Most of the systems we use were designed for a person clicking through them.
The screen tells that person what they can do.
The menu shows where everything is.
And the error messages explain, more or less, what went wrong.
When another type of actor tries to operate the same system, those cues are no longer enough.
I do not believe the answer is to teach it to imitate clicks at any cost.
First, we need to make explicit what many interfaces hide:
Which actions exist.
What information each one requires.
Which permissions enable it.
What result it should produce.
And when a person must intervene.
That also improves the human experience.
A process with clear rules, understandable states, and visible exceptions is easier to operate from a screen, an integration, or a conversation.
Designing for artificial intelligence does not begin by adding an assistant.
It begins by no longer depending on someone knowing every hidden path through the system.
Screens do not disappear.
But they stop being the only place where the operation can be understood.