Domain 4 — Workflow Integration and Solution Design (16%)

16% of the exam, roughly 10 questions. The second-heaviest domain, and the one that most resembles a management interview.

4.1 Finding the right task to give Claude

Good candidates share a shape: language-heavy, repeated often enough to be worth setting up, tolerant of a draft-then-review cycle, and not dependent on information Claude cannot reach.

4.2 Designing the workflow around a review step

A workflow that works in practice looks like this, and the review step is not optional:

1. Trigger — what starts the task, and how often.
2. Inputs — which documents or data are supplied, and where they come from.
3. The Claude step — the prompt or Project doing the work, written once and reused.
4. Review — who checks what, proportionate to the stakes.
5. Output — where the result goes and who owns it.

Where a workflow fails, it is usually because step 4 was left implicit and therefore skipped once the novelty wore off, or because step 2 depended on someone remembering to attach the right file.

4.3 Fitting into what people already do

The exam consistently favours integration over replacement. A process people already follow, with a Claude step inserted where the drafting happens, gets adopted. A parallel new process that requires everyone to change how they work gets abandoned, however good it is in isolation.

Similarly, when rolling something out to a team, the exam favours: start with one well-understood task, document the prompt or Project so it is reproducible, prove it works, then expand. Not: convert the whole department at once.

4.4 Communicating value and limitations

This is an explicit task statement in the blueprint, and it comes up as questions about what you tell a sceptical manager or an over-enthusiastic one. The examinable answer is always the honest, specific one:

• Say what it does well, concretely — "cuts the first draft of the weekly report from two hours to twenty minutes", not "transforms productivity".
• Say what needs checking and by whom, without hedging it into meaninglessness.
• Say what it cannot do — reach systems it has no access to, guarantee factual accuracy, make decisions the organisation requires a person to own.
• Where output is customer-facing or consequential, be clear about how AI involvement is disclosed.

Watch for — options that promise a specific percentage of time saved, or that describe the tool as removing the need for review. Overselling is scored as a wrong answer just as firmly as refusing to use the tool at all.

4.5 Measuring whether it worked

Adoption arguments need evidence. Useful measures are the ones tied to the task: time to first draft, number of revision rounds, throughput of a queue, consistency of format across a team. Vague measures — "team sentiment", "innovation" — do not survive a budget conversation, and the exam knows it.

← Previous: Domain 3 — Product and Model Selection  |  Next: Domain 5 — Configuration and Knowledge Management →