← All posts Guide · Nov 20, 2026

How to match customer product descriptions to your own article numbers

Lennard Kooy··16 min read
How to match customer product descriptions to your own article numbers

To match customer product descriptions to your own article numbers, keep a cross-reference of each customer's descriptions and part numbers against your articles. Let an AI agent propose the match on every incoming order line from that history and your product master. A person confirms only the uncertain lines, and each confirmation teaches the next order.

At Lleverage we think the 2 fixes most companies try first both fail in practice. Fuzzy matching on the text alone picks the article whose description looks most alike, which is often the wrong size or the wrong pack. Asking customers to order with your codes works for a month, until the next buyer at the customer starts and types what they always typed. What works is a mapping per customer that keeps learning, kept next to the ERP you already run. You need no new PIM for it. Our order management agents do this work on the order desk every day, and you can book a demo to see them run on your own orders.

Picture the order inbox of a wholesaler in technical parts. Brakveld Techniek writes "lager 6205 2RS C3, 10 st". Noordkamp Lagers B.V. sends a PDF with its own code, NKL-44871, and no description at all. A third customer pastes a line from a quote you sent last spring. All 3 mean the same article in your ERP. Someone on your desk knows that, because they have typed it 200 times. This guide shows how to get that knowledge out of their head and into a place where an agent can use it.

Why do customers describe the same product differently?

Customers describe products in the words of their own systems and habits, and those words rarely match your item master. Each buyer has learned your range from a different angle: a catalogue from 2019, a quote, a sales rep's spoken name for a product, or their own purchasing system.

The differences fall into a few groups, and each one breaks a different matching method.

  • Abbreviations and shorthand. "verz." for verzinkt, "SS" for stainless steel, "2RS" written as "2 RS" or "RS2". Each buyer has a style, and it is consistent per buyer.
  • Old article numbers. If you renumbered your range a few years ago, many customers still order with the codes from before, because those codes sit in their own ERP.
  • The customer's own part number. Larger customers order from their purchasing system, which prints their internal code on the purchase order. Your number may not appear at all.
  • Supplier and manufacturer codes. A customer orders by the manufacturer's part number, or by the code of a competitor they bought from before. Both point to an article you stock under your own number.
  • Units and pack sizes. "1 doos" can mean a box of 10 or a box of 100. "3 rol" of cable can be 3 coils of 100 metres or a typo for 3 metres.
  • Typos and free text. "Same as last time but 40 instead of 25" carries no product name at all, only a reference to an earlier order.

A desk that has worked with a customer for years resolves all of this from memory. The cost shows when that person is on holiday, or when a new colleague takes months to learn which customer means what. At SIG Benelux, a building materials distributor, around 3% of orders carried an article number and roughly 47% had only a description. The rest came in as messages and photos that only made sense to someone who knew the customer.

Where does the customer item cross-reference live in your ERP?

Most ERPs already have a table built for this job: a list that links a customer's own item number and description to your article. It goes by a different name in each system, and in most companies it holds far fewer rows than the desk's memory does.

ERPWhere the cross-reference livesWhat it can hold
Microsoft Dynamics 365 Business CentralItem References, opened from the item cardThe customer's or vendor's item number and description, plus a unit of measure and variant code that are copied to the document line
SAP (ECC and S/4HANA)Customer-material information records, created with transaction VD51 or the Manage Customer Materials appThe customer-specific material number and description, and delivery tolerances agreed with that customer
Other ERPsOften a customer article table, sometimes an add-onAsk your ERP partner which fields it holds and whether order entry reads it

In Business Central, item references link the terms you and a customer use for the same item. When someone fills in the item reference number on a line, the customer's description, unit of measure and variant code are copied to the document. In SAP, the customer-material information record stores the material number a customer uses where it differs from yours. During order entry, a line can be entered with either number, and the system finds your material. So the function is there in both systems, and in most companies it is close to empty.

Why does the cross-reference table stay empty?

The cross-reference stays empty because filling it is extra work at the worst possible moment. The desk person who has just worked out what a customer meant is in the middle of an order, with 30 more in the inbox. Opening the item card to add a reference costs another minute, and the order is already done.

So the knowledge goes into the person's head and stays there. A few other things keep the table thin:

  • The ERP table matches a customer's part number well. It does nothing for a free-text description such as "kabelgoot 60x100 verz. 3m", which changes slightly every time the customer types it.
  • It has one row per code. A customer who writes the same product 4 different ways would need 4 rows, and nobody adds them.
  • Nobody owns it. Sales thinks master data owns it, and master data thinks sales does.
  • It goes stale. When you replace an article with a successor, the old references point to a blocked item and the next order fails.

The order history holds most of what the table lacks. Every sales order your desk created from a customer email is a solved example: this customer wrote these words and got this article. SIG Benelux found exactly this. Years of order history held the translations the sales desk had made by hand, so the matching could start from there instead of from rules someone had to write down first.

"I knew the technology made this possible. My biggest question was the quality of our data. It was a positive surprise to see the order history was enough of a base to get results this fast." Ditte Smit, Marketing Director, SIG Benelux

How do you match customer product descriptions to article numbers?

You match customer product descriptions by combining several methods in a fixed order: exact codes first, then the customer's own cross-reference, then the customer's order history, and only then a search on the text. Each method catches lines the others miss, and each one is wrong in a predictable way.

MethodWhat it matches wellWhere it goes wrongUse it for
Exact code lookupLines that carry your article number, an old number you still map, or a supplier code you holdFails on any line without a code, and on codes that changedThe first check on every line
Customer cross-reference in the ERPCustomer part numbers from purchasing systemsCovers codes only, and only the ones someone enteredLarge customers who order from their own system
GTIN or EANLines that quote a barcode number, common in retail and electricalOnly works when the customer quotes it and your item master holds it, per pack levelA strong signal when present
Fuzzy text matchingSmall spelling differences and word orderPicks the article that reads most alike, often the wrong size, colour or packA fallback that suggests options, never an automatic match
AI matching with contextFree-text descriptions, shorthand and reordersNeeds order history and a clear product master to reason fromEvery line the methods above cannot settle

Exact codes and the customer's cross-reference

Start with what can be looked up exactly. A line with your article number, an old number you still map to its successor, or a customer code from the cross-reference resolves without any judgement. Keep these lookups first, because they are cheap and almost never wrong when the data behind them is current.

GTIN and EAN codes

A GTIN is the barcode number GS1 assigns for trade items, which GS1 defines as products or services that are priced, ordered or invoiced at any point in the supply chain. According to GS1 Canada, each GTIN is unique to a single packaging level, so a pallet, a case and an inner pack each carry their own. That makes a GTIN a precise match when a customer quotes one. It also means your item master needs the GTIN for each pack you sell, or the case code finds nothing.

Why fuzzy text matching alone fails

Fuzzy matching scores how alike 2 strings are. "Kabelgoot 60x100 verz 3m" scores close to "kabelgoot 60x100 verz 2m" and to "kabelgoot 60x150 verz 3m", because they differ by one character. To a fuzzy algorithm those are near misses. To your customer they are the wrong product. Text similarity has no idea that the length and the width carry the meaning, or that this customer has ordered the 3 metre version 40 times and never the 2 metre one.

AI matching with the customer's context

An AI agent reads the description the way your desk does. It splits it into what it says about type, size, finish and pack, compares that with the attributes in your product master, and then checks the candidates against context:

  • what this customer has ordered before, and how they described it then;
  • the quantity, because a customer who always orders 10 of the 3 metre version is unlikely to want 1,000;
  • the price on the line, if the customer quoted one from your earlier quote;
  • the rest of the order, because a line of mounting brackets next to a cable tray narrows the tray down.

The order history carries the most weight, because it is specific to the customer. At SIG Benelux, 88% of order lines match the correct article first time, including lines that arrive without an article number, because years of order history do the translation.

How should confidence decide which lines a person checks?

Confidence should decide line by line. The agent scores how sure it is of each match. Above a line you set, the match goes through on its own. Below it, a person decides. Only you move that line, and you move it on evidence from your own orders.

A useful way to set it up:

  1. Start supervised. In the first weeks, every proposed match goes to a person, who confirms or corrects it. You learn where the agent is right and where it hesitates.
  2. Look at accuracy per band. If every match the agent scored very high was right over several hundred lines, those lines can pass without a check. Lines in the middle band stay with a person.
  3. Treat some lines as always checked. New customers, articles with safety or certification requirements, and lines above a value you choose can go to a person whatever the score.
  4. Leave doubtful lines blank. A wrong article that slips through costs a return, a credit note and a call. A line the agent leaves open costs a person a few seconds.

Xpol, a fresh flower wholesaler, built its order agent around that last rule. The agent prepares a draft that a person signs off, and ambiguous items are left blank for that person to fill. Each blank field also points to data that needs fixing, such as a customer code that needs cleaning up.

When the agent cannot place a line, it should ask the right person instead of waiting in a queue. In Lleverage, the agent can ask a colleague in Microsoft Teams or Slack which article a customer means and carry on with the order once it has the answer. At Topa Bathroom Products, an unrecognised article code is flagged for human review. Our guide to sales order automation with AI covers how the rest of the order moves into the ERP.

How each confirmation teaches the next order

When a person picks the right article for "lager 6205 2RS C3, 10 st" from Brakveld Techniek, that pair becomes a row in the customer mapping: this customer, these words, that article. The next time the same words arrive, the agent finds the row and matches with high confidence.

Keep that mapping in a structured table, one row per customer, description or code, and article, so anyone can read and correct it. In Lleverage, that mapping can live in a Table, which agents and workflows read and keep up to date. Changes to how the agent works, such as a new rule for one customer, go through review: the agent proposes the change, and a person on your team approves it before it goes live. It also pays to copy confirmed customer codes into the ERP's own cross-reference, so the item references in Business Central or the customer-material records in SAP fill up over time.

How do units and pack sizes fit into the match?

Units and pack sizes are part of the match, because the same article number with the wrong unit is still a wrong order. A customer who orders "5 doos" of screws has matched the right screw, but your ERP needs to know whether that is 5 boxes of 100 or 5 boxes of 200.

Treat the unit as a second question for every line:

  • What unit did the customer use? Box, roll, metre, pallet, piece, or nothing at all.
  • Which units does this article sell in? Your product master should list them with their conversion, such as 1 box is 200 pieces.
  • What has this customer ordered before? A customer who always orders cable by the 100 metre coil and writes "3" means 3 coils.

Business Central's item references can carry a unit of measure, so a customer code can point to "box of 200" directly. For free text, the agent works the unit out from the description, the quantity and the customer's history, and converts it before the line reaches the ERP. When the quantity looks wrong for the unit, such as 3,000 metres from a customer who orders 300, the line goes to a person.

Xpol writes this kind of knowledge down as customer rules. Its agent runs 25 customer rulesets, covering unit conversions, label text that depends on the weekday and splits across depots, which used to live in the heads of senior staff.

What does your product master need before you start?

Your product master needs to be consistent enough for the agent to tell articles apart. It does not need to be perfect, and a full clean-up project before you start can take months. Fix the fields that matching depends on and leave the rest.

Check these 6 things first:

  1. Blocked and replaced articles are marked. An agent that matches a customer to an article you stopped selling creates a sales order nobody can ship, so every replaced article should point to its successor.
  2. Descriptions follow one pattern per product family. "Kabelgoot 60x100 verzinkt 3m" and "KG 100x60 VZ L=3000" for 2 articles in the same family make both harder to find.
  3. Attributes sit in fields. Size, material, finish, colour and length in their own fields let the agent compare like with like. Attributes buried in the description still work, less reliably.
  4. Units and conversions are complete: every unit an article sells in, with its conversion to the base unit.
  5. GTINs are filled per pack level for the articles where customers quote them.
  6. Duplicates are merged. 2 article numbers for the same physical product split the order history in half.

The order history matters as much as the master. If your desk has typed orders into the ERP for years, that history already holds thousands of solved matches. For a wider look at keeping articles, customers and suppliers in sync across systems, see our guide to master data management software, and our master data management agents for the work itself.

How do you start with your top customers and most-ordered articles?

Start where the volume is. On many order desks, a small group of customers sends most of the order lines, and a small share of the range makes up most of what they order. Matching those well covers most of the work long before the long tail is done.

A practical sequence:

  1. Pull 6 to 12 months of sales order lines from the ERP, with customer, article, quantity and unit. Add the original order emails or PDFs where you still have them.
  2. Rank customers by order lines and pick the top 10 to 20.
  3. Build the first mapping from that history: for each of those customers, the descriptions and codes they used and the article they got. Where the original orders are gone, the sales order lines still show which articles each customer buys and in which units.
  4. Run the agent supervised on live orders from those customers. Every match goes to a person, who confirms or corrects it, and every confirmation adds to the mapping.
  5. Measure the first-time match rate per customer each week: the share of lines the agent got right without a correction.
  6. Move the confidence line for customers where the agent has been right for long enough, and add the next group of customers.

The first process typically runs supervised within a few weeks. SIG Benelux started with one location, one process and the central order mailbox, and ran its project with our partner SPAIK as a 4-week Kickstart.

Two things help early on. Include your hardest customers in the first group, because they show where the method breaks. And keep a short list of lines the agent got wrong, with the reason, because most errors trace back to a handful of master data problems you can fix once.

What changes for the order desk

The order desk spends its time checking drafts. A person opens a draft order with each line matched and the uncertain ones marked, each with the reason for the doubt. The email on one screen and the ERP on the other go away.

The knowledge also stops depending on who is in that day. When the mapping holds what the desk knows, a new colleague inherits it on their first morning, and a holiday no longer means slower orders for the customers only one person understood.

Koninklijke Dekker works this way in Business Central. Orders arrive as PDFs, Excel sheets and plain emails, and the agent checks products, availability and pricing before the order is created, with only low-confidence exceptions going to a person. For how the whole flow from inbox to sales order fits together, see our guide to order entry software for wholesalers, and the Intelligence layer of our platform for how agents use your history and mappings.

Frequently asked questions

What is a customer item cross-reference?

A customer item cross-reference is a table that links the item numbers and descriptions a customer uses to your own article numbers. Business Central calls it Item References, and SAP calls it customer-material information records. It lets an order with the customer's code resolve to your article without anyone looking it up by hand.

Can fuzzy matching handle customer product descriptions on its own?

Fuzzy matching alone is too risky for orders. It scores how alike 2 pieces of text are, so a 2 metre and a 3 metre version of the same product look almost identical to it. Use it to suggest candidates, then decide with the customer's order history, the quantity and the price before any line goes into the ERP.

Should we ask customers to order with our article numbers?

Asking helps with a few large customers who order from a purchasing system and can store your codes. For most customers it does not last: buyers change, old habits return and the long tail never adopts your codes. Map their own descriptions to your articles instead, so customers can keep ordering the way they already do.

How accurate can AI matching of customer descriptions be?

Accuracy depends on your order history and your product master. At SIG Benelux, 88% of order lines match the correct article first time, including lines without an article number, because years of order history do the translation. Measure the first-time match rate on a week of your own orders before you let any line through unchecked.

Do we need a PIM before we can automate product matching?

No new PIM is needed. Matching needs a product master that marks blocked and replaced articles, lists units with their conversions and keeps descriptions consistent within a product family, plus the order history already in your ERP. A customer-specific mapping next to the ERP does the rest, and it grows from every line a person confirms.

Start with one customer's last 6 months of orders

Pick the customer whose orders cost your desk the most time. Export their last 6 months of sales order lines, and collect a week of their original emails and PDFs. Compare the 2: every line where their words differ from your article description is a row your mapping needs.

Bring that sample to a demo and we will show you how our order agent matches those lines against your articles and that customer's history, which lines it would send to a person and why, and what the first supervised weeks on your ERP would look like.

Give your back office an AI workforce.