Skip to content
Not Your Job

SOLUTION / 001Documents → DataWENT LIVE 02·08·2026A MANUFACTURER

It reads everything. It guesses nothing.

Ten order layouts, a hundred statements with no template, three years of mail: a manufacturer received them in a mailbox and in shared folders. One system reads them, ties them to the same catalogue, the same contracts, the same due dates, and says what it could not work out.

BUSINESS VIEW

  1. 01The day-to-day

    Open the catalogue and the contract for every order, read every email, copy every statement, file every invoice.

  2. 02The problem

    Every document means looking something up somewhere else.

  3. 03The solution

    Reads, matches by rule first and by resemblance last, checks against the contract, and says what it could not work out.

  4. 04Not your job anymore

    The reading, the matching, the copying, the classifying, the renaming.

  5. 05The person keeps

    Approving every order, settling every doubtful case, reviewing every price change.

SYSTEM VIEW

  1. BEFORE
    01 · INPUTMAILBOX · FOLDERSemail, PDF, Excel; read-only
  2. SOLUTION
    02 · AIREAD
    • amount
    • date
    • number
    • sender
    a model reads what has no structure, once per document
  3. 03 · RULESMATCHING
    • EXACT_REF
    • EXACT_EAN
    • MAPPING
    • ALIAS
    • FUZZY
    the cascade stops at the first stage that succeeds; resemblance last
  4. 04 · RULESCERTAINTY
    • GREEN · pre-approved
    • AMBER · one click
    • RED · nothing
    a degree, not a result, never “roughly”
  5. AFTER
    05 · HUMANVALIDATIONapproves, corrects, decides; the only moment the system learns
  6. 06 · OUTPUTDATABASE · WORKBOOKSits own database, its Excel workbooks; never the ERP

01What disappears

02Data flow

The mailbox, processed step by step.

Nine named steps, in the order the work is done. Each one has a state: waiting, running, done, failed.

  1. 01 · DÉPILAGE

    Check the mailbox. New messages are collected; the report says how many.

    DONE
  2. 02 · IMPORT

    Import the messages. Messages and attachments enter the database, once each.

    DONE
  3. 03 · CLASSEMENT

    Classify new messages. Invoice, reminder, statement, other: a rule on words and domains first, a model second.

    DONE
  4. 04 · EXTRACTION

    Read new documents. Amounts, dates, numbers, sender. A document already read is not read again.

    DONE
  5. 05 · DÉLAI

    Re-read payment terms. The term written on the invoice, if there is one.

    DONE
  6. 06 · NATURE

    Re-read document type. Invoice, credit note, or not a payable at all.

    DONE
  7. 07 · ÉCHÉANCE

    Set due dates. A date set when it can be derived; otherwise, nothing.

    DONE
  8. 08 · RATTACHEMENT

    Match payables. Each invoice finds its supplier in the reference list, or is marked as sender not recognised.

    DONE
  9. 09 · REGISTRE

    Enter new invoices into the queue. OK, information missing, or unclassifiable. Never “roughly”.

    DONE

A degree of certainty, not a result.

  • GREEN Pre-approved.

    Exact reference, EAN code or a previously approved match: the system is certain, so it prepares it.

  • AMBER One candidate, one click required.

    A resemblance: the system shows what it proposes and writes nothing until a person has clicked.

  • RED Nothing pre-selected.

    No stage succeeded: the line stays empty, and says so.

A step that fails is named, with its reason, and does not stop the others. A failure is never shown as a success. That rule had one exception, from 28 August to 2 September: a failure showed green. Measured, fixed, dated.

03Human intervention

What it does not do

No import deletes anything.
A line missing from a file is deactivated, within a scope that is shown. Two sources that disagree produce a flagged conflict. An import can be undone.
No uncertain match is approved by default.
Green means pre-approved, amber asks for a click, red pre-selects nothing.
The system only learns on approval.
A hypothesis that was looked at is not a hypothesis that was approved. Only the correction a person approved is kept.
The same file is only extracted once.
Every document carries its fingerprint; its extraction is cached and never paid for again.

What the person keeps

  • They approve every order: the only moment the system learns.
  • They click when a line is amber, choose when it is red, fill in what is missing.
  • They review every price change, every deactivation, every conflict.
  • They extend the reference lists from the screen; nothing overwrites them at startup.
RULES
The system does what is certain, shows what is doubtful, names what is missing.
HUMAN
The person approves, corrects and decides.

04What proves it

02·08·2026
went live, the day after the first commit
118
updates reviewed and shipped in seven weeks
145
entries in a log written for non-technical readers
64
changes to the database, applied automatically each time a new version goes live

05Data

READS
PDF, Excel; the accounts payable mailbox, with a permission restricted on the Microsoft side to that one mailbox; Teams and SharePoint folders, read-only.
WRITES
Its own database, its own document store, the Excel workbooks it produces.
DOES NOT WRITE
The ERP, the mail system, SharePoint.
KEEPS
The source document for six months, then a daily pass deletes it. The extraction, the lines and the approval remain.
LOGS
Filing, approval and every change in a document's state are recorded with their author, in the same transaction. Nothing in the system deletes that log.

The common rule is here. →

06Where it actually stands

Went live on 2 August 2026, observed in production until 14 September. Filing invoices into SharePoint is prepared and simulated, not yet run. Time saved has not been measured, and this page does not estimate it.