The problem first, the tool second

The director tells the facilities team: “Start using AI.” Nobody knows what for. A week later, everyone has tried something different, nobody has measured anything and the report to the director says that “AI has a lot of potential.”

The mistake was starting with the tool. Claude pays off when it solves a concrete problem someone can measure.

Start with the problem

Before you open Claude, answer:

• Which tasks take the most time? The ones repeated every week, the ones with manual steps, the ones that produce errors.
• What does a good result look like? Fewer hours, fewer errors, a faster response. If you can’t measure it, you won’t be able to show it improved.
• Who decides? Priorities and final approval stay with people.

With that list, the use case chooses itself: the task that eats the most time and where Claude can help without deciding for anyone.

Claude also helps define the problem

For sorting out requirements, Claude is very useful. If a project manager has interview notes from five stakeholders, she can ask it to:

• extract the requirements from the notes;
• flag where the stakeholders contradict each other;
• list the open questions and assumptions that need confirming.

What it shouldn’t do is decide the final priorities or add requirements “they surely forgot” as if they were confirmed.

Augment before you replace

Once the case is chosen, there are two ways to bring Claude in:

	What it is	Example
Augment a workflow	Claude joins an existing step and speeds it up	In accounts payable, Claude matches invoices against purchase orders and delivery notes; the two approval levels stay the same
Redesign a workflow	The way the step is done changes	Instead of reading 2,000 comments by hand, Claude groups them by theme with counts and quotes, and the analyst checks samples and interprets

The same rules apply in both cases:

• Start with the step that takes the most time, not the flashiest one. Using Claude only to polish the final report leaves the four-day step untouched.
• Keep the controls that already exist. The person who reviews stays in place (human in the loop), especially when audits, money or customers are involved.
• Run a pilot, measure and expand based on results.

Common traps

• Choosing the most advanced feature and looking for a task to use it on.
• Asking everyone to “use Claude for an hour a day” with no defined problem.
• Replacing the whole process at once and removing the approvals.
• The opposite extreme: touching nothing until “after the audit,” or using Claude only for final formatting.
• Adopting a generic “best practice” without giving Claude the real data about the process.

On the exam

If the scenario opens with a vague mandate (“use AI”) or with someone who wants to automate everything at once, look for the option that first identifies the time-consuming tasks and defines how success is measured, or the one that starts with the costliest step, keeps the controls and expands after a pilot. Rule out the ones that start with the tool, the ones that remove human review and the ones that leave Claude in a corner where it changes nothing.