How to build a data-driven company

Same month, three numbers. Sales reports that June was strong, finance that it was average, operations that it was a record. All three come “from the system”, and all three people are right — each queried their own source, their own cut, their own definition of revenue. The meeting meant to decide the next quarter spends its first twenty minutes deciding whose number is correct.

The damage is not the wrong number. A wrong number gets corrected. The damage is what the company learns that morning: that “the system” is not an authority but one party in a negotiation. From that day everyone keeps a copy — sales its spreadsheet, finance its report, operations its board — because each has seen the numbers disagree once and is now defending itself. And every new copy guarantees the next disagreement.

A data-driven company is usually sold as a tool: a BI platform, a board of indicators, a report that refreshes itself. But a data culture is not software you buy. It is a habit — that a number in the system is reliable enough that nobody checks it again in a spreadsheet. This piece is about how that habit is built and what makes it impossible, because it rarely fails on technology and usually fails on the first disagreement left unexplained.

One source of truth — and why it is hard to have

A single source of truth is not a single database. It is one place where a fact is authored, and every view is derived from it — never copied. A trial balance is not a copy of the ledger; it is the ledger seen from one angle. A customer statement is not a transcript of invoices; it is those invoices, summed. The test of a culture is simple: when a number surprises the person seeing it, what do they open first — the source or a blank spreadsheet? Which reflex wins is the culture, regardless of how many boards the company owns.

A copy is where divergence begins

The moment a number is retyped into a spreadsheet or piped into a second system, two truths exist. Today they agree. They part the first time someone changes the original and forgets the copy — a reversed invoice, a corrected discount, a backdated credit note. From then on, two reports “from the system” are both from the system, only from different copies of a moment long gone. Nobody is lying; each is reading a different photograph. So the list of places a copy quietly appears is worth more than any board:

• An Excel export that is then edited by hand and forwarded — from that email on, it is a separate, dead datum.
• A second reporting database refreshed overnight — correct until midnight, wrong for every change made during the day.
• A metric typed into a slide — a figure with no source, alive exactly as long as the file.
• A parallel commission spreadsheet kept by sales, because “the system is wrong anyway”.
• A hand-maintained customer list beside the one in the system, diverged before anyone noticed.

Three properties that make a number a datum

A number a culture may use for a decision differs from a nice-looking figure by three properties. Each is a mechanism with its own depth; here they matter as properties of the datum, not as engineering. If any one is missing, the number can still sit on a board — it simply cannot survive a question.

The datum knows whose it is

The account says what kind of value arose; the dimension says whose it is — which project, which cost centre, which channel. Without that field, “profitability by project” is not a report waiting to be requested but a wish: the datum simply has no shape a project could enter. That is why the dimension is chosen before the first entry rather than added when someone asks — because history has nothing to be filled from.

The datum knows when it was valid

A rule is remembered as it applied then. The VAT rate applied to a 2012 supply is the 2012 rate, and the VAT review for that period, regenerated today, must return the number that was filed — otherwise the report quietly rewrites the past. The same holds for contribution bases, exchange rates, the non-taxable salary amount. A datum that does not remember when it was valid has no past, and accounting operates exclusively in the past tense.

The datum knows where it came from

Every record answers who created it: a human, a deterministic rule, an accepted automation draft. Without that provenance, written at creation, a number is an assertion — as reliable as the memory of whoever defends it. Provenance is what turns a disagreement from a fight into a check: the question stops being “who is right” and becomes “which record was created how”.

Why three disagreeing reports kill trust

Trust in a number is asymmetric: slow to build, broken by a single disagreement. One meeting where two “identical” numbers differ is enough for the whole company to learn the lesson — the system is not the last word. And that lesson is self-fulfilling: the moment people stop trusting the source, they start keeping copies, and copies guarantee the next disagreement. Distrust feeds on itself, and it is the one part of this whole story that genuinely scales.

The cost is paid in meetings. A decision about price, stock or hiring turns into reconciling a number before the conversation about the decision even begins. The most expensive people in the company spend the first twenty minutes agreeing what revenue was, instead of what to do with it. And almost none of these disagreements is an arithmetic error — here is where they actually come from:

Every row is a difference of definition, cut or basis — and that is exactly why it is dangerous: both numbers are “correct”, so the dispute is not settled by checking but by an agreement that should have been made before the first report. A data culture is, to a large degree, the discipline of writing those definitions down once and versioning them, instead of renegotiating them each time in a room where everyone is defending a copy.

Decisions from data, not from a feeling

The goal is not to remove judgment from a decision. A director's sense that a customer has turned risky is valuable — as a hypothesis. A data culture asks only that the hypothesis be checked against a number with a denominator, a dimension, a time basis and a provenance. A feeling that survives the test becomes a decision; a feeling that does not is an opinion wearing a percentage. The line is not whether intuition is allowed, but whether it may be cashed without a check.

There is a quieter benefit. When numbers are trusted, an error is sought in the data, not in a person. “Margin dropped” stops being an accusation and becomes a question: which line, which customer, which month. Blame moves from people to definitions, and a definition is fixed without a meeting where someone defends themselves. A company that can be wrong out loud gets out of being wrong faster — because admitting a mistake costs nobody their face.

How the habit is built

1. 1

   One number, one definition

   “Revenue” means one thing, written down and versioned, not whatever each person assumed. A definition agreed after the first disagreement is honoured by nobody; agreed in advance, it is the only reason two reports can be compared at all.

2. 2

   Every number leads to a list

   A figure that cannot be clicked is a slogan. Behind every indicator there must be the list of records it was derived from — not because anyone will always open it, but because being able to open it changes the relationship to the number.

3. 3

   Kill the second copy by making the source better

   A parallel spreadsheet is born not of spite but of distrust or an awkward source. A ban does not help; what helps is a source that is correct, fast, and that stamps the definition and the cut moment onto every export, so even a dead datum knows when it died.

4. 4

   A disagreement is a bug, not a debate

   When two numbers differ, that is a defect traced to its cause and logged, not a debate somebody wins. A company that investigates every disagreement once stops having them; a company that negotiates them has them forever.

5. 5

   The same number is seen by everyone

   An indicator seen only by management becomes a slide. The same number must reach the person doing the work behind it — otherwise nobody reports that the figure is pretty and the reality is not.

How NG One approaches it

NG One is built around a single source of truth, not around a board added at the end. Every view — trial balance, customer statement, stock on hand, an indicator on the home screen — is derived from the same posted lines and documents rather than copied into a separate reporting layer. The three properties from this piece sit in the core: the dimension on the line, the parameter with a validity range, the provenance written at creation. On the home screen a figure is stated in language the user understands — “since the last sign-in, 83% of the bank statement was matched”, for instance — and every such sentence drills through to the records behind it. Exports carry the definition and the cut moment, because a datum that leaves the system should at least know when it stopped being alive.

A data-driven company is not one with more dashboards. It is one where a number is not negotiated — where a meeting starts from a shared figure and spends its time on the decision, not on whose spreadsheet is less wrong. That difference is a habit, and software can only make it cheap or impossible. Software that derives every view from the same source makes it cheap; software that permits five copies of the same number makes it, quietly, impossible.