An AI agent can read each customer order email, including the PDF or spreadsheet attached to it, match the customer, the delivery point and every line to your ERP master data, and create the sales order without a person retyping it. At a global fresh-produce company, our agents entered more than 1,000 retailer orders in their first seven weeks: about 650,000 cases of fruit for 63 distribution centres.
TL;DR
- Most order entry at large suppliers is still manual because the orders that do not come through EDI arrive as emails and PDFs, each customer with its own layout and its own codes.
- An order-entry agent watches the order inbox, reads every message, maps each line to your item numbers and units, checks your rules, and writes the sales order into the ERP.
- When it cannot resolve a field, it holds that one order and tells the owner which field and why.
- IT approves the mailbox and the ERP connection once; the agent writes only there, and every order it enters links back to the email it came from.
Why is order entry from email still manual at large companies?
EDI covers the large, stable flows between big trading partners. Everything else lands in a shared inbox. A retailer sends a PDF generated by its own purchasing system; a wholesaler pastes a table into the email body; a depot manager replies to last week’s confirmation with “same again, plus 20 cases for Thursday”. Each of those has to become a sales order with the right customer, the right ship-to address, your item numbers, your units and the right delivery date.
The people who do this know every customer’s quirks. They know that one retailer’s depot numbers map to your ship-to codes through a list kept in a spreadsheet, that another orders in pallets while your ERP counts cases, and that a customer order number appearing twice usually means the customer resent the same order. That knowledge is why the work has resisted automation. Template OCR and screen bots cope with one fixed layout at a time, and the inbox never holds still for long. We compare the approaches in AI agents vs RPA for order processing.
What does an order-entry agent do, step by step?
- Watch the order inbox. The agent reads each new message in the mailbox IT connected, including forwarded orders and attachments.
- Decide what the message is. A new order, a change to an existing order, a cancellation, or something else entirely, such as a delivery query or a newsletter. Only orders and changes continue.
- Identify the customer and the delivery point. The agent matches the sender and the delivery address or depot code to the customer and ship-to records in your ERP.
- Map every line. Customer article numbers and descriptions become your item numbers; pallets, layers or cartons become the unit your ERP expects.
- Check your rules. Cut-off times, delivery days per depot, minimum quantities, and the customer order number, which serves as the guard against entering the same order twice.
- Write the sales order. The agent creates the order through the ERP interface, with the customer’s order number in the external document field.
- Keep the receipt. Every order the agent enters links back to the email and the page of the attachment it came from, so anyone can check a line against its source in seconds.
What should the agent do when it is not sure?
It should stop on that order and say exactly why. “Item 40021 on line 3 has no match in the customer’s article list” gives the owner something to fix in a minute. A silent guess on an item or a depot can cost a truck, a missed delivery slot or a credit note days later.
This is an exception queue, and it is small by design. The agent enters the orders it can resolve completely and routes only the ones it cannot, each with the field and the reason, to the person who owns order entry. That person fixes the master data once, for example by adding the missing article mapping, and the next order from that customer goes through on its own. Approval on every order would put the team back to checking each line by hand, which is the work you set out to remove.
How does IT stay in control of an agent that writes to the ERP?
IT approves two connections: read access to the order mailbox, and the right to create sales orders in the named ERP companies. Anything else, such as posting an invoice, would need a connection IT has not granted.
Every run leaves an audit trail: which email the agent read, what it extracted, which records it matched, what it wrote and when. When finance or an auditor asks where an order line came from, you send them the link to the source email. We describe the model in more detail in what an AI agent audit trail needs.
Which customers should you automate first?
Rank your order senders on three questions before the pilot starts:
| Question | Start here | Leave for later |
|---|---|---|
| How many orders a week? | The senders behind most of the typing | A customer who orders twice a year |
| How stable is the format? | System-generated PDFs and spreadsheets | Handwritten or photographed order forms |
| How clean is the master data? | Customers with a maintained article list | Customers whose codes live in one person’s head |
For most suppliers, a handful of large retail customers account for most of the order lines. Getting those right first gives the team its time back quickly and builds the master-data mappings the rest will need.
What should you measure in the first weeks?
- Orders entered with no human touch, as a share of all orders the agent saw.
- Exceptions by reason: missing article mapping, unknown depot, past cut-off. Each reason points to a fix in the master data or the rules.
- Time from email to sales order, compared with the same customers before the agent.
- Corrections after entry: any order someone had to change after the agent created it, which is your clearest measure of accuracy.
What do the fresh-produce numbers show?
In their first seven weeks, our agents at the fresh-produce company entered 1,000+ retailer orders for three retail customers, covering about 650,000 cases (about 11,000 tonnes, estimated from typical carton weights) for 63 distribution centres. Those are the volumes our agents handled for those customers; they say nothing about the size of that company’s whole order book. The same company also runs our agents on quality claims, which we describe in automating a quality claims inbox.
FAQ
Can an AI agent enter sales orders into our ERP without a person retyping them?
Yes. The agent reads each order email and its attachments, matches the customer, delivery point and items to your ERP master data, and creates the sales order through the ERP interface. A global fresh-produce company running this kind of agent entered 1,000+ retailer orders in its first seven weeks.
What happens when the agent cannot match an item or a delivery point?
It holds that order, names the field it could not resolve and the reason, and sends it to the person who owns order entry. Every other order in the inbox keeps moving.
Does someone have to approve every order before it reaches the ERP?
No. The agent works on its own inside the connections IT approved. Teams in regulated industries can switch on an approval step for specific agents, but order entry usually runs without one, with exceptions routed to a person.
Which ERP systems does this work with?
Any ERP the agent can reach through an approved connection, for example Microsoft Dynamics 365 Business Central or SAP. The agent writes through the same interfaces your integrations already use.
How long does it take to get the first customer live?
A pilot starts with one recurring workflow, your real order emails, the current rules, and the known exceptions. Start with the customers that send the most orders in a stable format, then add the harder layouts.
Where to go next
Read which back-office work fits an AI agent today, see how Atlas works, or book a pilot with one order inbox and your real emails.