Page 208 - CyberSecurityEssentialsEnglish

● Remember: Section summary
• Key idea: A workflow is a set of clear commitments about the beginning, the execution,
the decision, and the end.
• Practical decision: Write a definition card for the process before choosing an
automation tool.
• Mistake to avoid: Documenting the normal path while neglecting exceptions and

responsibilities.

5.4 Where Do We Place AI Inside the Process?

◆ Key idea: give each step to the executor that suits it.

Deterministic steps follow fixed rules and belong with rules or conventional tools.
Generative steps suit understanding and drafting and need a check afterwards. Human
steps carry the judgment and the authority.

Not everything AI could carry out should be assigned to it. Strong design divides the steps
into three layers: deterministic, generative, and human judgment. Deterministic steps follow
fixed rules and can be tested easily, such as checking that a field exists, computing a

percentage, or renaming a file. Generative steps deal with language and ambiguity, such as
summarizing, flexible classification, and proposing wording. Human judgment remains

where approval, responsibility, and high-impact exceptions lie.

When we use a language model to carry out a deterministic step, we add uncertainty for no
benefit. We do not ask a model to sum a column when a spreadsheet can do it exactly.
Conversely, we do not try to write hundreds of rules to summarize varied complaints when
a model can extract the themes and a person can then review them. Choosing the type of

engine for each step is one of the most important workflow decisions.

Structured outputs reduce the difficulty of connecting AI to the rest of the process. The
common format for this is JSON, a plain-text way of writing data as named fields and values

that programs can read directly. There is a difference between asking a model for a table
inside prose, asking for free JSON, using a JSON schema that states which fields must be
present and what type each one is, and enabling strict schema adherence.

Adherence to form makes programmatic connection easier, but it does not prove that a

figure, a source, or a judgment is correct. A gate for meaning and evidence therefore comes
after it [4][5].
In sensitive processes, we use a human in the loop to stop the path at a defined point and

request a decision or a correction before continuing. The principle is not that a person
reviews everything, but that a person appears at the point that changes the result or
prevents harm. Approval may be required before sending an external message, updating an

official record, executing a transaction, or handling a low-confidence case [6].

207