When Cleaning Up Means Betting Blind

Declaring a clean target architecture is not a cost-only cleanup — it is a blind decision that risks destroying value you cannot see

The first three articles built to an uncomfortable place. Projects fail for missing information about how the work runs (one); much of that information was never decided and must be declared, an act that changes the process (two); and those declarations are entangled and costly, with the bill falling in a paradoxical order — you must decide before you can discover (three).

There is an obvious way out. Then stop discovering piece by piece. Declare the whole target architecture once, design the best process from the start, and rebuild toward it. The evidence even seems to encourage it: the AI deployments that succeed are overwhelmingly those sitting on a coherent, well-architected foundation. So architect everything up front, and be done.

This article is about why that move is not the clean escape it appears to be. The rebuild has a cost — that was the third article. What this one adds is that it also has a risk, of a kind you cannot price in advance, because it lives in exactly the part of the estate you cannot see.

You would be designing blind

To design the best possible target, you need to know your current process. But you don’t — that was the whole argument. The most ordinary symptom is scale: organizations that believe they run two hundred applications discover six hundred, and counting. If you cannot say how many systems you have, you cannot say how complex your process really is, and you cannot claim to be designing the best version of something you have not finished finding.

This is not ordinary under-information that more effort dissolves. It is structural: fully resolving the current state would require the very declarations the target was supposed to settle. The design and the discovery it depends on are tangled into each other, so the target is chosen under irreducible uncertainty about the thing it is meant to replace.

The risk: you cannot tell the valuable mess from the worthless mess

Here is what turns cost into risk.

It is tempting to assume that whatever you don’t know about your estate is junk — redundant systems, dead weight, accidental sprawl. Much of it is. But not all of it, and that is the trap: you cannot tell which undocumented parts are worthless and which are load-bearing, precisely because they are undocumented. Both look identical from the outside — unmapped, unexplained, still there for some reason. So when you re-architect through them toward a clean target, you are not removing known waste. You are spending value and waste together, at a rate you cannot measure.

And some of what is hidden there is not waste at all. The undocumented corners of an organization are where its hard-won fit to its own reality accumulates: the special-case handling, the workaround that quietly saved a deal, the step that works for a reason no one wrote down. This danger is well charted. It is the logic of Chesterton’s fence — do not remove what you do not yet understand the purpose of. It is what James C. Scott, in Seeing Like a State, documented repeatedly: imposing a clean, legible design destroys the illegible local know-how that was actually holding the system up. It is why seasoned engineers fear the great rewrite — the old system is also an undocumented record of every problem it learned to survive.

So rebuilding blind is not merely expensive. It carries a downside you cannot bound in advance: the chance of deleting the thing that was working, without ever knowing it was there.

With AI, the risk has a sharper edge

There is a further turn when the rebuild aims at an AI-ready, standardized target. Standardizing a process toward a common template tends to erase the very idiosyncrasies that differentiated it — and a good share of an organization’s real edge lives in those idiosyncrasies, in the undocumented last mile. A generic, best-practice process is, by definition, one that everyone else can also run.

This is the old lesson that operational best practice is not strategy: when everyone adopts the same template, competition converges and the advantage erodes (Porter), and organizations drift toward sameness over time under ordinary institutional pressure (DiMaggio and Powell). AI sharpens this because it accelerates the move to shared templates. The qualification matters, and we make it plainly: AI built on genuinely proprietary data and process can sustain differentiation; it is standardization that erodes it, and AI erodes differentiation only to the degree it is adopted as a common template. We hold this as a conjecture, not a law — but the direction is clear enough to weigh.

The dilemma this leaves

Set beside the third article, this closes a trap rather than opening a door. The cautious move — incremental rationalization — is safe, but it reaches mostly what was already dead; it can retire a live system where the information happens to be sufficient, but it leaves the deepest entanglement untouched. The decisive move — declaring a clean target and rebuilding — could touch that entanglement, but only by betting blind on which unseen parts were waste and which were quietly carrying value. The careful path does not fix the problem; the path that could fix it is not safe.

None of this argues for never rebuilding, or for treating disorder as a virtue. It argues for treating re-architecting as what it is: a risk-bearing investment made under uncertainty, to be sized against the value it puts at risk — with the cruelty that the value at risk is the very thing the disorder has hidden from view.

Put to the field

Part of this claim rests on practice rather than published study, so we put it to the practitioners who would recognize it. Have you watched a clean target architecture, faithfully delivered, quietly erase something that used to work — an edge case, a relationship, a capability no one had documented and no one missed until it was gone? Where this matches what you have seen, say so; where it doesn’t, say that too. This is the kind of risk that shows itself only in the field, and that is where it has to be judged.

This describes a present risk in the act of rebuilding, not a trend over time; the one temporal element — competitive convergence — is a separate question, and the most tentative part of the claim.

Sources and influences

• Illegible local knowledge destroyed by imposed order — Scott, J. C., Seeing Like a State (1998); Polanyi, M., The Tacit Dimension (1966); Hayek, F. A., «The Use of Knowledge in Society» (1945).
• Do not remove what you do not understand — Chesterton’s fence (G. K. Chesterton, 1929).
• Best practice is not strategy; competitive and institutional convergence — Porter, M., «What Is Strategy?» (Harvard Business Review, 1996); DiMaggio, P. & Powell, W., «The Iron Cage Revisited» (American Sociological Review, 1983).