ITLine

In development · looking for pilot partners

Article numbers from e-mail

A configurable AI layer between incoming orders and your own product catalogue.

Orders arrive by e-mail, as free text, every partner in their own language. Today someone does that translation by hand. This is the layer we are building — both directions, configurable per partner, with human approval.

Request the two-hour reviewShow me the numbers first

What we have measured so far

line items processed
1082
from real orders that arrived by e-mail — 3.9 items per order on average
order needed manual entry
1
in the entire period — order entry as manual work stopped, it did not merely shrink
line items correct first time
80,6%
on hard, real-world letters — up from 78.6% after a single improvement round
weeks missed
0
weekly: 57 · 55 · 35 · 50 · 39 · 20 · 40 orders — not a demo, an operation

Measured: August 2026. In production since June 2026, in a working environment, on real orders.

Start with a real letter

A typical order, exactly as it arrives — no article number, no form, not even a full sentence.

Subject: re-ordering

Hi! I need two more of the last lot, anthracite, the 120 one.

And three of the one that got left off then.

Thanks!

The letter is invented, but its shape is not: this is precisely the kind of input the system learned on.

Five to ten minutes, for one short letter. Someone digs out the previous order, works out what "the 120 one" is, checks which anthracite exists in the catalogue and looks up the article numbers line by line.

This is what it becomes, before anyone looks at it

One screen showing the incoming letter and the line items extracted from it, with article numbers, side by side. The green row is done. On the amber one the system does not choose for you — it lists the candidates and the operator decides with one click.

This is what it becomes, before anyone looks at it

Incoming letter

Subject: re-ordering

Hi! I need two more of the last lot, anthracite, the 120 one.

And three of the one that got left off then.

Thanks!

Extracted items

PR-120-07profile, 120 mm, anthracite2 pcs

unambiguous — ready to approve

two candidatesthe "left off" item3 pcs

two products fit the letter — one click

  • PR-120-05profile, 120 mm, grey
  • PR-140-07profile, 140 mm, anthracite
RK-120-07fixing kit, anthracite2 pkg

this partner always orders it with the above — the system offered it

3 items · 1 decision neededEditApprove
Illustration, with invented article numbers.

That is the difference: nothing to look up — only something to approve.

Today, and with the layer

Today

8–15 min / order

  1. reading, interpreting1–2 min
  2. finding the partner and history1 min
  3. article lookup, line by line4–8 min
  4. entry, confirmation2 min

With the layer

1–3 min / order

  1. review1 min
  2. correction, where amber0–1 min
  3. approvalseconds

The longest step — the lookup — is the one that goes away.

What this is worth in money

Fill it in with your own numbers. Every assumption is visible and adjustable — that is the only way the calculation is honest.

Assumptions

20
8 min
HUF 4,000

Calculated with 21 working days.

Time freed per month
56 hours
In wage cost per month
HUF 224,000
Per year
2.7 M Ft

Wage cost is the floor, not the reality

This measures what the work costs to perform — not what the time it consumes is worth. If an administrator fills the freed hours, this is the right number. If a salesperson does, it significantly understates it. We free up capacity; we do not promise revenue growth.

The second leg: the cost of a wrong article number

The second leg is the cost of a wrong article number, and it matters more. A wrong code is not an error message — it is the wrong goods delivered: freight both ways, re-picking, explaining yourself to the partner and less patience on the next order. Nothing in the current process books this cost, yet it is real, and it does not scale with order volume.

What we measured on our own system

This section comes early because every other claim stands or falls on it. Anyone telling you today that their system recognises orders with 100% accuracy either has not measured it, or is not telling you what they count.

The model's confidence versus reality

Of the line items in a given confidence band, how many were actually correct

0255075100%98.1%1,00n = 10781.4%0,90–0,99n = 28551.4%0,80–0,89n = 3752.9%0,50–0,79n = 34
  • Measured accuracy — share of lines that were actually correct
  • What the confidence score promises

The interface labels this band "high AI confidence". In this band every second row is wrong.

241 live orders replayed, 465 paired line items. Correctness is measured against the item the operator actually approved.

Data as a table
Confidence bandnMeasured accuracy — share of lines that were actually correct
1,0010798.1%
0,90–0,9928581.4%
0,80–0,893751.4%
0,50–0,793452.9%

81%

of erroneous line items arrived with high (≥ 0.80) confidence

68%

of those silently — the system did not ask back at all

Four fifths of the errors are confident, and two thirds of those are silent.

The safety net is not the score. The safety net is that every article number is validated against the catalogue, and that a human approves before anything financial happens.

Where this is going

GoalWhat it rests on
Over 95% first-time-correct at tuned, returning partnersthe partner dictionary and the correction history narrow the error space every round; this is the strongest signal today
Effectively flawless recognition of recurring phrasingswhat was corrected once is stored as a concrete mapping — never re-estimated
The remaining uncertainty must not be silentthis is the most important open work: today only a quarter of erroneous lines trigger a question back

A 90% system that flags the remaining 10% is more useful than a 95% one that stays quiet.

Why "just hand it to an AI" is not enough

Hand the whole letter to a language model and ask it to "return the article numbers", and the model always answers. Faced with an unknown code it does not say it does not know: it produces a plausible-looking article number and attaches confidence to it.

Naive approach

  • Letter
  • AI: "what's the article number?"
  • Article number — inventable

What we build

  • Letter
  • AI: attributes
  • Code builds the key
  • In the catalogue?
    • Article number
    • Ask back

The model reads out attributes. Code builds the article number. And the catalogue closes the chain.

The assembled code is not an answer but a lookup key: the system immediately checks it against the catalogue, and if no such article exists it does not return the constructed code — it applies the cutting and rounding rules, or asks back. A constructed but non-existent article number therefore physically cannot leave the system.

Two cases from our own log

The one the AI got wrong

A letter contained an article number verbatim. The system did not recognise it, decomposed it into attributes, did not know one of the type codes, invented a meaning for it — and returned a different product, with 0.72 confidence. It was confidently wrong.

What we built from it

The fix was a deterministic matching layer placed ahead of attribute decomposition: no rule decides what counts as an article number — the system extracts a permissive candidate set and the catalogue decides.

The one the AI did not get wrong

In one order the size was stated correctly in the text, and the model read it correctly. The deterministic mapping, however, found no size-specific key, fell back to a general pattern and picked the alphabetically first of the matching articles. The customer had asked for a larger size and received 1,150 pieces of a smaller one.

What we built from it

There is zero AI error in this. "Smallest element wins alphabetically" was not a decision but the absence of a default. What we built from it: if more than one article matches a mapping rule and there is no distinguishing data, that is uncertain by definition — it earns a question back, not a sort order.

How it works

We use AI where there is linguistic ambiguity, and only there. The full chain has nine steps, and six of them are deterministic code, not a model.

StepPerformed by
Letter normalisation, attachment extraction, thread handlingcode
Classification, partner identificationmodel + rules
Extracting article-number candidates from the textcode
Matching against the cataloguecode + database
Free text → attributesmodel
Review — checking the extracted itemsa second model
Attributes → lookup key → catalogue itemcode
Context filtering — what excludes it, what belongs with itcode + knowledge base
Final decisionhuman

Three outcomes, not one

StateWhat it meansWhat happens
Unambiguousone candidate, and the rules agreeready to approve
Several candidatesmore than one product fits the textflagged, candidates listed — one click
No matchnothing in the catalogue fitsa new product, a typo or genuinely missing

The goal is not that there is never an amber row. The goal is that what is uncertain shows up amber — and not green.

Every correction is a one-off

What the operator corrects once, the system remembers — not as a statistic but as a concrete mapping for that partner, and it takes effect on the very next letter. No retraining, no months of data collection: the daily work produces the training material.

concrete mappings
895
all from human corrections, in the live system
partner rules
154
per partner, for their own phrasings

This is not a one-off rollout but a continuous shift toward automation. Start now, and in six months you have a tuned layer, trained on your own partners.

Not new software to learn

This is the most important difference from a usual software rollout: it is not a third interface to maintain, train people on and get accepted.

  • It fits your existing mail

    Microsoft 365 or Google, exactly as you use it today. Not a third interface to maintain, train people on and get accepted.

  • With permissions

    Who can see what, who can approve, who can change a rule — and two-factor authentication can be required for settings.

  • It can hand data on

    The output does not stop at a screen: it can pass data on to your ERP or invoicing system, if that has an API.

  • Callable from your own AI assistant

    Under construction

    The aim is for the mapping knowledge to plug into the assistant your company already uses and trusts — ChatGPT, Claude or Microsoft Copilot — over the open MCP standard. No separate login: called from there.

What to compare it against

Yes, this category exists, and it is worth knowing what you are choosing between. In the enterprise market it is well established, and in 2026 the ERP vendors entered too — Microsoft Dynamics 365 Business Central’s order agent became generally available in November 2025. That is good news: the market has settled the question of whether this layer is needed. The only question is what you do if you have no ERP for it to be built into — or you do not want to replace your platform for a single feature.

What you can buy todayWhat it assumes
Enterprise order-automation platformERP integration, an annual platform fee, per-trading-partner pricing
Order agent built into an ERPthat you run that vendor’s ERP
Simple document-extraction toolsthat no per-customer tuning is needed — precisely the thing that matters here
An in-house automation projectthat maintenance afterwards becomes your problem

Four things that make this layer different

  1. 01

    It does not require an ERP

    It works even if you run on spreadsheets and a mailbox today.

  2. 02

    Built for Hungarian, informal, dictated text

    Not for clean PDFs. The dictionaries and unit rules are built on how Hungarian trade letters are actually written.

  3. 03

    It learns per partner

    Not one global model trying to understand everybody.

  4. 04

    Open

    The mapping knowledge will be callable from your own AI assistant over an open protocol.

A note on timing

A timing note for any company on Microsoft 365: Microsoft is retiring EWS, the legacy mail interface. Shutdown begins on 1 October 2026 and completes by April 2027. Buy an EWS-based integration today and you will be migrating next year. This system is built on its successor, Microsoft Graph, from the start.

When this is genuinely not the answer

Plainly: if your orders do not arrive as free text but in a machine format — EDI, a webshop basket, a completed form — there is nothing to translate and this layer is redundant. And if you have no returning partners with their own phrasings, the learning loop that drives the whole thing cannot get going.

Where this runs in production today

WPC Pont

WPC Szaküzletek Kft.

June 2026

The core of this layer runs in production today, in a working environment, on real orders, continuously since June 2026. Not a prototype: there are delivered goods behind it. The system is WPC Pont — a complete business system we built for WPC Szaküzletek Kft., and where this interpretation layer first went into live operation.

The full reference →

What we offer, and what we ask in return

First step — a two-hour joint review, free of charge

We sit down for two hours, go through 20–30 real letters and measure how much time their processing takes today and how much would remain. This is not a presentation and not a sales conversation: at the end we hold a number.

What happens in those two hours

  1. 15 minyour process today: where orders come from, who handles them, where it jams
  2. 45 mingoing through 10–15 real letters together
  3. 30 minthe typical traps: article mismatches, missing data, confirmation
  4. 20 mintiming and savings estimate — the calculator above, with your numbers
  5. 10 mindecision: is a pilot worth discussing

About the data: we can review the letters on the spot, over a screen share — nothing needs to be sent in advance. If samples are needed later, we ask for them under a separate agreement or anonymised.

Second step — a four-week pilot

WeekWhat happens
1.connecting the mail source, capturing article rules for the top 3–5 partners
2.first live processing, tuning the approval screen
3.measuring accuracy on real letters, extending dictionaries and rules
4.reverse translation onto outgoing documents, evaluation

At the end you hold a measured result: how many letters went through, how many lines were right first time, how much time it actually took. The accuracy figure becomes your own number, not a promise lifted from a brochure.

What we ask in return

What we ask in return: real letter samples (they can be anonymised), one hour a week from the colleague who does this work today and — if it works — the right to cite you as a reference, with the details approved by you.

The first step is free of charge. We ask nothing for the joint review and the measurement, and it carries no obligation. If it is not worth it, we will say so.

Next step

A two-hour joint review, with 20–30 real letters. Nothing is binding until then.

Or directly:

Tatár Gábor

gabor.tatar@itline.hu

+36 20 391 1669