Monthly review, second Tuesday of the month. Sales reports 42 million in revenue, controlling has 39 on the slide. For the next half hour the room debates whose number is right instead of which decision follows. Half an hour every month, eight executives at the table: 48 hours of top management per year in which nobody decides anything.
And the monthly review is only the visible symptom. “Is customer X actually profitable?” takes weeks in many companies, because someone has to assemble the answer by hand. And the new AI assistant that was supposed to speed exactly that up delivers a third number to the same question. Fluently worded, without hesitation, without provenance.
Yet nobody is bad at maths. The fight over the correct number is almost never a calculation error. It is a translation error. Your company speaks five languages, it just hasn’t noticed yet.
False friends: one word, three numbers
Linguists call these false friends: same word, different meaning. The English billion is a thousand million, the German Billion a thousand times more. Your company has the same problem, but doesn’t notice, because everyone believes they speak the same language. Sales, logistics, production, controlling, accounting: on paper, one language. In reality, every department speaks its own dialect, and the false friends are called “customer”, “product” and “revenue”.
Take the “customer”. For sales, that’s the company placing the order. For logistics, the address the goods are shipped to. For accounting, the group that ultimately pays the invoice. Three legitimate views, three customer numbers, three data sets. Same game with “revenue”: order intake here, invoiced revenue there, with or without subsidiaries. 42 and 39 are both right, in their respective languages.
The root sits in the most inconspicuous data of all, master data: customers, products, suppliers, cost centres. Every department keeps its own private dictionary, and nobody has ever compared the entries. The moment a question crosses departmental borders, the keys don’t fit: if the same customer carries two numbers in two systems, there is nothing to join. So the mapping table is born in Excel, maintained by the one colleague who happens to speak both dialects. Interpreting by hand. Hence the weeks.
And the AI assistant? It doesn’t get an interpreter’s job, it gets a guessing job. The AI learns all five dialects at once and builds a sixth out of them, one that sounds like each of them and belongs to none. It speaks it fluently, but nobody can look it up, because it is in no dictionary. Five quiet ambiguities become one confidently delivered one.
A dictionary, not a world language
The classic reflex is a single language: one ERP for everyone, one numbering world, a multi-year harmonisation programme. Call it the Esperanto approach. It sounds logical and fails regularly, because the dialects exist for good reasons: logistics genuinely needs the delivery address, accounting genuinely needs the invoice recipient. Impose one language on everyone and you don’t get consensus, you get shadow Excel.
International teams do it more cleverly. Nobody forces the colleagues in Lyon to learn German. You agree on a shared glossary for the terms that concern everyone, and internally each keeps speaking as before. That is exactly what good master data management is: not a systems project, but a dictionary with owners. A manageable number of shared terms – customer, product, supplier, organisational unit –, one binding definition, one owner and one leading source per term, plus translation tables into the dialects of the individual systems.
For that, a dictionary needs three things: someone who decides which words go in and who writes them; a place where it lives and everyone can look things up; and an operation in which translation actually happens. In your company the three are called data and AI strategy, data catalogue and data platform. None of them can do the others’ job.
The strategy decides which words go into the book first
No company clarifies all its terms at once, and whoever tries ends up back at Esperanto. The data and AI strategy provides the order, and that order comes from the use cases: profitability accounting needs products and cost centres, the churn model needs the customer hierarchy, the invoice-checking agent needs suppliers. Each first use case pays for each first chapter. In practice: when you prioritise use cases, there is a third column next to value and effort, which terms the use case presupposes. That column is your dictionary plan.
It also decides which dialects need to exist at all. That logistics needs the delivery address is a good reason for a view of its own. That two plants carry the same product under two numbers because somebody set it up that way twenty years ago is not. Such dialects get dismantled: one system fewer, one numbering world fewer, one dictionary entry fewer.
And it settles who writes: one owner per term, from the business, who decides the definition and settles disputes, not IT and not the data team, with a mandate from top management, because the agreement always hurts one of the parties. Business leads, technology enables: with master data this holds more literally than anywhere else.
The data catalogue is the dictionary itself
The definition the owner and the departments have agreed on has to live somewhere everyone can find it. Otherwise it lives in the head of the colleague with the Excel sheet, and it retires with her. That place is the data catalogue. This is where it says what a “customer” means, which system is the leading source, who the owner is, what the term is called in each individual system and in which reports and models it is used. That is metadata: not the data, but the description of the data.
It sounds like admin, but it is the difference between a glossary that exists and one that gets used. The catalogue shortens the monthly review because every figure points to its definition: “revenue” on the slide means invoiced revenue excluding subsidiaries, it says so right there, discussion over. And it is the point where the AI assistant stops guessing. An agent that knows “customer” means the ordering company in the sales system and the invoice recipient in accounting can translate between the two. One that doesn’t pretends they are the same thing.
The data platform is the interpreter’s booth
That leaves the operation: somewhere the translation has to actually happen, every day, for every report and every agent. That is the job of the data platform. This is where the translation tables run as pipelines instead of Excel: the customer from CRM, from logistics and from accounting is mapped to the one entry in the dictionary, once, checked, for everyone. The colleague’s private mapping table becomes a maintained reference table with quality rules on which everything downstream depends.
This is where AI helps in the other direction. Spotting duplicates, matching spellings and proposing hierarchies is something it now does far better than any rule engine. There is just one thing it cannot do: decide what a customer is in your company. Hence the order: a platform without a dictionary is a very expensive room in which five dialects sit side by side. Only when strategy, catalogue and platform work together does “revenue with customer X” have exactly one answer, and it arrives in seconds rather than weeks.
And now the asterisks
First: a tool is not yet a dictionary, and a dictionary is never finished. MDM software and catalogue tooling help, but they are the book cover, not the content. And the content ages: a new product, an acquisition, a new sales region, and three definitions no longer hold. Catalogues rarely die at the start; they die in year three, when ownership meant “decide” but not “maintain”. And because no board celebrates a cleaned-up article number: work in the slipstream of value-creating use cases, chapter by chapter, never as an end in itself.
Second: measure the right thing. What counts is not the number of cleaned records, but whether the symptoms disappear. Does the monthly review still start with reconciling numbers? Does the answer to “Is customer X profitable?” still take weeks, or hours by now? Those are incorruptible metrics, and they are pleasantly untechnical.
Your monthly review negotiates million-euro questions in five dialects every month, and its glossary is empty. The good news: you don’t need a world language or a mammoth project. You need a strategy that says which terms come first, a catalogue in which they are written down, and a platform that translates them every single day. Then the review starts with what follows from the 42 million, not with whether it is 42. And your AI no longer speaks a sixth dialect, but the one that is in the dictionary.
First the terms. Then the numbers.