Where AI fits in a design process and where it breaks

Most agency writing about AI is defensive, and you can feel it. The argument always arrives at the same place — that creative work is somehow beyond automation, and you should therefore keep paying for it.

That is not what we have found. What we have found is more specific and more useful: AI has collapsed the cost of making things and left the cost of deciding things almost untouched. Which parts of a project fall on which side of that line is the whole question, and it is worth knowing before you commission anything.

What it is genuinely good at

Divergence. Fifty directions in an hour. Most of them are bad — that is fine, it was always true of the sketching phase. You are buying volume of options at the point in a project where volume is exactly what you need. This is the single biggest change to how the first week now works.

Reading. Category sweeps, competitor audits, forty-page brand documents, a decade of customer reviews. Synthesis of material you would otherwise skim. Fast, and accurate enough to work from, provided someone checks the load-bearing claims.

First drafts of words. Tone exploration, headline variants, ten ways of saying the same thing so you can find out which one you actually mean. Never the final line. Always a useful start.

Scaffolding. Placeholder imagery, textures, backgrounds, layout stubs — anything that exists to be replaced. Also implementation code, once the design decisions are made. An interface someone has already resolved gets built markedly faster than it did two years ago.

Naming longlists. Two hundred candidates in ten minutes. Not the shortlist. The shortlist requires judgement about trademark, category, mouthfeel and what a word does to a person, and that is not in there.

None of that is small. If someone tells you AI has no place in a brand project, they are either not using it or protecting a rate card.

Where it breaks — and why

The failures are worth understanding mechanically, because "it lacks soul" is a non-answer that tells you nothing about when to trust it.

Typography. These models generate type that looks like type. What they do not have is a feedback loop on measurement — kerning pairs, optical alignment, the relationship between a headline and the body text three sizes below it. Type is a metric craft judged optically, and generation optimises for plausibility. The result reads fine at a glance and falls apart the moment it is set at 11px in a real paragraph.

Logos. A mark is not a picture, it is a constraint problem. It has to survive 16 pixels, one colour, embroidery on a jacket, engraving, motion, and the side of a building. A model optimises for one attractive still at one size. Those are different jobs — and the second one only reveals itself as a failure months later, at the point where changing it is expensive.

Systems. A brand is a set of rules that has to hold across two hundred applications made by people who were not in the room. Generation is per-artefact. Consistency between artefacts is precisely what it cannot hold, because each output is sampled independently. You can generate a beautiful poster. You cannot generate the fortieth poster and have it belong to the first thirty-nine.

Editing. Models are trained toward addition. Ask for improvement and you get more — more elements, more copy, more gradient. Knowing what to remove is a judgement about the whole, and the model only ever sees the prompt. Restraint is not something you can request. It is something someone has to decide, repeatedly, against the grain of the tool.

Convergence. This is the important one. A model returns something near the centre of its training distribution. That is not a flaw, it is the mechanism working correctly — and it is why AI-assisted work from six different companies arrives looking like one company. If your goal is to be unmistakable, a tool that regresses toward the average is working against your objective by design. You can push it off-centre with effort. It drifts back.

What this means for what you buy

The ratio has moved. Execution used to be most of the cost of a brand project. It is now a smaller fraction of it. What has not moved is the cost of deciding — what the thing should be, what to leave out, which two of the fifty directions matter, and whether the result still holds in three years.

So the honest advice is this. Use the tools for everything on the first list, and do it yourself if you like — most of that list is self-serve now, and anyone charging you for a naming longlist is charging you for ten minutes of work. Pay for the second list. Or do not do the second list at all, and understand what that means.

What we used on this site

Backgrounds and textures: generated, then heavily edited. Implementation code: partly assisted. Research and competitor analysis: assisted. The type system, the layout, the mark, the writing you are reading, and every decision about what to leave off: not.

That is the split we would expect on most projects. It is also the split we would tell you to run yourself.