When the Question is Not “What Will Happen?” but “What Should We Do?”

From Data to Decision

In a bank, the decision may seem straightforward: should we approve a loan application or reject it? In logistics, the question is different. If demand is expected to increase in a particular area, where should we position our vehicles? Do we need to expand the fleet at all? In ecommerece, the decision might be whether to lower prices, increase marketing spending, or simply leave things as they are.

In risk management, the challenge goes beyond measuring the size of a risk. The real question is whether it is worth taking action in the first place, and whether the cost of that action is actually lower than the potential loss. The same logic applies to insurance, clinical research, and even environmental studies. If, for example, artificial lighting is introduced into a natural area, the ecosystem will change in some way. The question is how, and when.

These fields may seem entirely different, yet the underlying problem appears again and again. We have data. We have models. We may even be able to make reasonably good predictions about the future. But there is still something that data alone cannot answer: what should we do with this information?

Data can tell us what happened. Models can sometimes help us estimate what might happen next. But the actual decision usually begins somewhere else: with the alternatives available to us.

Prediction Is Not a Decision

Suppose a bank builds a model that predicts the probability of a customer defaulting. That is undoubtedly useful information, but it does not tell the bank what to do with it.

Should the application be rejected? Should the loan be approved at a different interest rate? Should the customer be offered a smaller amount? Or should additional guarantees be required?

The model may be excellent at prediction, but it does not choose between these alternatives. In fact, that is not what it was designed to do.

The same applies to logistics. If a model tells us that demand in a particular city is expected to increase by 15% next month, the decision does not suddenly become obvious. We still need to know how many vehicles will be required, where exactly they should be deployed, how much it will cost to move them, and what happens if the expected growth does not materialise.

Marketing presents a similar challenge. We may observe that spending on a particular channel is associated with higher sales, but that does not necessarily mean increasing spending will produce a proportional increase in sales. The relationship we observe may be causal, or it may not be. Conditions may change.

These are precisely the questions that make the transition from data analysis to decision-making far more complicated than it first appears.

The Effect of the Decision Itself

There is an important difference between observing that two things are related and knowing that one actually causes the other.

If sales increase after a marketing campaign, the campaign may have caused the increase. But it could also be due to seasonality, a change in market conditions, or the fact that the campaign targeted customers who were already more likely to make a purchase.

This is exactly the kind of problem that causal inference attempts to address.

Rather than asking, “Which factors are associated with sales?”, we ask a more precise question: What would sales have looked like if the campaign had never been launched?

In healthcare, the equivalent question is whether the improvement we observe was genuinely caused by the treatment. In environmental research, we might ask whether a reduction in pollution resulted from a new regulation, or whether other factors happened to change during the same period.

The answer is rarely straightforward because we cannot observe reality and its alternative at the same time. We cannot see a company both launching and not launching the same campaign simultaneously.

Experiments, data, and statistical models can help us get closer to an answer. And that matters because many decisions require more than knowing what is associated with an outcome. We need to understand what actually changes it.

Testing Decisions Before Making Them

Some decisions are too expensive or risky to test directly in the real world.

Imagine a logistics company considering opening a new distribution centre. Instead of immediately committing to the investment, it could build a model of its operations and test different scenarios: changing warehouse locations, increasing the number of vehicles, dealing with a sudden rise in demand, or accounting for supplier delays.

The company can then observe what might happen before making any real-world changes.

This is the basic idea behind simulation and digital twins.

A digital twin does not necessarily have to be a visually impressive 3D replica of a factory or a city. More fundamentally, it is a representation of a system that allows us to test changes before implementing them in reality.

A manufacturing company might use one to understand what happens when production increases. An energy company might explore the impact of declining efficiency in one of its assets. A city could test changes to its road network before construction begins.

In environmental research, these models can also be used to study scenarios we would rather not test directly, such as increasing light pollution or changing environmental conditions over several years.

The model is no longer simply describing reality. It becomes a space where decisions can be explored before we commit to them.

The Future Is Not a Single Number

Many models give us a single number: expected sales of ten million, projected demand of twenty thousand orders, or an 8% probability of default.

These numbers are useful, but they can sometimes give us a greater sense of certainty than we actually have.

The future rarely arrives as a single number. Demand may be higher or lower than expected. The effect of a decision may differ from what we predicted. Even the conditions on which the model was built may change altogether.

This is where probabilistic modelling becomes particularly useful for certain types of decisions.

Rather than treating an outcome as certain, we try to understand the full range of possible outcomes. What could happen? How likely is each outcome? And what risks are we taking if we are wrong?

This way of thinking is particularly important in insurance, risk management, banking, healthcare, and any other field where the cost of making the wrong decision can be high.

Where Should We Actually Start?

Today, we have no shortage of tools: machine learning, artificial intelligence, probabilistic models, simulation, digital twins, causal analysis, and optimisation.

That abundance makes it very easy to start with the technology itself.

Should we use AI? Do we need a machine learning model? Should we build a digital twin?

But perhaps the better starting point is much simpler: What decision are we actually trying to make?

Only after answering that question do the other questions begin to fall into the right order.

What information do we need?

What do we not know?

What alternatives are available?

And what is the cost of being wrong?

The goal is not to use the most sophisticated technology available. Nor is it to build the most complex model possible.

What matters is whether the analysis actually helps us answer the question that matters.

Because in many cases, the problem is not a lack of data. Nor is it a lack of models.

The problem is that the decision is bigger than the model itself.

And perhaps this is where data, statistics, and artificial intelligence begin to come together in their most useful form: not simply to understand what happened or predict what will happen, but to understand the choices available to us before deciding what to do.