks — so nothing tells you which numbers are wrong.
DocWise
Learns each document type from your samples, checks every value it returns, and tells you exactly which ones a person should look at.

02How it works

From a few samples to a pipeline you can trust.

  1. Step 1

    Teach it

    Upload 3–20 samples, or clone a template. DocWise reads each one and suggests the fields: names, types, choice lists and table columns, with example values. A field is required when it appears in every sample.

    PDF · PNG · JPEG · TIFF

  2. Step 2

    Test it

    Run the draft on your samples, confirm or fix each value, and DocWise scores itself against what you confirmed (target 90%). It also checks the new type isn't confused with your others. Each publish is a version you can roll back.

    Test bench · accuracy · versions

  3. Step 3

    Run it

    Every new document is classified, extracted, normalized and checked. Clean results go to your systems by API, webhook or export. The rest go to a review queue.

    Classify · extract · check · route

03Every document

Seven steps. Five of them can stop and ask.

DocWise doesn’t force a match or fill a gap with a guess. When a step can’t vouch for its result, the document waits for a person.

  1. 01Intake

    File type checked from its bytes, not its name. Malware scan. Page count against your plan.

    Infected or unsupported files are blocked.

  2. 02Read

    The text layer of digital PDFs is read directly; scans and photos go through OCR.

  3. 03Classify

    Matched against the document types your workspace has enabled.

    Unknown, ambiguous and mixed documents go to review. A mixed PDF comes with suggested page splits.

  4. 04Extract

    Output is checked against your type's JSON Schema and retried once if it doesn't match. Long documents are read in overlapping 10-page windows.

  5. 05Check

    Dates, amounts and codes normalized. Formats, ranges and cross-field checks such as line items summing to the subtotal.

    A failed check sends the document to review.

  6. 06Evidence

    Each value's quoted text is located on the page, so a reviewer can see exactly where it came from.

    A quote that can't be found on the page is flagged.

  7. 07Route

    If every value clears your threshold (default 0.85) and every check passes, the result is delivered.

    Anything else waits for a person.

04Review

People only where it matters.

Reviewers see the page beside the fields, lowest confidence first, with the evidence highlighted. They accept, correct, change the type, split a mixed PDF into documents or mark it out of scope. Corrections are stored as new versions, never overwrites, and a review.completed webhook tells your systems when it’s done.

Why a document waits for review

  • Low confidence
  • Check failed
  • Evidence not found
  • Unknown type
  • Ambiguous type
  • Mixed document

Thresholds are set per document type. Raise one while you build trust in a new type, lower it once the review queue shows DocWise is getting it right.

05Integrate

Results arrive where your work happens.

REST API
Upload, poll and fetch results with scoped, expiring API keys. Retries are safe with an idempotency key.
Webhooks
Signed with HMAC-SHA256, retried five times with backoff, and replayable from your workspace.
Exports
CSV, JSON or Excel per document type, with line items as one row each.
Read the quickstart
A signed delivery to your endpoint (abridged)
POST /webhooks/docwise
X-Webhook-Id: evt_…
X-Webhook-Timestamp: 1790812800
X-Webhook-Signature: v1=5b2e…   # HMAC-SHA256

{
  "event": "document.completed",
  "data": {
    "documentId": "cm…",
    "result": {
      "document_type": { "name": "Invoice" },
      "data": { "invoice_number": "INV-20931", "total_due": 1342.29, … }
    }
  }
}

06Templates

Or start from a template.

Ready-made field lists, classifier hints and checks. Clone one, add your own samples, test and publish.

  • Invoice
  • Receipt
  • Purchase order
  • Bank statement
  • Utility bill
  • Bill of lading
  • Form W-2
  • Form W-9
  • Form 1099-MISC
  • US driver license
  • Passport
  • Bank check
  • ACORD 25

07Your data

Careful with what you send us.

  • Customer documents are not used to train AI models.
  • Every upload is scanned for malware before processing.
  • Source files deleted after 30 days, results after a year — both configurable.
  • Staff access to a workspace needs a time-limited, recorded grant.
  • Roles for owners, builders, reviewers, developers and viewers; two-factor sign-in.
How we handle data

08Pricing

Priced by the page.

Every plan includes a monthly page allowance, whichever document types the pages belong to. The free plan includes 100 pages a month. Paid plans start at $49 a month for 1,000 pages.

See plans

09Questions

What files can DocWise read?

PDF, PNG, JPEG and TIFF. Digital PDFs are read from their text layer; scans and photos use OCR. File size and pages per document depend on your plan.

How many samples does a new document type need?

At least 3 to publish, up to 20. Include the variations you actually receive — different senders, scanned and digital copies.

What happens when a document matches none of my types?

It is marked unknown and goes to review instead of being forced into the closest type. Ambiguous and mixed documents are handled the same way.

Which AI models does it use?

Anthropic and OpenAI models, chosen per task (classification, extraction, field suggestions) with automatic fallback to another model if one is unavailable. DocWise manages the models; you don't need your own AI account.

Do you train AI models on my documents?

No. Customer documents and results are not used to train AI models. Source files are deleted after 30 days and results after a year by default, and you can change both.

Do I need a developer?

Not to start. You can upload in the browser, review results and export CSV, JSON or Excel. A developer can connect the REST API and webhooks when you want results to flow into your systems automatically.

Is there an SDK?

Not yet. The REST API is documented with an OpenAPI reference, and the quickstart has curl examples and webhook verification in Node.js and Python.

Bring three samples.

Create a workspace, upload a few documents you receive and see the fields DocWise suggests. Processing high volumes? Talk to us.

Start freeAPI quickstart