Data science: from raw data to a decision someone makes

A distribution company spends eight months on a dashboard: twenty-two charts, daily refresh, access from a phone. It is presented to the committee and put up on the boardroom screen. Six months later someone notices the inventory turnover figure has been frozen for three weeks, because a field was renamed in the source system. Nobody had noticed, and that says more than the twenty-two charts: for three weeks, no decision depended on that dashboard.

That outcome is more common than a noisy failure, and worse, because it teaches nothing. The cause is not technical. Nobody defined, before starting, which decision would be made differently once the number was available.

The one-question test

Before spending anything, ask: which decision will be made differently when this number exists, and who makes it? If the answer is “we will have visibility”, the project has no recipient. A useful answer sounds like this: when an item turns over less than once a quarter, the purchasing manager stops reordering it automatically and decides case by case. It has a verb, an owner and a threshold, and in three months you can check whether it happened.

The second question is just as uncomfortable: what is done today instead. There is always something, even if it is someone’s experience or a spreadsheet. That is the baseline, and without it there will be no way to know whether the project worked.

When in the cycle you are buying

In January 2026, Thomas Davenport and Randy Bean published their five trends for the year in artificial intelligence and data science in MIT Sloan Management Review. The first is unusually blunt for that publication: the artificial intelligence bubble will deflate and the economy will suffer (MIT Sloan Management Review). For a mid-sized company the reading is not alarmist but a matter of timing: what is bought at the peak of the promise is paid in full and half used. The decisions that survive are the ones made on a concrete problem with a prior measurement.

Before the model, the data

A model trained on inconsistent data produces consistent, wrong answers, which is worse than having no answer, because nobody distrusts a number. For each piece of data that will be used there are three questions:

• Who owns it? Not who enters it, but who answers for it being correct. It is always someone in the business, because whoever knows whether a customer is correctly classified is whoever sells to them.
• Does everyone define it the same way? If sales counts a sale on shipping and accounting on invoicing, both figures are correct and they do not match.
• How do you know when it is wrong? With automatic checks that alert when a value stops arriving, goes out of range or changes format.

According to McKinsey, around two thirds of organisations had still not scaled artificial intelligence beyond pilots in 2026 (McKinsey, State of AI trust in 2026). The difference lies in how each is fed. A pilot runs on a file someone cleaned by hand the week before. In production the data arrives by itself, every day, with that day’s errors.

What can be asked, and what cannot yet

What happened, why it happened, what will happen and what should be done are four different questions, and they are answered in that order. Most data projects that do produce a return stay with the first two, and there is nothing wrong with that. Predicting requires enough history, a pattern that repeats and reliable data, because the model learns past errors as if they were facts. Buying prediction on data that still cannot answer what happened produces a demo that looks good, because the demo uses the past, which is already known.

How to buy this without buying smoke

An analytics proposal is hard to evaluate for anyone outside the trade, because almost everything it says is true and almost nothing can be checked. Four requirements change that, and they are asked for before signing:

• The decision and its owner, written in the proposal and not just said in the meeting.
• The baseline, measured before signing.
• A short, paid test with your own data, with its odd cases and gaps, and with the success criterion written before starting.
• Who watches the model after delivery. A model nobody reviews stops being right without warning, because the world it learned from changes.

The value of data is not in having it. It is in the decision it changes.