https://www.millabs.net/blog/adding-a-cui-tier/

Here is a pattern that repeats. An organization operates one or more classified environments and no environment at the controlled unclassified level. A great deal of the work does not need to be on the high side, but there is nowhere else to do it, so it goes there. Everyone knows a CUI tier is needed. Eventually one gets funded.

Then cost goes up, and nobody can explain why the obviously sensible move made things worse.

The failure mode is arithmetic

If you add a CUI environment and the classified side stays exactly what it was, you now operate three separately authorized boundaries instead of two. Whether they answer to one AO or three changes how hard each acceptance is to obtain. It does not change how many boundaries there are to assess, and that count is what drives the cost.

The new tier arrives with its own authorization program, its own assessment cycle, its own continuous monitoring, its own operations staffing, its own transfer procedures into and out of it, and its own recurring administrative load. If it is stood up as another on-premises enclave, it inherits nothing, exactly like the two you already have.

Meanwhile the high side lost nothing. Same seats, same images, same licenses, same operations team, same authorization burden, because no work actually moved down. The people who could have worked at the CUI level are still provisioned on the high side, because that is where the tooling and the data and the habit are.

So you have added a boundary to an estate that already could not staff the boundaries it had. This is not a subtle effect and it is not a modeling artifact. It is addition.

The version that works

The move that pays is not "add a tier." It is change what the high side is for.

The new tier takes over the work that never needed to be up there. The classified environment shrinks to carrying only the genuinely classified delta: the artifacts, analysis and data that are actually classified, and the seats of the people who actually need them.

Now the arithmetic runs the other way. The most expensive boundary you own gets smaller in every dimension that costs money. Fewer seats. Fewer images to maintain. Less storage. A smaller operations footprint. Less to assess. And the work that moved down lands somewhere with better options underneath it, where you can inherit controls from a platform instead of building every one yourself.

That is the difference between a tier that adds a boundary and a tier that redistributes an estate. The infrastructure looks identical. The outcome is opposite.

The precondition, which is the thing to check first

That whole play rests on one property, and it is not an infrastructure property.

Is the classified content in your artifacts actually separable from the unclassified content?

For some organizations the answer is yes and always was. Work is organized so that classified material is discrete, identifiable, and lives in specific places.

For others the answer is no. Classification is distributed through the artifacts. A model, a dataset or a document is classified in aggregate because classified detail is threaded through it, and pulling that apart is not a data migration, it is a re-engineering effort on the artifacts themselves.

This is an engineering question about how your work was built, not a question about your infrastructure, and it is the largest single risk in any plan of this kind. If separability does not hold, the high side cannot shrink, and adding a tier gives you the failure mode above no matter how well the tier is built.

Check it first, on real artifacts, with the people who produce them. It is a week of work and it decides whether the rest of the plan exists.

While you are choosing the delivery model

Two things worth knowing, because they are usually argued about in the wrong order.

A third on-premises enclave is the worst of the available options. It inherits nothing, and inheritance is invisible in a per-seat comparison. Compared on cost per seat it can look competitive. Compared on total cost to authorize and operate, it is not close.

Which hyperscaler is not the decision. Where a cloud premium applies to a government-wide vehicle, it tends to apply to the major providers alike, which makes it a poor discriminator between them. The delivery model - dedicated persistent desktops against a fleet sized to actual concurrency, managed hosting against self-operated - moves the answer far more than the provider does. Organizations routinely spend the decision cycle on the choice that matters least.


Next in this series: the single authorization everyone asks for, when you can actually have one, and what to build when you cannot.

Robert Burckner is the founder of Millabs Corporation. He has served as ISSM and ISSE for legacy weapon systems and as Division Chief at the Space Warfighting Analysis Center (USSF/NRO).

Share this post
Email LinkedIn
GET IN TOUCH WITH US

Find out what the gap between your product and an authorized environment actually looks like.