AI Governance Misses the Question That Fixed a 97.3% Delivery Rate

Delivery performance reached 97.3 percent in three months. No new system.

On Wernfeldt's rings, the author's answer to where that question belongs is nowhere. It is not data quality, not metadata, not ownership, not lineage, not a definition, not a policy, not a control, not an output you can defend. At that defense operation every ring could have been built to a high standard, and the technicians would still have been batching orders at four o'clock. Technician incentives are not a data governance problem, so nothing in the framework asks about them. If the organization has not established how it actually operates, the framework will govern beautifully, an operating model whose assumptions nobody ever checked.

His alternative sequence, built on a theory called Strand Commonality, runs observe, explain, validate, govern, automate. Automate last, after causal relationships are demonstrated rather than assumed. Most organizations run a different order: process, then data, then AI, then governance, with the first real check arriving after everything is built. Automate before establishing the causal relationships and you do not remove the problem, you encode it and run it faster.

That mattered less in 1998, when an incorrect assumption affected human decisions one at a time. Today the same assumption sits inside an autonomous system making thousands of decisions at machine speed, the liability question at the center of AI Rogue Agents: Who Bears Liability for Breaches. AI does not reduce the need for the first step, the author writes. It raises the cost of skipping it.

An old story does the closing work: a truck wedged under a low bridge. Engineers discuss cutting the bridge or calling in a crane. A boy asks why they do not let some air out of the tires. The boy knew less than the engineers; what he had was no investment in the frame everyone else was working inside. Air pressure is not on anyone's diagram. Neither is time of day.

His own framing, the map you design and the map you trace, makes the same point. Those rings govern the map you designed. Nothing in them checks whether the map you trace matches it. An elaborate framework gives an unvalidated understanding greater authority, he argues. The more rings you draw around an assumption, the less anyone questions it.

"Sometimes the most important question is the one the existing model gives you no reason to ask," the author writes.