The short answer
Gartner describes four types of business analytics: descriptive (what is happening), diagnostic (why did it happen), predictive (what will happen) and prescriptive (what should we do). Most companies stop at the first two. That is reporting. The value for operations is mostly in the last two, because that is where a decision can still change the outcome.
The four levels, with examples
| Level | Question | Example in a print or manufacturing operation |
|---|---|---|
| 1. Descriptive | What happened? | 768 blanks were spoiled last month. |
| 2. Diagnostic | Why did it happen? | A third of them were black tees from one vendor batch, pretreated by hand on the night shift. |
| 3. Predictive | What will happen? | Job 48213 has a 6.1% chance of spoilage, about three times the shop average. |
| 4. Prescriptive | What should we do now? | Raise pretreat to 36 g and pull 8 spare blanks before the job starts. |
A report tells you about last month. Intelligence tells you about the job on the press this afternoon, while there is still time to act. The numbers in the example are illustrative; the shape is what matters.
Why most companies stop at reports
- The tools were built for it. Reporting software is very good at totals and charts, and not designed to make a call on one order or one job.
- The data sits in separate places. Predicting spoilage needs the job, the blank batch, the press, the shift and the reprint history together.
- Nobody owns the action. A report goes to a meeting; a prediction needs a person who will do something today.
- Trust. People act on a recommendation when they can see why it was made.
What it takes to reach levels 3 and 4
| Ingredient | What it means in practice |
|---|---|
| History | At least a year of the records that describe the outcome you care about |
| Joined-up data | Orders, jobs, purchasing and quality records linked by the same IDs |
| A clear decision | One question with money attached, and a person who acts on the answer |
| A reason with every flag | ‘High risk because: dark garment, hand pretreat, new vendor batch’, not just a score |
| A feedback loop | Record what happened after each flag, so the model learns and you can measure it |
How to tell if it is working
- Decisions changed: how many flags led to an action.
- Money saved: the cost of the problems avoided, measured against your baseline.
- False alarms: how many flags turned out to be fine. Too many and people stop reading the list.
How to test it
- Choose one outcome to predict, such as spoilage, late orders or loss-making orders.
- Pull the related history, read-only, ideally a year or more.
- Build the model and test it on a past period. Would it have flagged what actually went wrong?
- Run the daily list live and record what people do with each flag.
You do not need a new system for any of this. The data is usually already in the one you run. See Do you need a new ERP to use AI?
Want to see what your data could predict? Bring us one problem. We’ll show you on your own data.
Bring us one problem →Frequently asked questions
What is the difference between reporting and intelligence?
Reporting tells you what happened and why. Intelligence predicts what will happen and recommends what to do while there is still time to act.
What are the four types of analytics?
Descriptive (what is happening), diagnostic (why it happened), predictive (what will happen) and prescriptive (what should we do), as described by Gartner.
Do I need a data warehouse first?
Not always. Many first predictions can be built from a read-only extract of the system you already run.
How much history do I need for predictions?
Usually at least a year of the relevant records, so that seasonal patterns are included.
How do I know predictions can be trusted?
Test them on a past period first, show the reason with every flag, and track false alarms once they are live.
