Food distributors automate retailer orders with an AI agent that reads each order from the inbox, works out which depot it is for, maps the retailer's article numbers and pack sizes to the supplier's items, and enters it in the ERP before the cut-off. In their first seven weeks, our agents at a fresh-produce multinational entered 1,000+ retailer orders, about 650,000 cases (about 11,000 tonnes) of fruit, for 63 distribution centres.
TL;DR
- Retailer orders for fresh produce arrive per depot, often as email text or PDF, with a cut-off time and a shelf life that leaves no room for a backlog.
- The hard parts are the ship-to for each depot, the mapping from retailer articles to your items and pack sizes, and changes to orders that are already entered.
- The agent does that work end to end and sends only the unclear cases to the order desk, each with the reason.
- You start with one retailer: replay its order history and compare the result with what your desk entered.
Why is food order entry harder than generic order entry?
A generic order has a customer, a few lines and one delivery address. A retailer order for fresh produce has the same parts, but each one carries more risk. The order is for a single distribution centre, and one retailer may run dozens of them. Each depot has its own ship-to in your ERP, its own delivery day and often its own order reference. The goods are ripe fruit with a short life, so an order that sits unentered until the afternoon can mean a missed truck.
The mistakes also cost more. A wrong depot sends a pallet to a warehouse that did not ask for it, and the depot that did ask for it receives nothing. A missed cut-off means the order does not ship that day. Neither can be fixed by correcting a field the next morning, because by then the fruit is somewhere else or is no longer sellable.
What makes retailer orders hard to read?
Four things come up at almost every supplier.
- Many depots, one sender. The same retailer sends orders for different depots from the same address. The depot is in the subject, the attachment name, a header on the PDF or a code in the table, and the place differs by retailer.
- Different article numbers. The retailer calls an item by its own number and description. You sell it under your item number, in your pack sizes. One retailer line may be a case of 18 kg, another a pallet, another a weight in kilos, and each needs converting to the unit your ERP expects.
- Changes the same day. A quantity goes up after a promotion is confirmed, a depot cancels because of a delivery problem, or the retailer re-sends the whole order with one line changed. The change has to land on the order already entered.
- Promotions. Promotion volumes arrive in the same format as normal orders but follow a different schedule, and a promotion line can carry a different price or reference.
The general version of this work is covered in how to automate sales order entry from customer emails. This article covers what changes when the goods are perishable and the customer is a retailer with many depots.
What does the agent do with an order, from inbox to ERP?
- Reads the order from the email body, the PDF or the spreadsheet, whichever the retailer sends.
- Identifies the depot and matches it to the right ship-to and delivery date in the ERP.
- Maps the articles: retailer article number to your item number, and the retailer's pack unit to your cases, pallets or kilos, using the mapping your desk already follows.
- Checks for an order already entered for that depot and date, so a re-sent order or a change updates the existing one.
- Enters the sales order in the ERP with the lines, quantities, dates and the retailer's order reference, then files the source email with it.
- Sends exceptions to the order desk with the reason: an article it cannot map, a depot it does not recognise, an order that arrives after the cut-off.
For Microsoft Dynamics 365 Business Central, the write path is described in how an AI agent enters sales orders in Business Central.
How does the agent handle changes and cancellations?
A change is an order that refers to another order. The agent looks for the earlier order from the same retailer, depot and delivery date. If it finds one, it updates the lines or cancels the order, and it keeps the earlier version in the record. If it finds none, or it finds two that could match, it does not guess. It sends the email to the order desk with the reason, because a change applied to the wrong order is worse than a change applied a few minutes late.
The same applies to the cut-off. When a change arrives after the point where the ERP no longer accepts edits, the agent says so in the exception, so the desk can ring the depot while there is still time to act.
What does the agent leave to your people?
It enters what the customer ordered. It does not plan allocation. When supply is short, deciding which depot gets how much of the fruit that exists is a commercial call for the people who own the retailer relationship, and it stays with them. What changes is the starting point: every order is already in the ERP, correct and complete, when they sit down to plan.
Agents run on their own by default. Approval steps exist only where a customer asks for them. The order desk sees only the exceptions, each with its reason.
Where do these numbers come from?
The order figures at the top of this article come from our agents at a fresh-produce multinational. The same company runs a second agent on its claims inbox, which has read 7,000+ emails so far and logged 1,700 quality claims from depots in 17 European countries. That work is described in what happens when an AI agent reads every email in a claims inbox.
How do you start with your own retailers?
- Choose one retailer, ideally one with many depots and a steady order pattern.
- Collect a few weeks of its orders, including the changes, cancellations and re-sends, and the matching sales orders from your ERP.
- Give the agent the lists your desk already uses: depots and ship-tos, retailer articles and your items, pack-size rules.
- Replay the history. For each order, compare what the agent would have entered with what your desk entered, line by line, and list every difference.
- For each difference, decide whether the agent or the desk was right. Fix the rule or the list, and run the history again.
- Go live on that retailer only, and add the next one when the exceptions are down to a short list.
If your retailer also sends EDI, the choice between channels is covered in EDI versus email orders. For the wider picture, see AI agents for back-office operations.
FAQ
How do food distributors automate retailer orders?
An AI agent reads each order as it arrives by email or PDF, finds the retailer, the depot and the delivery date, maps the retailer article numbers and pack sizes to the supplier item numbers, and enters the order in the ERP with the right ship-to. Changes and cancellations are matched to the order already entered. At a fresh-produce multinational, our agents entered 1,000+ retailer orders in their first seven weeks, about 650,000 cases (about 11,000 tonnes) of fruit, for 63 distribution centres.
What happens when an order changes after it was entered?
The agent finds the original order for the same depot and delivery date and updates it, or cancels it, instead of entering a second order. If it cannot tell which order a change belongs to, or the order is already past the cut-off in the ERP, it sends the case to the order desk with the reason.
Does the agent decide which fruit goes to which depot when stock is short?
No. The agent enters what the customer ordered. Allocation when supply is short stays with the people who plan it, and the agent hands them an accurate, complete picture of the orders.
How do you start with one retailer?
Pick one retailer, collect a few weeks of its real orders, and run the agent on that history. Compare each order the agent would have entered with what your desk entered, line by line. Then run it live on that retailer only.
Where to go next
Read how to automate sales order entry from customer emails, or see how Atlas works. When you have one retailer in mind, book a pilot and bring a few weeks of its orders.