The Polanyi Problem

That was the observation made by Michael Polanyi. It sounds abstract. It is not.

It shows up every time a decision feels wrong, even when the numbers look right.

You can ride a bicycle. You cannot fully explain how you balance. You can recognise a face instantly. You cannot list the variables that led you there. You can look at a plan, a model, a set of outputs and say: something is off. And you are often right.

Most systems are built on the opposite assumption.

That knowledge is explicit. That it can be written down. That it can be measured, modelled, and optimised. That if something matters, it will show up in the data.

This is where the gap opens.

Models work with what can be defined. Decisions depend on what cannot. The more precise the system becomes, the more it excludes. Not because it is wrong. Because it is incomplete.

In marketing, this is easy to miss.

The model runs. The outputs stabilise. The ROI lines up. And yet the plan drifts. Not dramatically. Gradually. Toward what is easy to capture. Toward what responds quickly. Toward what the system can see.

This is how you end up optimising the wrong thing, very efficiently.

The cheapest media looks best. The most attributable channels expand. The plan starts to reflect the measurement system, not the market.

Nothing is broken. But the outcome is.

This is not a modelling flaw. It is a structural one.

The model is doing exactly what it is designed to do. It is the decision framework around it that narrows.

Tacit knowledge does not disappear. It just stops being heard.

You see it in small moments.

When someone hesitates before approving a plan. When a result is accepted, but not believed. When experience contradicts output, and gets overruled.

The instinct is to fix the model. More data. Better variables. More granularity.

But the missing piece is not more information. It is knowledge that does not translate cleanly into variables.

The work is not to eliminate that gap. It is to recognise it.

To treat models as inputs, not answers. To create space for judgement, not override it.

To ask, consistently: What does this actually change?

Because the risk is not that the model is wrong. The risk is that it is right, and the decision is still off.

This is where human analysis comes in. Not as a check on the model. Not as a fallback when it fails. As the bridge.

Between what can be measured and what can be known. Between output and judgement. Between the system's answer and the right one.

It is the question the model cannot ask about itself: what happens if we keep making this move?

Both are necessary. Neither is sufficient alone.

That is not a limitation to manage. It is how good decisions actually get made.