One lesson keeps reappearing in complex organisations.
When something goes wrong, the approval is rarely where the decision began.
Start there. Then work backwards.
You begin with the person who approved the outcome. Their name is on the record. They approved a recommendation. The recommendation came from the system. The system ranked the available options and surfaced the highest. The ranking followed a model. The model was optimising for an objective someone set, based on a problem statement someone wrote, to reflect a business intent someone approved when commissioning the build.
That person was senior enough to authorise the system.
They are rarely visible in the governance record of any decision it went on to make.
Approval.
Recommendation.
Ranking.
Objective.
Training data.
Problem definition.
Business intent.
At some point in this investigation, you stop finding people.
You find a parameter.
An objective function documented in a system design from eighteen months ago. A confidence threshold that nobody currently employed remembers setting. A data inclusion decision made during a sprint that finished before most of the governance committee had heard of the project.
On one occasion, I found a comment in a code file.
"Not sure about this. Will revisit."
Nobody had.
Authority did not disappear from these decisions. It became distributed across a decision architecture that almost nobody sees as a whole.
The person who approved the outcome held authority in a formal sense.
The people who built the system held authority in a consequential sense.
They were not the same people.
They were rarely in the same meeting.
The formally authorised person saw the last moment of a decision journey that had been quietly assembling itself for months, sometimes years, before they were asked to approve it.
The objective function was written before the governance framework existed.
The training data was assembled before the use case was approved.
The model was deployed before the risk register included it.
This is not a story about negligent organisations.
Most of the organisations I have observed were trying to do this well.
The engineers were diligent. The governance teams were engaged. The executives were informed.
What none of them had was a view of the decision architecture as a complete thing.
Each group saw their part of it.
Nobody saw all of it.
The question "who made this decision?" presupposes that decision making happened in one place, at one time, by one authority.
In AI enabled organisations, that presupposition is increasingly wrong.
I wonder whether that is the question we should be asking at all.
Or whether the more useful question is simply this.
When did the decision you approved this morning actually begin?
Question for the Field
Choose a consequential AI supported decision your organisation made in the last year.
Start at the approval. Work backwards.
Recommendation. Ranking. Objective. Training data. Problem definition. Business intent.
Follow it until the trail becomes unclear. Until you stop finding authorised decision makers and start finding system parameters, design documents, and assumptions nobody currently employed remembers making.
That point is worth knowing.
That is probably where your investigation really starts.