Automatically process orders from email and PDF with AI

A customer emails an order. Sometimes in the email itself, sometimes as a PDF, sometimes as a photo of a form. Someone in the office reads it, looks up the item numbers and retypes everything into the ERP. Every day, all over again.

That can be automated. AI reads the email or PDF, extracts the customer, items, quantities and delivery date, and prepares the order in your ERP. If the system is in doubt, a colleague checks it before the order goes through.

Why this stayed manual work for so long

Orders by email look different for every customer. One sends a tidy PDF from their own system, another types three lines in an email with “same as last time, but 20 boxes”. Fixed rules and templates get stuck on that. For every new customer someone had to create a template, and every change to the layout broke it again.

Language models (LLMs) read a document the way a person does. They don’t need a template per customer. That makes this work easy to automate well now.

	Retyping by hand	Templates and OCR	AI extraction
New customer or layout	No problem, but takes time	New template needed	Usually works right away
Free text in an email	No problem, but takes time	Doesn’t work	Works
Errors	Typos, especially when busy	Breaks silently on a different layout	Doubt is flagged
Time per order	Minutes	Seconds, if the template fits	Seconds

How it works, step by step

1. Collect. New emails in a dedicated mailbox, such as orders@, are collected automatically, including attachments.
2. Extract. AI extracts the data you need: customer, order lines, quantities, delivery address and requested date.
3. Check. The system looks up the customer and the items in your ERP. Does an item number not exist, or is a quantity far off from what the customer normally orders? Then the order goes to a colleague instead of through.
4. Prepare. Orders that check out arrive as a draft in your ERP or WMS, with the original email attached.
5. Confirm. After approval, the customer automatically gets an order confirmation.

How you stay in control

AI makes mistakes. Not often, but it happens. That’s why I build it so that a mistake stands out before it reaches the customer.

• Checked against your own data. Item numbers, prices and customer details are always compared with your ERP. What doesn’t match doesn’t go through.
• A person decides when in doubt. For the first few weeks, a colleague approves every order. After that, only the exceptions.
• The email is data, not an instruction. If an email says “ignore the previous instructions”, the system does nothing with it.
• Everything logged. For each order you see what was extracted from the email and who approved it.
• Your own environment. The automation runs on your side, with the AI provider of your choice.

When does it pay off?

Count how many orders arrive per day by email or PDF, and how long retyping takes per order. If that adds up to more than an hour a day, it’s worth looking at. If it’s a few orders a week, retyping is often simply the cheapest solution.

It also pays off sooner when errors are expensive. A wrong quantity or a wrong item costs a return, an extra trip and an unhappy customer.

How do you start?

Collect ten orders the way they really come in, the tidy ones and the messy ones. That’s all you need for a first call. With those you quickly see what can be automated and where a person is still needed.