AI can analyse thousands of numbers in seconds.

It cannot rescue a number that was never clearly defined.

I explored to try to get to the root of the issue.

Imagine three people are asked to measure how long a process takes.

One starts the clock when the application is submitted.
Another starts when someone opens it.
One stops when a decision is made.
Another stops when the final document is collected.

Every number may be accurate.

They still cannot be compared.

That is the problem hidden underneath a surprising amount of data. Before anyone calculates an average, identifies a trend, or asks AI for an explanation, someone has to answer a much less glamorous question:

What exactly are we counting?

A visitor could mean an overnight guest, a cruise passenger, or every arrival through a port.

A complaint could mean one received, one formally recorded, or one that remains unresolved.

A price could mean the advertised price, the amount paid, or the total cost after fees.

A waiting time could begin when someone arrives, checks in, submits a form, or finally reaches the correct desk.

Those definitions are not footnotes. They determine what the number means.

This becomes especially important with AI because AI is exceptionally good at making mismatched information look coherent.

Give it three spreadsheets containing a column called “processing time,” and it can quickly calculate averages, produce charts, identify apparent trends, and write a convincing explanation.

But unless someone checks how each organisation defined “processing time,” the analysis may be comparing completely different things.

The arithmetic can be flawless while the conclusion is useless.

That is why the most valuable part of a statistical system is not necessarily producing more numbers. It is establishing common definitions, consistent methods, and enough transparency that people know what each number includes—and what it leaves out.

That work is slow, technical, and easy to overlook.

It is also the foundation underneath every useful AI analysis that comes afterwards.

Before giving a dataset to AI, ask five questions:

1. What exactly is being counted?
Write the definition in plain language.

2. Who or what is included—and excluded?
A statistic can change dramatically depending on where the boundary is drawn.

3. What is the denominator?
“Twenty incidents” means something different among 100 people than among 100,000.

4. What period does it cover?
A week, a quarter, and a year can produce very different stories.

5. Has the definition changed?
A trend can appear simply because the counting method changed halfway through.

A useful prompt is:

Before analysing these numbers, define each metric in plain language. Identify what is included and excluded, the denominator, the time period, the source, and any changes in methodology. Flag any figures that cannot be compared directly. Do not calculate trends until the definitions match.

That last instruction matters.

AI wants to continue with the task. If the numbers are present, it will often begin working with them. It may infer missing definitions rather than stop and demand them.

A careful human process has to create that pause.

Good statistics are not trustworthy because they contain decimal points or appear in a polished chart. They are trustworthy because someone did the quiet work of making sure each number consistently means what people think it means.

AI can accelerate the analysis.

But first, we have to decide what counts.