While the assistant works it shows what it is doing, a step at a time.
The messages are real
"Checking account balances", "Pulling recent transactions", "Running profit and loss" — each names an actual lookup being performed against your workspace. They are not a loading animation with words on it.
That is why they vary between questions. A question needing one figure shows one step; a question comparing several months across customers and costs shows several.
Why an answer can take a moment
Most questions need more than one lookup, and each has to finish before the next can begin — you cannot compare two periods until both have been fetched.
So the time is mostly the work rather than the writing. A question that reads simply can require several steps behind it, and the progress messages are how you can tell which.
What "thinking" means
Working out what to look up before looking it up. On a vague or broad question this stage is longer, because more of the answer is deciding what the question actually is.
Asking for something specific is usually faster for the same reason.
Why it shows its work at all
So the answer can be checked. Seeing which figures were pulled is the first half of trusting the result, and the records behind them are the second — see the evidence chain.
An answer that arrived without touching the data it claims to describe would be indistinguishable from one that did, unless the steps are visible.
When a step is missing
Steps you might expect and do not see are usually a permission rather than an omission. The assistant cannot look up what your role cannot reach — see roles and permissions.
Longer pieces of work
Anything with several stages becomes a plan it can carry across sessions rather than holding you at a progress message.