Every company I work with arrives with the same sentence: we have a lot of data and we are not sure what to do with it. That is a fair place to start. The confusion usually is not about the data itself. It is about not knowing which kind of question you are asking.
Analytics splits into four types. They build on each other, they answer different questions, and they cost very different amounts of time and money. Knowing which one you actually need will save you a quarter of wasted work.
1. Descriptive analytics: what happened?
This is the foundation, and it is the one most businesses already have in some form. Revenue last month. Churn rate by plan. Support tickets per week. Descriptive analytics counts, sums, and averages what already occurred.
It sounds basic, and people often want to skip past it toward the impressive machine learning work. Do not. If your monthly revenue number is wrong because two systems define a sale differently, every model you build on top of it inherits that error. Most of the value in a first analytics engagement comes from getting these numbers trustworthy and agreed upon.
Typical output: a dashboard, a weekly report, a single source of truth for core metrics.
2. Diagnostic analytics: why did it happen?
Descriptive tells you conversion dropped 12 percent in March. Diagnostic tells you it dropped because a checkout change broke the mobile payment flow for one browser.
This is where segmentation, drill-downs, correlation and cohort analysis live. You take the number that moved and you split it apart until the movement is explained by something specific enough to act on.
Diagnostic work is where a good analyst earns their fee, because the answer is rarely the first thing you find. The first correlation you spot is usually a coincidence or a proxy for something else. The discipline is in continuing to ask why after you find an answer that feels satisfying.
Typical output: a root-cause analysis, a cohort breakdown, a clear explanation of a change in a KPI.
3. Predictive analytics: what is likely to happen next?
Now we move from the past into probability. Which customers are likely to churn in the next 60 days? What will demand look like in Q4? Which leads are worth a salesperson calling first?
This is where statistical modelling and machine learning come in. It is also where expectations need managing. A predictive model does not tell you what will happen. It gives you a probability, and that probability is only as good as the history it learned from. If your business changed fundamentally last year, a model trained on three years of data is partly learning a company that no longer exists.
The honest prerequisite: predictive work needs clean, consistent historical data at sufficient volume. If steps one and two were skipped, this stage will disappoint.
Typical output: a churn risk score, a demand forecast, a lead prioritisation model.
4. Prescriptive analytics: what should we do about it?
The final step recommends an action, and sometimes takes it automatically. Not just this customer is likely to churn, but offer this customer this specific retention discount, because at this price point the expected value is positive.
Prescriptive analytics combines a prediction with business constraints and costs: your margins, your inventory, your capacity, what you are willing to spend to keep a customer. Dynamic pricing and automated inventory reordering are the familiar examples.
Few mid-sized companies need to start here, and starting here without the earlier layers is how organisations end up automating a bad decision at scale.
Typical output: a recommendation engine, an optimisation model, an automated decision rule.
So where should you start?
Almost always one level below where you think. If you are asking for prediction, the practical first question is whether your descriptive layer is solid. If two departments quote different revenue figures in the same meeting, that is the project.
- No agreed metrics? Start descriptive. Expect weeks, not months.
- Metrics agreed, but surprises keep appearing? Start diagnostic.
- Solid history, repeatable decisions? Predictive is worth the investment.
- Predictions already trusted and used? Prescriptive is the natural next step.
The four types are a ladder, not a menu. The companies that get real value from data are rarely the ones with the most sophisticated models. They are the ones who climbed the ladder in order.