A couple of years ago, most of the conversation around AI and design was framed as a threat: would the tools eventually just do the job. That conversation has mostly moved on, at least in the day-to-day work I do. The more useful question turned out to be narrower, not "can AI design this," but "where in my process does AI actually remove friction, and where does it just move the friction somewhere else."
Prompt engineering, in practice, is less about clever phrasing and more about a skill designers already had: taking a fuzzy problem and giving it enough structure that someone or something can act on it. The difference now is that the "someone" can be a tool that returns ten directions in the time it used to take me to sketch two. That's genuinely useful. What it doesn't do is tell me which of those ten directions actually solves the user's problem, respects the technical constraints, or holds together as a system six months from now. That part hasn't moved.
Where it earns its place
Throwaway exploration. Getting from a blank canvas to five rough directions is where I've found the most honest time savings, not because the output is final, but because reacting to something is faster than generating from nothing.
Restructuring rough input. Turning messy stakeholder notes or a half-formed content dump into a workable component structure is a task I'm happy to hand off entirely, then edit.
Variation, not decision. Drafting flow variations for me to evaluate is a good use of the tool. Picking the winning variation isn't.
Where it still doesn't
No tool I've used yet can sit in a usability session and notice the half-second hesitation before someone clicks the wrong thing, or knows which technical constraint a stakeholder mentioned in passing three meetings ago and forgot to write down. Those are still the parts of the job that decide whether a product actually works, and they're the parts I'd be nervous handing to anything that hasn't sat in the room.
The job didn't get smaller. The part of it worth paying for shifted from producing every pixel to deciding which ones matter.
If I had to give one filter for evaluating a new AI design tool, it's this: does it remove friction before a decision, or does it try to make the decision for you? The first category has earned a permanent spot in how I work. The second, I'm still skeptical of, not because the technology isn't capable, but because the decisions are the part of the job I actually want to keep doing.
