I asked Claude Code to rebuild this site from a design handoff. Its first draft included a line explaining what the side states on a Jira board were for. It sounded right. I never said it.
A mentor of mine calls this the 75% trap. The first 75% of any piece of work keeps getting easier, and the last 25% is where the work lives. AI produces a rough, working draft embarrassingly fast. Getting from that draft to something I'd put my name on is still my job.
What the last 25% looks like
My rule is to treat AI like an intern. Nobody at Agile Defense puts their name on what it gives on the first pass. I fact-check it, ask it to cite sources when something doesn't sit right, and push back. (After pushback, AI sometimes swings to the opposite answer just to please you, so the skill is knowing enough to recognize a good one.)
For this site, I had every page checked twice: once in a browser at desktop width and at 320 px, and once line by line against my writing rules. The browser check found tables that stretched the mobile layout. The writing check found the Jira line, along with other copy that went beyond the facts I'd given Claude.
Review can also live inside the tool. On our program management course, the team noticed recurring patterns in how the scripts were written. We gave our internal AI tool a skill that flags them for us or the subject matter experts to verify. Editing time dropped by about half, because patterns that used to need their own review round get caught automatically.
Who's ready for this
I think designers and project managers are better prepared for this than almost anyone.
Designers spend their careers judging unfinished work. Every review comes down to one question: does this solve the problem for the person who has to use it? A polished draft doesn't answer it.
Project managers spend their careers in the gap between done and delivered: who signs off, what was promised, what slipped. A draft that looks finished doesn't close that gap.
AI is good at making things look finished. Both jobs are about what sits behind that. Plenty of people review well. For these two, review is the main job.
Before a team decides what to generate with AI, it should decide who reviews the result.