AI changes the speed. Taste decides the direction

The constraint early on was never imagination. It was how many ideas you could afford to make visible before the budget said choose — concepts, moodboards, visual directions, layout studies and image treatments all cost roughly the same amount of time each.

They no longer do. So the shortlist gets wider before it gets narrower: more directions drawn, more layouts tried at real proportions, more alternatives put up next to each other where they can be compared rather than imagined.

The work is still made properly once a direction is chosen. What changes is that the choice is made against fifteen options instead of three, and the ones that were only ever somebody's first instinct get found out early, while changing course is still cheap.

A model will return two hundred variations in an afternoon and hold no opinion about any of them. Volume is the easy half. The useful half is deciding what is relevant, what actually sounds like the brand, what is distinctive rather than merely competent, and what to throw away.

Most of what comes back is average by construction — it is drawn from everything, so it lands in the middle of everything. Average is the one thing a brand cannot afford, and recognising it takes someone who has formed a view about what good looks like.

So the ratio matters more than the volume. Generating is cheap now; editing, refining and knowing when to stop are not, and that is where the design work has moved.

An interface over a model is a different problem from an interface over a database. The output is not deterministic — the same input can give a different answer twice — so a design that quietly assumes correctness will mislead people confidently.

That puts the weight on a few things. Showing what the system can and cannot do, before somebody asks it for something it will fail at. Showing that it is working, rather than leaving a silence. And making the result correctable, because the answer to a wrong output is an edit, not an apology.

Confidence has to be legible too. People need enough signal to decide whether to trust an answer or check it, and a system that never admits doubt teaches people to trust everything or nothing. None of this is exotic: it is clarity, feedback and control, applied to something that is sometimes wrong.

The honest test is whether a step got better, not whether AI touched it. It earns its place on the repetitive parts: first-pass variations, carrying a system across formats, organising research into something readable, drafting copy a person then rewrites, standing a prototype up quickly enough to be worth testing.

It earns nothing by being added to a stage that already worked. Pushed into every step because it is fashionable, it slows the work down and adds a review cycle for output nobody asked for.

So it goes where it removes friction and stays out of where the thinking happens. That is a judgement per project rather than a policy, and it is worth revisiting as the tools change — which they do, faster than any policy would survive.