Data Visualization Services Without Four Hundred Truths
Author : James Smith | Published On : 02 Nov 2026
Self service analytics has one well known failure mode. You give three hundred people the ability to build charts, and within a year you have three hundred versions of revenue.
We went through that and came out of it. Here is how we now structure Data Visualization Services so self service works instead of multiplying.
The problem is not the tool
It is what the tool is pointed at.
Our first attempt gave people a capable platform connected directly to source systems. Each analyst then wrote their own logic for margin, for active customer, and for region.
All of them were competent. None of them agreed, because the source systems do not contain those concepts. They contain fields that have to be interpreted, and everybody interpreted them slightly differently.
Build the layer people build on
The single most important investment.
Between the source systems and the charts there is now a modelled layer. Margin exists in it once. Active customer exists in it once. Region exists in it once.
Analysts build freely on top of that. They cannot build their own margin, because margin is not a calculation available to them. It is a column.
An AWS data warehouse modernization is what gave us that layer, with the definitions applied centrally, documented, and permission aware. Self service before that layer existed was the thing producing the four hundred versions.
Certify a small set and leave the rest alone
Two tiers, clearly marked.
Certified content is owned, tested, and safe to quote outside the team. Everything else is explicitly marked as personal exploration and carries a visible label saying so.
We have forty one certified items and several hundred personal ones. Nobody minds the personal ones, because nobody mistakes them for official figures.
Our previous estate had no distinction, so every chart carried the same implied authority as every other chart.
Make the certified version easier than building your own
The behavioural lever.
People rebuilt things because finding the existing version was harder than starting fresh. Search was poor, naming was inconsistent, and half the catalogue was abandoned work.
We cleaned the catalogue, adopted a naming convention, and put the forty one certified items on a single page. Rebuild requests dropped sharply, because the alternative became genuinely quicker.
Good Data Visualization Services work includes the catalogue, not only the charts.
Teach the three mistakes
Not a full statistics course.
Percentages on small bases. Comparing a partial period against a complete one. And reading a rising absolute number as a rising share.
Those three accounted for most of the wrong conclusions we saw. A ninety minute session covering them, run twice a year, was the highest return training we have done.
Let people ask for promotion
The route from personal to certified.
An analyst who has built something useful can request certification. We review the logic, the definitions, and whether anything similar already exists.
About one in six requests gets certified. The rest usually turn out to duplicate something, which is a useful conversation rather than a rejection.
Keep permissions in one place
The failure that is harder to recover from.
Our first attempt had permissions set per dashboard, by whoever built it. Several contained figures their audience should not have seen, and nobody noticed for months.
Permissions now live in the modelled layer and are inherited. A chart cannot show somebody data their role does not allow, regardless of who built the chart or how.
Be realistic about generated charts
A question that now arrives weekly.
Several AI Tools For Business will build a chart from a question typed in plain language, and they do that competently.
Over our modelled layer that is useful, because the definitions are already correct. Over raw source systems it is the four hundred versions problem with a faster engine, which is why we did not deploy it until the layer existed.
Review usage, not creation
The metric to watch.
Counting how many items people build measures enthusiasm. Counting what gets opened, and by whom, measures value.
Our review archives anything uncertified and unopened for six months. Nobody has objected, which tells you what those items were worth.
What our arrangement looks like
- A modelled layer holding every shared definition.
- Two tiers, certified and personal, visibly labelled.
- A clean catalogue with a naming convention.
- A promotion route from personal to certified.
- Permissions inherited from the layer, never set per chart.
- Training on the three common misreadings.
- Archiving based on usage rather than age.
Common questions
Does self service cause the duplication problem?
No. Self service over unmodelled data causes it. The same people building on a modelled layer do not produce conflicting numbers.
How many certified items should exist?
Few. Ours is forty one, and the discipline is in keeping that number down rather than growing it.
Who certifies?
A small analytics group with the relevant business owner. Data Visualization Services partners can prepare the case and should not hold the final say.
What about tools that generate charts automatically?
Useful on a modelled layer. Before that they accelerate the problem rather than solving it.
Should the platform choice come first?
No. Agree the modelled layer and the definitions, then pick. Data Visualization Services selected before that happens get configured around whatever the source systems happen to contain.
Does this apply to assistants as well?
Yes. The AI Tools For Business we deployed over the modelled layer work well, and the ones pointed at source systems produce the same conflicting answers our analysts used to.
How long did the layer take?
About seven months, and it was the precondition for everything else. Data Visualization Services work before it existed was effort spent on charts we later rebuilt.
How do you stop people exporting to spreadsheets?
Mostly by making the charts answer the question. Exports usually mean the chart was close but not quite right.
Where we are now
Forty one certified items, a modelled layer with the definitions in it once, inherited permissions, and several hundred personal charts nobody confuses with official numbers.
The platform never changed. We changed what it was pointed at, and the disagreements about whose revenue figure was correct stopped within a quarter.
