A business intelligence framework with governed, curated, and sandbox tiers

Most arguments about data inside a company are really arguments about status. Is this number official? Am I allowed to use it in a board deck? Can I build my own version? Who has to be consulted before it changes?

A framework that answers those questions is worth more than one that describes components. This page sets out a three tier model of publication: what each tier promises, how something moves up, and how something gets pushed back down.

What to take away

The three tiers

The tiers are not a quality ranking of the people who build in them. A sandbox analysis by a skilled analyst can be better work than a governed metric. The tier describes what the organization is willing to stand behind, which is a different property.

Why three and not two

Two tiers, official and everything else, is the arrangement most companies drift into, and it fails in a specific way. The official tier is slow to enter, so real work happens outside it, and everything outside it is undifferentiated. A number from a careful analyst and a number somebody made up in a spreadsheet look identical to a reader.

The middle tier exists to make that distinction visible. Curated means: built on defined concepts, by someone who knows the model, and reviewed by a second person, but not promised to be stable or supported forever. That is an honest description of most useful analytical work, and it gives readers a category they can calibrate against.

What each tier actually requires

Governed. A written definition, a named owner, tests on the underlying table, a change process that announces before it lands, and a stated position on whether history gets restated when the definition changes. If any of those is missing, it is not governed, whatever it is labeled.

Curated. Reads its core concepts from governed definitions rather than recomputing them. Reviewed by one other person who knows the data. Carries the name of a builder and a date. Can break without an incident being declared.

Sandbox. One rule only: it must be visibly marked as sandbox wherever it appears. The entire value of the tier is that it is fast, so do not add process to it. Add labeling.

Promotion: the gate that keeps the top tier honest

Something moves from sandbox to curated when it is going to be used more than once, and from curated to governed when a decision with real consequences will rest on it.

The promotion request should be short and should answer four things: which recurring decision this serves, who will own it, what it excludes, and what would have to break for someone to notice. If the requester cannot answer the fourth, the artifact is not ready, because nobody has thought about how it fails.

Promotion has to have a cost, and the cost is the owner's time in perpetuity. A governed metric is a standing commitment to maintain a definition, answer questions about it, and manage its changes. Charging that cost honestly is what stops the governed tier from swelling until it means nothing.

Demotion: the part nobody builds

Artifacts should move down as well as up, and there should be named triggers.

Demotion is not deletion. Moving a governed metric to curated tells readers it is still there and no longer guaranteed, which is far more useful than leaving a stale promise in place.

Who decides

Keep this small. One person from the data function and one from the business area, meeting briefly on a regular cadence, can run promotion and demotion for a company of considerable size. A large committee turns the framework into the approval process it was designed to avoid.

Publish two lists: what is governed, with owners, and what was demoted recently, with the reason. The second list does more for credibility than the first, because it proves the tiers mean something.

Making the tier visible to readers

A framework nobody can see does not change behavior. The tier must appear where the number appears: on the page, in the export, in the file name.

Treat it as metadata that travels with the figure, not a policy to remember. A reader deciding whether to put a figure in a customer-facing document needs its status without asking anyone.

Two practical rules: sandbox artifacts should be visually distinct enough that nobody accidentally screenshots one into a deck. They must differ by more than color alone, since color is not a reliable sole signal for every reader.

Any export from a curated or sandbox artifact should carry its tier in the file; the export is where context is normally lost.

Starting from where you are

Related reading on this site

The function that runs this, and its request economics, are covered in business intelligence, while definition ownership and change control are covered in data governance. Several tier problems first appear as the shapes in recurring patterns in intelligence work.

For how this fits a project's build order, see the six layer model, and for tooling implications, see analytics tools.

Common questions

Is this the same as a data catalog?

No. A catalog describes what exists. This assigns a promise to each thing and a process for changing it. A catalog with no tiers lists a governed metric and somebody's draft next to each other.

What if a business team refuses to use the tiers?

They will use them the first time a sandbox number is challenged in a meeting and the label protects the person who produced it. Lead with that benefit rather than with compliance.

How many governed metrics should there be?

Few enough that the owners can be named from memory. If the list needs a search box, the tier has stopped meaning anything.

Does the middle tier not just become a dumping ground?

Only without the demotion triggers. Review the curated list on a fixed cadence and remove anything with no reader, and it stays a working shelf rather than an archive.