NRY Blog4 min read

Automating order processing with AI: from e-mail and PDF into your ERP

How AI takes orders from e-mails, PDFs and spreadsheets into your ERP: the flow step by step, the exceptions that decide success, what to measure, and why order intake is often the best first AI process.

By Laas NuryPublished 29 September 2026

In many companies, a large part of the order desk's day looks the same: open an e-mail, open the attachment, find the customer in the ERP, type the lines in, look up the customer's own article codes, check the price, check the delivery date, confirm to the customer. Hundreds of times a week, with a new layout for every other customer.

It is also one of the best places to start with AI. The volume is high, the work is mostly rules with exceptions, the result is easy to check, and a draft order can always be corrected before it is booked. This article walks through the flow, the exceptions that decide success, what to measure and how to start.

Why order intake is a good first AI process

High volume. Dozens to hundreds of orders a day make a small saving per order add up quickly.
Checkable. Either the order in the ERP matches what the customer asked for, or it does not. That makes testing simple.
Reversible. A draft order can be corrected before it is booked or shipped.
Unstructured input. E-mails, PDFs and spreadsheets in every format are exactly what fixed automation struggles with and AI handles well.
Clear value. Faster confirmation for customers, fewer typing errors, and an order desk that spends its time on the orders that need judgement.

The flow, step by step

Receive. The AI watches the order inbox, the EDI exceptions queue or a portal, and picks up every new order, including orders in the body of an e-mail.
Read. It extracts the customer, reference, delivery address, requested date and every order line, from PDFs, spreadsheets, scans or plain text.
Identify the customer. It matches the sender and details to the customer in the ERP, including customers with several delivery addresses.
Match the articles. It maps each line to your article numbers, including the customer's own codes, old codes and descriptions instead of numbers.
Check against agreements. Prices, discounts, minimum quantities, packaging units, what this customer may order.
Check availability. Stock and the earliest possible delivery date, against what the customer asked for.
Create the draft. A draft sales order in the ERP, through the official interface, with a note on anything uncertain.
Route the exceptions. Everything it cannot match or that deviates goes to the order desk, with the reason and a proposal.
Confirm. Once approved, the order confirmation goes out, in the customer's language.

The exceptions that decide success

The straightforward orders are easy. These are the ones that make or break the project:

Exception	What the AI should do
Unknown article or customer code	Propose the most likely match, send to the order desk
Price differs from the agreement	Flag the difference, do not change the price itself
Quantity does not fit the packaging unit	Propose the nearest valid quantity, flag it
Requested date cannot be met	Propose the earliest possible date, flag it
One e-mail with several orders, or a change to an earlier order	Split or link, flag the change
New customer, or a new delivery address	Send to a person; never create master data on its own
Unreadable scan or incomplete order	Ask the customer, or send to the order desk

Write these down with the order desk before the build starts. The list is also a good description of the process, and often the first time anyone has written it down.

What to measure

Measure	Why it matters
Share of orders that need no correction	The core measure of accuracy
Share of orders that go to a person	Shows how much work remains, and where
Time from receipt to confirmation	What the customer notices
Errors that reached the customer	The cost of mistakes, not just their number
Time per order for the order desk	The saving, including checking time

Measure the current process the same way before you start. Without a baseline, nobody can say later what changed.

EDI, OCR or AI?

EDI is the best route for large customers who can send structured orders. Keep it.
OCR tools with templates work for a few customers with fixed forms, but every new layout means a new template, and free-text orders are out of reach.
AI covers the long tail: every customer, every format, including orders written in the text of an e-mail. It also does the matching and checking that OCR leaves to people.

In practice: EDI for the customers who can, AI for everyone else, and one exception queue for the order desk.

How to start

Collect a few hundred recent orders with the order as it was eventually booked. That is your test set.
Write down the rules and exceptions with the order desk (the table above is a start).
Run the AI in the background for a few weeks: it reads and proposes, the order desk works as usual, and you compare.
Switch to drafts: the AI creates draft orders, the order desk approves and corrects.
Let it book the straightforward orders below agreed limits, once the numbers justify it.

If you use Exact Online or AFAS, connecting AI to Exact Online and AFAS explains how to connect safely. If you run SAP or Dynamics 365, read AI on top of SAP and Microsoft Dynamics. And if you want to know what a build like this costs, having an AI agent built explains what moves the price.

Questions and answers

Can AI read orders from e-mails and PDFs reliably?

For most orders, yes: current AI reads orders in the body of an e-mail, in PDFs, in spreadsheets and in scans, without a template per customer. Reliability comes from what happens next: matching against your customer and article data, checking against agreements, and sending anything uncertain to a person.

We already use EDI. Do we still need this?

Often, for the rest. EDI handles the customers who send structured orders. Most companies also have a long tail of customers who order by e-mail, PDF or phone follow-up. That tail is usually where the order desk spends most of its time, and where AI helps.

What is the difference with OCR software for orders?

Classic OCR tools read fixed positions on a page and need a template per layout; a new layout or a changed form breaks them. AI reads the document the way a person does, so a new customer or a new layout is not a new project. It also handles orders written as free text in the body of an e-mail.

Does the AI place orders in our ERP by itself?

It can, but it should not start that way. The safe route is: the AI creates draft orders, the order desk approves them, and only once the numbers show it can be trusted does it book straightforward orders by itself, below agreed limits. Exceptions always go to a person.

How long does it take to automate order intake?

For one ERP and a clear set of customers, a first version can run within weeks: first in the background next to the order desk, then creating drafts. Most of the time goes into agreeing the rules for customer-specific prices, article codes and exceptions.

About the author

Laas Nury, Founder, NRY Partners

Laas Nury is the founder of NRY Partners.