20 Orders a Day. 3 Part-Time Staff. And Still, They Couldn't Keep Up.
Our client — a mid-sized manufacturer with around 150 employees — faced a problem familiar to many manufacturing companies: orders arrived via email. As PDFs. As free text. Sometimes via WhatsApp. Sometimes even by fax.
And every single one had to be manually typed into the ERP system.
Item number, quantity, delivery date. Line item by line item. Some orders had 10 line items. Others had over 100.
Three part-time employees were exclusively occupied with this. And still, it sometimes took hours for an order to be in the system.
The Problem Nobody Talks About
In almost every manufacturing company, there's someone who spends all day typing orders. Often in internal sales. In smaller companies, the sales manager does it themselves — or even the CEO.
On average, manual order entry in the manufacturing industry takes 17 minutes. At an hourly rate of 60 Euros, that's almost 20 Euros per order — just for typing.
With 20 orders a day, that's 400 Euros daily. Over 8,000 Euros a month. Almost 100,000 Euros annually — for an activity that creates no value whatsoever.
And those are just the direct costs. Add to that:
- Typing errors that trigger incorrect goods — returns, re-manufacturing, delivery delays
- Slow processing that frustrates customers and, in the worst case, leads to lost orders
- Employee frustration — hardly anyone enjoys typing item numbers all day
- No growth without more staff — every increase in order volume requires new people
Why EDI Doesn't Solve the Problem
The obvious answer would be: implement EDI. Electronic Data Interchange. Standardized formats. Direct system-to-system communication.
Perfect in theory. In practice:
- Too expensive for many mid-sized companies — setup and operation quickly cost five figures
- Too rigid — each customer needs their own connection, every format change becomes an IT project
- Never complete — even companies with EDI still receive some orders via email and PDF
EDI solves the problem for the largest customers. But the long tail of smaller orders remains manual.
The Other Way: Forward Emails Instead of Typing
Our approach was deliberately pragmatic. No big IT project. No system change. No training.
The process is simple:
- An order arrives via email — as always
- The clerk forwards it to a special address — one click
- Our AI reads the email and the PDF attachment, extracting all line items: item numbers, quantities, customer data
- It cross-references everything against the item and customer master data from the ERP — does the item exist? Is the number correct?
- The order is created as a draft directly in the ERP
- The clerk receives a reply email with the processing status and a link to the ERP
Instead of 15-30 minutes of typing: Forward, quickly check, approve. Done.
And the best part: nothing changes for the end customers. They continue to order via email, in their format, in their language. No customer portal, no EDI onboarding, no "Please use our new order form from now on."
What Happens When the AI Is Uncertain?
That's a valid question — and one we often hear. The answer: It tells you.
If an item number isn't found, a quantity seems implausible, or a scanned PDF is poorly legible, the system flags the affected line item. The clerk immediately sees where they need to intervene — and where not.
The Numbers After 6 Months
In the first months, we processed a total of 16,500 order line items automatically. Of these, 51 errors were reported.
This corresponds to an accuracy of 99.7%.
For comparison: The average error rate for manual data entry, depending on the study, is between 1 and 5%. For 16,500 line items, that would be 165 to 825 errors.
The three part-time employees who previously spent all day typing? They now only check the results and handle genuine exceptions. The rest of their time goes into tasks that actually create value — customer support, complaint processing, quote generation.
What We Learned from 16,500 Line Items
Every order that passes through our system teaches us something. A few insights that even surprised us:
Customers often order with outdated item numbers. Products are renamed, numbers change — but the customer's buyer still has the old number in their Excel list. A human sometimes recognizes this. Sometimes not. Our AI cross-references it against the current master data and suggests the correct item.
Complexity lies in the structure, not the data. Many orders have nested formats: a header with the item number, below it several sub-items with different quantities and delivery dates. Or line items that are crossed out but still appear in the PDF. Or tables that span page breaks. These are the cases where simple OCR solutions fail — and where an AI that understands context makes the difference.
Free-text emails are surprisingly well-processable. We had expected PDFs to be easier than free text. The opposite is often the case. While free-text orders are unstructured, the information is clearly legible. Poorly scanned PDFs, on the other hand, are the real challenge.
Orders arrive via more channels than you think. Email and PDF are standard. But in practice, orders also come via WhatsApp, fax, or as a photo from a mobile phone. Yes, really. And yes, those can also be processed.
From First Idea to First Processed Order: 3 Weeks
Setup took three weeks. This includes:
- Export of item and customer master data from the ERP
- Configuration of email forwarding
- Training the AI on the specific order formats of end customers
- Testing phase with real orders
After that, we continued to optimize iteratively — with every reported error, the system gets better.
The effort on the customer's side? Minimal. A data export from the ERP and an email forwarding rule. No IT project. No new software. No training.
Who This Works For — And Who It Doesn't
This approach is suitable for manufacturing companies that:
- Regularly receive orders via email or PDF
- Currently type these manually into the ERP system
- Have a certain volume (from approx. 5-10 orders per day, it becomes noticeable)
- Sell standard items (item number-based)
It is less suitable for:
- Purely configurable products without fixed item numbers (we have other solutions for these)
- Companies that only accept orders via EDI or webshops
- Very low order volumes (1-2 per day — the effort isn't worth it)
The First Step Costs Nothing
If you recognize yourself in the description above, there's a simple way to find out if automation works for you: Send us 5 real orders (anonymized, if preferred). We'll show you within 48 hours what our system makes of them.
No sales pitch. No commitment. Just a test with real data.
