Analysis identifies that utilization predicts margin. A reasonable executive concludes: raise utilization to raise margin. They raise it by cutting the bench, and delivery reliability collapses.
Nobody made an analytical error. The classification was never stated.
The three categories
Causal. The mechanism is understood, and ideally tested. Changing the input changes the outcome. You may act on it as a lever.
Correlational. The two move together. The mechanism is unknown, or runs the other way, or both are driven by something else. You may watch it. You may not pull it.
Predictive only. The variable improves the forecast and has no business meaning you would act on. Leave it inside the model.
The third category surprises people. A variable can be genuinely useful for prediction while being meaningless as a management signal — for example a data-entry timing pattern that correlates with project type without describing anything a manager could change.
Why the confusion is systematic
Predictive models are optimized to predict, not to explain. The variables they select are chosen for their statistical contribution, and nothing in the method distinguishes a cause from a coincidence that happens to be reliable.
So the output looks identical in all three cases: a list of drivers with weights. The classification has to be added by a human who understands the business, and if nobody adds it, readers will default to reading every driver as causal — because that is the only interpretation that suggests an action.
The reverse-causation trap
The utilization example above may run backwards. Low margin projects may require more hours, producing high utilization. In that case utilization is a symptom of the margin problem, not a cause of it, and raising it deliberately makes things worse while appearing to follow the evidence.
The direction of a relationship is rarely determinable from the correlation alone, and is often knowable from the business — which is why the classification is a conversation with operators rather than a statistical test.
The practical discipline
Every driver in an analysis gets one of three labels, stated in the readout, in writing.
For anything labeled causal, state the evidence: a mechanism, a natural experiment, a controlled comparison. "It is strongly correlated and it makes sense to me" is not evidence of causation — it is correlation plus a plausible story, which is the most persuasive and least reliable combination available.
The cost of skipping this is not an incorrect model. It is a correct model used to justify an action it never supported.
Practical check
For every driver presented to management, add two columns: classification and permitted use. A causal driver may support an intervention. A correlational driver supports monitoring and further investigation. A predictive-only variable stays inside the model. This small table prevents a statistically useful signal from being turned into an operational instruction it never justified in practice.
Predictive Intelligence engagements classify every identified driver as causal, correlational, or predictive-only, and state which of the three the evidence supports.