Customer references in orders: how AI finds the right item in your catalogue

When a customer orders using their own code, an old reference or just a description, AI does not look for an exact match. It weighs several clues (the code received, the description, the price quoted, this customer’s order history) against the item records in your ERP, and proposes the most likely item with a confidence score. Lines it is sure about go through according to your rules; ambiguous lines are flagged to the sales admin team.

One line, a thousand ways to write it

Take an electrical wholesaler. In its ERP, the item is called “Modular circuit breaker 1P+N 20 A C-curve 4.5 kA”, code DIS-1PN-20C. Here is how it reaches the sales admin inbox in a single week:

• The customer’s purchasing code: “Art. 004512”, the reference the contractor uses in their own purchasing software, which means nothing on your side.
• The manufacturer’s reference: the maker’s code, copied from a catalogue or a label.
• A damaged reference: “DIS1PN20C” without hyphens, or “DIS-1PN-2OC” with a letter O instead of a zero, typical of a scan.
• An old reference: the item was replaced last year, but the customer kept their old order template.
• A free-text description: “breaker ph+neutral 20A C curve”, in the body of the email.
• Nothing at all: “same as last time, twice over”.

None of these lines is wrong. Each one reflects the customer’s habits, tools and history with you. But none of them can be entered into the ERP as is. This is where much of sales admin’s time goes: searching, comparing, sometimes calling the customer back to be sure.

Why a cross-reference table is not enough

The usual answer is to keep a table that maps, customer by customer, their reference to yours. It is useful, and you should keep it. But in most companies it has the same weaknesses:

• it lives in a spreadsheet or in the heads of one or two people;
• it only gets updated after the fact, once an order has already caused a problem;
• it only covers regular customers and items they have already ordered;
• it says nothing when a customer switches software and therefore changes their codes.

A table answers the question “have I seen this code before?”. It cannot handle a free-text description, a typo or a replaced item. Yet these are exactly the lines that take time.

How AI reasons about an order line

Rather than looking for an identical code, AI treats each line as a bundle of clues and checks them against your data.

1. Read the whole line, not just the “reference” column. The code, description, quantity, unit and unit price written by the customer are extracted together. A description can correct a misread code, and vice versa.

2. Look at what this customer has already ordered. If the contractor has been buying the same 20 A breaker for two years, a line saying “brkr 20A C” very probably refers to it. Order history in the ERP is often the most reliable clue.

3. Compare attributes. For a free-text description, AI identifies the attributes that matter in your trade (rating, curve, number of poles, breaking capacity; or, in hygiene supplies, ply count, format and material) and compares them with the records in your catalogue.

4. Take replacements into account. If an old reference points to a substitute item in the ERP, the proposal is the active item, and the substitution is visible.

5. Assign a confidence score. Every proposal comes with a level of certainty. A customer code seen several times before, and consistent with the price, scores high. A description that fits three similar items scores low and goes for review.

This is what makes it possible to read orders without setting up a template for each document layout, as described on our order entry page. The same matching is used for price requests, where products are often described even more loosely: see quote management.

The clues people tend to forget

An experienced sales admin does not rely on the code alone. They look at everything else, and AI has to do the same.

The price. A customer who writes €6.20 next to an ambiguous line gives a valuable clue: between powder-free nitrile gloves at €6 a box and the heavy-duty version at €11, the doubt is gone. Conversely, a price far from the list price of the proposed item should lower the confidence.

The unit and the quantity. A line saying “4 cases” cannot match an item that is only sold by the piece. Quantity and pack size help identify the item, even before working out the quantity to enter. That second topic, converting between units, cases and pallets, has its own article: unit of measure conversion in orders.

The context of the message. “Same as the order of 3 October, except the fittings” refers to a specific order in the ERP. An AI that reads the whole email, not just the attachment, can find that order and work from it.

The customer themselves. The same wording does not necessarily mean the same item for two customers. A food distributor supplying both restaurants and school canteens knows this well: “35% cream” does not refer to the same pack size for every customer.

What should stay with the sales admin team

The goal is not for AI to decide at all costs. Some situations call for a human decision and should be presented as such:

• two items are equally plausible, and nothing in the history helps choose;
• the customer asks for a product they have never bought, or that is outside their usual range;
• the item has been discontinued and several replacements are possible;
• the reference is recognised, but the price or quantity is far from the customer’s habits.

In these cases, Eledone flags the line with a low confidence score and shows its proposals; the team decides. Everything else follows the rules defined with each customer: direct entry into the ERP for lines it is sure about, validation for others. And every decision the team makes adds to what the tool knows about that customer, so the same question does not come up again on the next order.

What helps AI in your catalogue

You do not need a perfect item master to get started. A few things do make a real difference, though:

• readable descriptions that mention the key attributes rather than internal abbreviations;
• replacements recorded in the ERP when one item supersedes another;
• inactive items clearly marked, so that products you no longer sell are not proposed;
• existing cross-reference tables, even incomplete ones, as a starting point.

The best test is still your own orders, with their customer codes, crooked scans and “same as usual”. You can submit one through our online trial to see how the lines are matched.

Frequently asked questions

Do we need to provide a cross-reference table to get started?

No, but if you have one, it will be used. AI relies mainly on your catalogue and on each customer’s order history in the ERP, which already contain most of the matches, even if nobody ever wrote them down.

What if the customer uses the manufacturer’s reference or an EAN barcode?

If those codes are stored in your item records, the match is direct. If not, AI falls back on the description, the price and the history, with a confidence score that reflects the uncertainty.

Can AI pick the wrong item without warning?

That is the risk the confidence score is there to limit. A line where several items are plausible, or where the price does not fit, is flagged instead of being entered. You set the level of strictness with us, customer by customer if needed.

What if the customer orders a replaced or discontinued item?

If the replacement is recorded in the ERP, the active item is proposed and the substitution is visible to the team. Otherwise, the line is flagged so that someone in sales admin can choose, and let the customer know if necessary.

Want to see how your own customer references are recognised?
Book a demo