Sales, service, logistics: where predictive analytics earns its keep
An accurate model is only the beginning. Profit appears where there is an action scenario and a threshold that triggers it.
Predictive analytics has become a noticeably more common topic in business conversations. Accessible tools have arrived, data has accumulated, and consultants have learned to put together compelling demonstrations. But the gap between "we have a predictive model" and "we are making money from it" is still wide - and it is worth understanding why before spending budget.
I have seen several projects that worked well technically but ended up on the shelf. The reason was almost never the quality of the model.
Where the gap comes from
A model predicts a probability or a value. On its own it does nothing - it only informs. To turn that into profit you need three things: someone or something has to receive the prediction, understand what to do with it, and actually do it.
When that chain does not exist, the model becomes a report that someone occasionally looks at - usually after the event has already happened.
The right question when starting any predictive project is not "how accurate is the model" but "what exactly will change in the behaviour of a person or a system when the model produces a result".
Where the economics work
In sales, predicting customer churn works when the company has a response scenario - a concrete step a manager takes with a customer who has a high probability of leaving, and a threshold at which that step is triggered. Without this the manager receives a list of a hundred "at-risk" customers and does not know what to do with it.
In service and maintenance, predicting equipment failure saves money when the company can actually plan preventive work based on the forecast rather than purely on schedule. Otherwise the prediction just adds information to an existing calendar that nobody is prepared to reorganise.
In logistics, demand forecasting reduces inventory and improves availability - but only if purchasing and replenishment processes are genuinely driven by those forecasts, rather than by "manager experience" that treats the forecast as just another input it can choose to ignore.
Where the pretty slides usually come from
A few patterns I see more often than others.
The model optimises a metric that is not an operational lever. For example, it predicts conversion probability, but the company cannot actually change anything in the process for different segments - all customers go through the same funnel regardless.
The action threshold is not defined. The model says "probability 0.67" - so what? The company has not decided at what value to act, who makes the decision, and who is responsible for the outcome.
The data the model was trained on does not match the data that will be available at decision time. This is the classic data leakage problem, which surfaces in production.
The model output requires manual interpretation. If every forecast needs to be explained by an analyst before anyone can act on it, there will be no scale.
How to evaluate a project before launching it
A few questions worth asking before approving a budget:
- What exactly will change in the company's actions when the model produces a prediction? Who, what, and when specifically?
- Is the action threshold defined - the value or signal at which a response is triggered?
- Does the team have the capacity and authority to act on the forecast?
- How will the result be measured - not model accuracy, but business impact?
- Who owns this process after launch?
If these questions have no answers, the project will most likely produce a nice dashboard and stop there. If they do, the chances of a real effect are substantially higher.
Predictive analytics works. But not on its own - only in combination with an operational process that is ready to rely on it.