Skip to content
Document extraction Free consultation

Document data extraction · IDP

Document data extraction that knows when to ask a person.

Invoices, receipts, forms and scans become clean fields and tables in your schema. Every value carries a confidence score. Anything under your threshold, like a smudged handwritten number, goes to a person before it is written anywhere.

Book a free consultation Got sample documents?

Free consultation. If it fits, the next step is a systems audit from $2,500, credited against the build.

No software license to buy. You own what we build.

KWESTRA · RS-01REC 000
  1. 01 Capture Inbox tray and pusher
  2. 02 Locate the fields Reading arm and scan bar
  3. 03 Read and score Lens and confidence dial
  4. 04 Human check Shutter and lever
  5. 05 Structured data out Record board and output ports

Example run on synthetic documents. Nothing is uploaded or stored.

  • Every field gets its own confidence score
  • 0.85 starting threshold, tuned on your documents
  • Your schema JSON, CSV or your ERP's fields
  • Yours code, prompts and test set

What it reads

Invoices, receipts, forms and tables, printed or handwritten.

If a person can read it, the system can usually read it too: native PDFs, phone photos, faxes and multi-page scans. Six of the documents we see most.

  • Invoices and credit notes

    • vendor
    • invoice number
    • line items

  • Receipts and expense slips

    • merchant
    • date
    • total and tax

  • Purchase orders and packing slips

    • PO number
    • SKUs
    • quantities

  • Bills of lading and delivery receipts

    • shipper
    • consignee
    • weights

  • Application and intake forms

    • names
    • checkboxes
    • handwriting

  • Bank and card statements

    • transactions table
    • dates
    • balances

Also: contracts and certificates (parties, dates, limits), tax forms, inspection reports. Ask about yours.

Intelligent document processing, step by step

From a scanned PDF to fields in your schema.

OCR turns pixels into text. A language model reads that text in context: which number is the total, which rows belong to the table, which date is the delivery date. Validation rules then check the math.

Example
Delivery receipt · Harbor Line Supply (example)

record.json

{
  "doc_type": "delivery_receipt",
  "vendor":   { "value": "Harbor Line Supply", "confidence": 0.99, "page": 1 },
  "doc_no":   { "value": "DR-0417", "confidence": 0.98, "page": 1 },
  "date":     { "value": "2026-10-14", "confidence": 0.97, "page": 1 },
  "lines": [
    { "sku": "HL-2210", "description": "Pallet wrap 18in",
      "qty": 12, "unit_price": 38.50, "confidence": 0.96 },
    { "sku": "HL-3305", "description": "Strapping kit",
      "qty": 6,  "unit_price": 96.10, "confidence": 0.95 },
    { "sku": "HL-4120", "description": "Corner boards",
      "qty": 4,  "unit_price": 61.50, "confidence": 0.97 }
  ],
  "total":    { "value": 1284.60, "currency": "USD", "confidence": 0.99 },
  "received_qty_line_3": {
    "value": 4, "confidence": 0.58, "status": "needs_review"
  }
}
  1. 1
    OCR and layout Text and positions from any scan or PDF.
  2. 2
    Read in context Fields found by meaning, not by template.
  3. 3
    Validate Totals add up, dates are real, IDs exist in your records.

Example document and values are synthetic. The same receipt runs through the machine above.

Extraction accuracy

Confidence scores decide what a person checks.

Every field gets a score. Above your threshold it goes straight through. Below it, a reviewer sees the crop from the page next to the value read, and approves or fixes it in seconds.

At 0.85: 9 go straight through · 3 go to a person

  • Handwritten note 0.58
  • Ship-to 0.76
  • Line 3 qty 0.82
  • Line 2 qty 0.88
  • Tax 0.91
  • PO number 0.93
  • Line 1 qty 0.95
  • Subtotal 0.96
  • Date 0.97
  • Invoice no 0.98
  • Vendor 0.99
  • Total 0.99

Threshold 0.85

Example batch, synthetic values.

Example

Received qty, line 3

Read: 4Confidence 0.58

Logged: reviewer, time, before and after.

  1. We measure accuracy per field, on your documents, not on a vendor benchmark.
  2. Rules catch what scores cannot: totals that do not add up, future dates, unknown vendors.
  3. Every correction becomes a test case for the next release.

Where the data goes

Structured data out: JSON, CSV or your ERP's own fields.

You pick the shape. We map every field to your schema, keep the page and position it came from, and deliver it by API, file drop or a direct write to your system.

{
  "doc_type": "delivery_receipt",
  "vendor":   { "value": "Harbor Line Supply", "confidence": 0.99 },
  "doc_no":   { "value": "DR-0417", "confidence": 0.98 },
  "date":     { "value": "2026-10-14", "confidence": 0.97 },
  "lines": [
    { "sku": "HL-2210", "qty": 12,
      "unit_price": 38.50 },
    "... 2 more rows"
  ],
  "total":    { "value": 1284.60, "currency": "USD" }
}

Reading the document is one step. Next step: automate the whole workflow, from the inbox to the posted entry.

Document processing automation

Security and data handling

Your documents stay inside a boundary you set.

Extraction runs in your cloud account or in an environment dedicated to you. Models are called by API under terms that exclude training on your data. How long files are kept is your setting.

Your environment
  1. Inbox or upload
  2. OCR
  3. Model call
  4. Review queue
  5. Your system
Model provider (API) No training on your data
  1. Encrypted in transit and at rest.
  2. Optional redaction of names and account numbers before any model call.
  3. A written data agreement before we touch the first document.
  4. Access by role. Every view and every correction is logged.
  5. Source files kept for the number of days you choose, then deleted.

Health or financial records? Tell us at the start and we scope the hosting and agreements around HIPAA, GLBA or your own policy.

Free, before anything else

Got sample documents? We will tell you the extraction rate.

Got a sample pack ready? Tick the box and we will send a secure way to share it after we reply. An engineer then runs ten real documents, messy ones included, through the extraction we would build and answers within two business days: the field-level rate, what would go to a person, and why.

  1. 1Tell us what you read and tick the box
  2. 2We reply with a secure way to share the pack
  3. 3Your rate and what goes to a person, in two business days

Nothing is uploaded here. Share files only through the secure link we send, stored encrypted, used only for this analysis and deleted on request. Redact names and account numbers if you prefer; the layout is what matters.

Extraction rate = fields read correctly at or above the threshold, out of all the fields you need.

A person replies within one business day. Not a sequence.

The next step

Then an audit, priced to your scope.

When the samples look right, the audit maps your document flow, measures the extraction rate on a real week of volume and ends in a written quote.

Work order · SA-01 Priced to your scope
Processes in scopeEach one is mapped on its own.
1
Roughly. It moves the price a little.
5 people
Systems it touchesERP, inbox, portal, spreadsheet, bank site.
2
Sites that run it differentlyBranches, entities or teams with their own way.
1
Regulated dataPick any that apply.

Count every system the documents come from or land in.

Book now

$2,500

Regular price $3,125

Launch offer: 20% off when you book by Dec 31, 2026. Your quote is held for 48 hours.

Work order
SA-01 Starter
Duration
2 weeks
Credited in full
$2,500
  1. SA-01 Starter $2,500
  2. SA-02 Team $3,000 to $6,000
  3. SA-03 Department $6,500 to $12,000
  4. SA-04 Division $13,000 to $20,000

What you hold at the readout

  • Document map Every document type, its volume and where its data has to land.
  • Measured extraction rate Field by field, on your own documents, at the threshold you choose.
  • Exception register What goes to a person, how often, and the rule behind it.
  • Scoped build and quote In, out, timeline and price in writing. If extraction is not worth it yet, that goes in writing too.

The work order lists what you entered. If the work turns out bigger, we tell you at kickoff and you choose: we stick to what you entered, or we quote the difference before doing it.

Skip the call

Start the audit now.

Pay the audit price and we start. An engineer reads your brief within one business day.

  1. 01 Pay your audit price By card or US bank account, through Stripe.
  2. 02 An engineer reads your brief You hear from them within one business day.
  3. 03 Kickoff within 2 business days Then the weeks on your work order to the process map and a written quote.

Credited in full: $2,500 against a build signed within 90 days of the readout.

SA-01 2 weeks $2,500

Regular price $3,125. Launch offer: 20% off when you book by Dec 31, 2026. Your quote is held for 48 hours.

Secure payment by Stripe, card or US bank account. Bank payments take 3 to 4 business days to clear. By paying you agree to our terms.

Questions

Document extraction questions buyers ask first.

Something else? Ask in the consultation form below.

What is the difference between OCR and intelligent document processing?

OCR turns an image into text. Intelligent document processing goes further: it finds the fields that matter, reads tables row by row, scores its confidence in each value, checks the values against rules, and hands you structured data instead of a wall of text.

How accurate is AI document data extraction?

It depends on your documents, which is why we measure it on them rather than quote a benchmark. We report accuracy per field at the threshold you choose. The number that matters for cost is how many fields need a person at that threshold; the sample analysis and the audit give you both.

Can it read handwritten forms?

Usually, yes. Clear block capitals read well; rushed numbers and struck-over digits read less reliably. Those come back with low confidence scores, so they go to a reviewer instead of into your system. The sample analysis shows how your handwriting mix performs.

Can you extract tables from PDFs and scans?

Yes. Line items, transactions and multi-page tables come out row by row with their columns, including tables that continue across pages. Each row gets its own confidence, and totals are checked against the sum of the lines.

Do we need a template for every vendor layout?

No. The model reads by meaning, so a new invoice layout from a new vendor does not need a new template. We still test on your real layouts before go-live and keep a test set so a model update cannot quietly change results.

Where do our documents go, and are they used to train models?

They stay in your environment or one dedicated to you. Models are called by API under terms that exclude training on your data, and retention is your setting. We sign a written data agreement before the first document.

What does document extraction cost to run?

The model cost per document is small: at October 2026 list prices, extracting a typical two-page invoice costs between a tenth of a cent and a few cents (see our AI automation cost guide). The real cost is the build and the review time, which the audit prices in writing.

Is this a software subscription?

No. We build the extraction for your documents and your systems, and you own the code, prompts and test set. We can run it for you afterwards if you prefer, as a service you choose.

Free consultation

Tell us which documents you need turned into data.

A sentence or two is enough: the document types, a rough volume and where the data should end up. A senior engineer replies within one business day.

What happens next

Today
Your brief goes straight to a senior engineer, not a sales team.
Within 1 business day
That engineer replies to you personally.
Then
You get a written answer on fit, scope and price. No email sequence.

A person replies within one business day. Not a sequence.