Start with the claim being answered, because it is usually stated wrong. The popular version is that AI is making design obsolete. That claim confuses design with drawing.
Drawing got cheap. Deciding did not.
Here is the mechanical consequence, and it is the part worth arguing about. When options were expensive, "we went with the one we had time to build" was a defensible answer in a review. A team can now put twelve credible directions on the wall by Wednesday, which turns that same sentence into a confession. Somebody has to say why this one and not those eleven, and say it in a form a team and an agent can both execute. That is a designer's job, it used to be maybe fifteen percent of the week hidden inside the making, and it is the whole week now.
So what does the work look like in practice.
Constraints, not principles. A principle has no failure condition, which is why "delightful, human, trustworthy" has never stopped anyone from shipping anything. A constraint is a rule you can catch being broken. Any suggestion a user can accept in one click needs an undo that is equally cheap. A confirmation cannot be the first time someone learns what the system already did. Both of those are checkable by a person, by a reviewer, and by an eval. Test: hand it to someone who was not in the room and ask them to point at the moment it was broken in a real output. If they cannot, rewrite it.
The range, not the screen. Generated output has no final state to draw, so an artifact showing one state is a sample of a distribution presented as a decision. What you write instead is three columns: best case, worst case you would still ship, and what the product refuses to do. Anything between the second and third is allowed and does not need your review, which is also how you stop being a bottleneck.
A rubric instead of a Thursday meeting. Most design taste lives in a recurring review on someone's calendar, which caps quality at the number of things that person can personally look at. Written down, scored, and versioned, it executes without you. Two people score the same five outputs independently: where they disagree, the rubric is vague, and where they agree but both feel wrong about the score, the rubric is measuring the wrong thing. The mechanics are the same ones behind The Eval Is The Spec, pointed at interface quality rather than model output.
The wrong path, owned. In a deterministic product the unhappy path was rare, so it went to whoever had capacity that sprint. It is common now. What the product does when it is unsure, how a person sees what already happened on their behalf, what undo means when an agent acted at 3am, and what gets said in the thirty seconds after the system is confidently wrong. None of that was in a design system in 2023, and most of it is still unnamed, which is exactly why claiming it is worth something.
The uncomfortable part for leaders: a team that is anxious about this cannot be fixed with reassurance, because reassurance is the optimistic version of the same passive frame. What works is a job description. Pick two rules, give each a named owner and a two-week deliverable, and put a thirty-minute check on the calendar. Anxiety about the future dissolves fastest when there is a specific piece of it somebody has been made responsible for.
One thing to do this week: take the last design decision you had to argue for, and write the three constraints that would get a competent stranger, or an agent, to that same decision without you in the room. Send it to your team. The full six-rule worksheet, with the failure condition and check for each, is at The Design Operating Rules, and the long argument is in Design Just Got Promoted.