Two years into designers having generative tools on the desktop, the honest summary is narrower than the marketing suggested and more interesting than the backlash allows. Some parts of the job genuinely changed. Others didn’t move at all, and a few got quietly harder.
Worth separating them, because the teams getting value are the ones who worked out which is which.
The cost of a second option went to roughly zero. This is the real shift. Exploring an alternative direction used to cost an afternoon, so most projects explored two or three and committed early. Now you can put fifteen directions on the wall before lunch. That changes what “explore” means as a verb — it stops being a budgeted phase and becomes something you do continuously, including after a direction is supposedly locked.
The blank canvas stopped being a problem. Not because AI produces good starting points — it mostly produces average ones — but because reacting to something mediocre is far easier than starting from nothing. Plenty of designers now generate a first pass they fully intend to throw away, purely to have something to argue with.
Production grunt work compressed hard. Resizing for six breakpoints, generating icon variants, writing placeholder copy that isn’t lorem ipsum, drafting alt text, producing the twelve empty states nobody budgeted for. None of this was ever the interesting part, and most of it is now minutes rather than hours.
That’s a real productivity gain. It is also, notably, all upstream and downstream of the actual design decision.
Knowing which option is right. Generating forty variants doesn’t tell you which one to ship. That judgement is still entirely yours, and it’s the part clients are paying for. If anything the volume makes it harder — see below.
Systems thinking. AI is genuinely poor at consistency across a body of work. It will happily produce a beautiful screen that uses a spacing value appearing nowhere else in your product, a near-duplicate of an existing component under a different name, and a colour that’s two hex values off your token. Each screen looks fine. The system rots. Anyone who has inherited an AI-assisted file six months later has seen this.
The bottleneck was never drawing. This is the one that gets missed. On most projects the slow part was never producing the artefact — it was deciding what to build, getting four stakeholders to agree, waiting on legal, discovering the API can’t return that field. Making the drawing part five times faster leaves those untouched. A team that was blocked on alignment is still blocked on alignment, now with more mockups.
Real constraints still come from the product. What states exist, what the data actually looks like, what happens when the list is empty or has ten thousand items, what the error copy says when the payment fails. AI doesn’t know any of that, and it will confidently design around the happy path every time.
Here’s what the workflow conversations tend to miss. If producing options is now cheap and choosing between them is not, the constraint has moved from production to curation — and curation is a harder, less teachable skill than most teams admit.
Reviewing forty generated variants well is genuinely difficult work. It requires holding a clear point of view while looking at a lot of plausible material, and plausible is exactly what these tools are optimised to produce. The failure mode isn’t picking a bad option. It’s picking the third-best one because by variant twenty-eight everything looked fine and you’d stopped seeing.
The practical answer is unglamorous: decide the criteria before you generate. Write down what this screen has to accomplish, for whom, under what constraint — then generate. Judging against a stated intent is a different cognitive task from judging against taste, and it survives volume much better. The patterns worth borrowing here are covered in more depth in this guide to AI UI design, which is more useful on the review side than the generation side.
It’s AI in the product you’re designing.
Using a generative tool to produce a screen faster is a workflow change. Designing an interface where the product itself behaves probabilistically is a different discipline, and it’s the one showing up in more briefs every quarter. What does a loading state look like when the wait is fifteen seconds and the outcome is uncertain? How do you show confidence without a percentage nobody trusts? What’s the undo affordance for something the system did on the user’s behalf? How does a user correct a model that’s wrong in a way that feels like collaboration rather than filing a bug?
None of those have settled conventions yet. The teams building these interfaces — including engineering-led shops like Netguru, where the design and the model behaviour get decided together rather than in sequence — are largely working them out in production, which is where most interface conventions came from anyway.
That’s the shift worth preparing for. The tooling change is real but it’s mostly a speed improvement on work you already knew how to do. Designing for systems that are sometimes wrong, and making that feel acceptable to a user, is new.
Use the tools for exploration volume, first drafts you intend to discard, and the production work that was never interesting. Don’t expect them to touch the parts that were actually slow. Budget real time for reviewing what you generate, because that’s where the constraint went. And keep your design system’s rules somewhere the AI can’t quietly ignore them.
The workflow changed less than the discourse suggests. What’s coming next changes more.
AI Articles
