Page 224 - CyberSecurityEssentialsEnglish

of the data. First we must define the unit of analysis, the period, the source, the meaning of
each field, and whether the values are complete and comparable.
When working with tables, we use computational tools for the operations and the model to

propose questions or interpret patterns after verification. For example: the spreadsheet
computes the completion rate, then the model helps to phrase a possible explanation, then

a specialist reviews whether the explanation is supported by the working context. We do not
let the model invent a cause merely because a correlation exists.

The work begins with data cleaning: removing duplicates, standardizing formats, handling
missing values, and checking bounds. You then build a data dictionary that explains the fields.

Compute the indicators in a way that can be re-executed. Finally, build a decision narrative
that separates what happened, what may explain it, what we do not know, and what action
is proposed.

AI can be used to extract fields from text or from images of documents, but accuracy must

be measured on a real sample. Critical fields such as an amount, an identity, or a date need
more verification than a descriptive field. If an error is rare but costly, an overall average of
accuracy is not enough; we need to know the types of error and where they occur.

A decision support pack does not offer only a recommendation. It usually contains the

question, the facts, the assumptions, the options, the criteria, the risks, the effects, the
sources, and the points that require judgment. In this way AI becomes an assistant in

organizing the evidence rather than a substitute for the decision owner.

TABLE 5-11: QUESTIONS BEFORE ANALYZING ANY TABLE

The question Why it matters

What is the unit of analysis? It prevents comparing different things.

What is the period and the version? It prevents mixing periods or versions.

What is the source of each field? It establishes the quality of the data.

How were missing values handled? It avoids reading a blank as a zero.

How was the indicator computed? It ensures the result can be reproduced.

What is the evidence for the
It prevents a correlation being turned into a cause.
interpretation?

What does the data not tell us? It defines the limits of the decision.

223