Every day, and in five distinct ways, none of which skip the step where a person decides what is allowed to ship, and one of which has a hard line around it.
Diagnosis. When a client reports a problem on a live site, most of the elapsed time goes into reproducing and locating it, not into the fix. A browser agent does that stretch: it reproduces the report, finds where the behaviour comes from, and reports what it found as evidence I read. This is the use that changed the job most, and it is the one that never touches the site on its own.
QA. Checking finished work before a client sees it. An agent goes through what was built and reports what it finds; whether anything needs fixing, and the fix itself, stay with a developer.
Scoping. For anything with an unknown in it, I scope with one model, hand the scope to a different model to find the holes, and bring the findings back to close them. The build plan that comes out has been argued with before anyone has paid for it.
Building. AI-assisted development, inside the normal branch and review flow, for the agency's own products: in-house applications that can later be sold. It is not used on client projects. A project quoted as human work is built as human work, by the developers who quoted it.
Learning. When the agency takes on technology nobody there has used before, I am the one who figures it out, and a model that has read the documentation is the fastest way to a first working version. The judgement about whether that version is good enough is still mine.
The pattern across all five is the same as the one on the agent systems I run: the model does the part where volume or breadth wins, and a person keeps the part where consequences live.
My own answers, from my own work. They change when the work does.