Client project · Automation

The line nobody has to find by hand anymore.

Our client orders from suppliers by email, and the confirmations come back by email. Someone has to check that each supplier understood what was ordered. We built the software that does the comparing. Version 1 is finished and waiting for approval from the client’s IT.

Discuss a project 
Input
Mailboxwhere the confirmations arrive
Documents
PDF & Excellaid out differently by every supplier
Price
Fixedagreed before we started
Status
Version 1finished, waiting for IT approval

Before

It almost always matches, and sometimes it doesn’t.

Orders go to the suppliers by email. Days later the confirmation comes back, also by email. A clerk finds it in the mailbox, opens the original order next to it and compares them line by line: dimensions, quantities, versions. Did the supplier understand what was ordered?

Every supplier writes its confirmations differently, in a different order, with its own abbreviations and with fields left out. You need to know the trade to compare them, and it is still dull work that makes you go blind over time. A measurement that is a few millimetres off is easy to miss in the third document of the day. If nobody catches it, the wrong part gets made and has to be made again, with the delay on top. That easily costs several thousand euros.

Before
  1. Order by email
  2. days later
  3. Confirmation from the supplier
  4. search
  5. Mailbox search it
  6. by hand
  7. Comparison line by line
  8. forward
  9. Release to production

What we built

No new system. The software reads the mailbox.

Nobody had to change how they work. Orders still go out by email and confirmations still come back by email. In between, our software reads the mailbox.

  1. 01

    Sort the inbox

    Incoming emails are recognised and passed to the right person. Order confirmations go into the check.

  2. 02

    Find the order

    For each confirmation, the software finds the matching order in the order folder, whether it is a PDF or an Excel file.

  3. 03

    Read both documents

    Items, dimensions, quantities and versions are read from both, even when the supplier formats them differently, abbreviates or leaves fields out.

  4. 04

    Compare and report

    Line by line. The clerk gets an email with a short PDF and decides what happens next.

  5. 05

    New suppliers without us

    There are too many document layouts for a fixed lookup table. The software is built to learn new layouts itself, so a new supplier doesn’t need us to step in.

Input
The existing email mailbox
Documents
Orders as PDF or Excel, confirmations as PDF, laid out differently by each supplier
Output
Email to the clerk with the comparison as a PDF
Operation
Runs without us and learns new document layouts itself

Under the hood

Years of documents. No training data.

The old documents were never meant for machine learning. Nothing said which supplier, which kind of document, which layout version and which fields each one contained. The software had to work that out from the documents themselves first.

It groups similar documents by layout and content, works out the supplier and the layout version in several passes, derives a shared set of fields from that and then reads each group in the way that suits it. Fixed rules where the layout never changes, statistics and models where it does. When a new layout turns up, the software adds it to what it already knows.

What happens before one result email

  1. 01
    Old documents

    Years of orders and confirmations, unlabelled, in changing formats.

  2. 02
    Document analysis

    The layout and content of each document are captured before anything is sorted.

  3. 03
    Grouping

    Similar documents are grouped by layout and content. The suppliers and their versions come out of the data, not out of a list.

  4. 04
    Supplier and version

    Several passes work out who sent a document and which version of their layout it is.

  5. 05
    Learning the fields

    The groups give a shared set of fields, which grows when new layouts appear.

  6. 06
    Reading each group

    Rules where the layout is stable, models where it isn’t. Items, dimensions, quantities, versions.

  7. 07
    Fixed checks

    Everything that was read is checked against fixed rules: is it plausible, is it complete, are the units right.

  8. 08
    Matching

    The confirmation finds its order, even when the reference is only in the email thread.

  9. 09
    Comparison

    Line by line, with the value from each document side by side.

  10. 10
    Decision

    The clerk decides. The software only prepares it.

Why the AI doesn’t decide

The client asked whether you couldn’t just give the whole thing to an AI. So we tested a current model and slipped it wrong values on purpose, like a dimension taken from a completely different field. In 22 of 135 cases it still said everything matched, and once it even explained why.

So in our software the AI is only allowed to show where a value is in the document. Whether it matches, the program calculates itself. The catch is that more fields stay open for a person. The clerk still checks, but only those fields, not every line.

Why the mailbox stays

The cleanest solution would have been to change how everyone communicates, with every supplier replying to the clerk who ordered and always using the same reference. That would be a project for the whole organisation, not a software project. So the shared mailbox stays, and the software works out what kind of email it is and who is responsible.

The features and the workflow are real. We don’t name the client, the suppliers or real values.

More about AI & data 
Confirmation 2026-0417 checked: 1 difference 1 differs
ItemOrderConfirmationStatus
Item no.4711-B4711-Bmatches
Quantity120 pcs100 pcsdiffers
Unit price€18.40€18.40matches
Deliveryweek 46week 46matches
What the result email to the clerk looks like. Sample values.

After

One email instead of two documents, with the problem at the top.

The clerk opens one email instead of two documents. Lines that match are ticked off. Lines that don’t are at the top, with the value from the order and the value from the confirmation side by side.

With JMK
  1. Mailbox is read
  2. automatic
  3. Matching order found
  4. structured
  5. Comparison line by line
  6. minutes
  7. Report email with PDF
  8. checked
  9. Decision by the clerk

Where it stands

What we can show today, and what comes later.

Price
A fixed price for the agreed scope, set before we started.
Status
Version 1 is finished and waiting for approval from the client’s IT.
Results
Figures from day-to-day use and a word from the client will go here once we have them.
Client
We don’t name the client yet. We are still clarifying what we are allowed to publish.

Elsewhere

The same problem, with other documents.

Two documents that should match, and a person who checks them by hand. Almost every business has that somewhere.

The first conversation is free. Describe the workflow and we’ll say if it’s worth it.

  • Delivery notes against orders
  • Incoming invoices against order confirmations
  • Supplier quotes against your requests
  • Field service reports against work orders

Which documentsdo you compare by hand?

Discuss a project [email protected]