DOCUMENT AI WORKFLOW AUTOMATION

The paper said ten. Nine arrived.

Document-triggered work, completed in the systems you already run. On this page one delivery note is classified, read, checked against the order, stopped, handed to the person who counted and written once.

Scope a document AI workflow See a delivery note read →

THE READ 56 lines · 6 carried

Three pages came stapled together. The reader finds the nine.

A delivery note of two pages, and a letter that does not belong to it. The letter is held for a person. Page two is read in 56 lines: the printed ten, the strike through it and the handwritten nine beside it are three of them.

pages in the staple

  1. unknown · held for a person
  2. delivery note · 1 of 2
  3. delivery note · 2 of 2
  • class delivery note · 2 of 2
  • supplier Nordvik Emballasje AS
  • note PS-2026-08871
  • item IBC-1000
  • ordered 10
  • received 9

Reading 56 lines on page two. The handwritten nine is read at 0.93 confidence, the receiver's name and the stamp at 0.99 and 0.97, and the signature at 0.97 with one letter wrong. A value read below the threshold agreed for its field goes to a person instead of being written. The organisation number on page one passes its checksum.

open-source OCR model about 1 s per page on one CPU

A recorded read by an open-source OCR model, on the flat page, on one CPU, on 18 September 2026. The boxes are drawn on the photograph where the model found each line. The check, the owner and the record below run live in your browser. An example note. Nordvik Emballasje AS and Fjellheim Varehus AS are invented companies.

THE STOP 10 ordered · 9 received

The value was read correctly. The delivery still stops.

Ten in the order, ten in the electronic invoice, nine in handwriting on the note. The comparison runs in your browser on the value the model read. Move the count and the verdict moves with it.

ordered
10 IBC-1000
invoiced
10 IBC-1000
received
9 IBC-1000

Stopped, one short

organisation number
312 678 454 checksum passes

the rule

Received must equal ordered before the receipt is written. A checksum that fails stops the note before the count is compared.

the bundled example, until you run the check
  1. RECEIVED 14:40
  2. READ
  3. CHECKED
  4. STOPPED
  5. RECORD

Runs in your browser. No customer data.

THE OWNER one exception · one owner

The exception goes to the person who counted. One answer releases it.

The stop reaches the goods receipt owner with the evidence attached: the handwriting, cropped from the page, and the rule the count broke. Their answer runs the check again and releases the record.

What the owner receives

the crop
the handwritten nine, cut from the model's box on page two
the rule
received 9 differs from ordered 10
the route
goods receipt owner · Fjellheim Varehus · port 3

the answer

Waiting for the owner.

Runs in your browser. No customer data.

THE RECORD 6 fields · written once

Nine received, one outstanding. Written once, even when the system was down.

The receipt is written into the system of record you name, with the page and the line each value came from. Switch the system of record off: the write waits, retries and is written once when the system is back.

Goods receipt WRITTEN · 1 OF 1
  • fields read 6
  • rules passed 1 of 2
  • decided by a person 1
  • written 1 of 1
order
ORD-45120 page 1, head
supplier
Nordvik Emballasje AS page 1, head
note
PS-2026-08871 page 2, head
item
IBC-1000 · IBC-container 1000 l page 2, line 4
received
9 of 10 page 2, line 4, handwritten
decided by
goods receipt owner · 17 September 2026, 14:40 page 2, signature and stamp

The receipt lands where the rest of your records are. A sample record from the example note. Move the count above and the record follows.

FOUR KINDS OF WORK 4 kinds · 1 workflow

Four kinds of work still arrive on paper. One workflow.

In Norway the business invoice becomes structured data by law. The note, the claim, the form and the contract still arrive as pages. Each is read, checked against what it must agree with, and handed to an owner when the two disagree.

Kind 01

Goods and invoices

what arrives
delivery notes, packing lists and foreign supplier invoices
must agree with
the order and the electronic invoice
who owns the exception
the goods receipt owner

Kind 02

Claims and cases

what arrives
claim forms, photographs and reports
must agree with
the policy and the case file
who owns the exception
the claims handler

Kind 03

Onboarding

what arrives
identity documents, forms and signed consents
must agree with
the register and the checklist
who owns the exception
the onboarding owner

Kind 04

Contracts

what arrives
signed contracts and their amendments
must agree with
the agreed terms
who owns the exception
the contract owner

Your document kinds are named at scoping. The workflow is the same one shown here.

BEFORE THE BUILD 3 rows · agreed first

What counts as accepted is written down first. The pilot is judged against it.

Three things are agreed before the first production document is read.

  1. 01 The classes and the schema Which documents arrive, what each one must yield, and what an unknown is held as.
  2. 02 The evaluation set Your documents, including the ones that must stop.
  3. 03 The rules, the thresholds and the owners Which rule stops a document and who owns each exception, with thresholds agreed in the statement of work.

THE BRIEF

Bring the documents and the decision they should trigger. Three lines are enough.

  1. What arrives The kinds of document, and how they come in.
  2. What it must agree with The order, the case, the register: the system of record.
  3. Who owns the exception The person a stopped document goes to.

The rules you sign start as these three lines.

Looking for work on AI projects? This form is for companies. Apply as a freelancer

The kinds of document, and how they come in. The order, the case, the register: the system of record. The person a stopped document goes to.

We use the information you provide to assess, respond to and manage your enquiry in accordance with our Privacy Policy.

A named project lead reads it and replies inside one EU business day with a first read on the classes, the rules and the evaluation set.

BEFORE YOU SCOPE

What operations, finance and IT ask first.

What does the workflow cover?

Classification, extraction, validation, review, routing and the system update, built as one workflow. On this page they are the read, the stop, the owner and the record.

Which documents can it read?

Printed pages, handwritten entries on them, scans and photographs of pages, and files that already carry text. Each class is named before the build, with what it must yield and what an unknown is held as.

What happens to a document the system does not recognise?

It is held as unknown for a person, together with the pages it came with. Extraction starts when that person names it.

Where does the extracted value go?

Into the system of record you name, once, after the rules have passed or a person has decided. The record carries the page and the line each value came from.

Who decides an exception?

The owner named for that rule. The exception reaches them with the crop from the page and the rule it broke, and their answer releases or corrects the record.

Does the invoice still need reading?

In Norway the invoice between bookkeeping-liable businesses becomes structured data by law. Foreign invoices, delivery notes, claims, forms and contracts still arrive as pages, and that is where the reading is.

Where does it run?

Inside the boundary you agree before release: your cloud, a dedicated EEA tenant, a hybrid split or your own hardware. The deployment page names the four and what each one keeps inside.

Paper will keep arriving. We build the workflow that reads each page, stops the wrong ones and writes the right ones once.

AI Implementation covers discovery, architecture and the deployment itself.

Scope a document AI workflow AI Implementation →