Where AI fits in a design process, and where it falls short

A lot of agency writing about AI is defensive, and you can feel it. It always ends up in the same place: creative work is somehow beyond automation, so you should keep paying for it.

We’ve found something more specific, and more useful. AI has made the cost of making things drop sharply, while the cost of deciding things has barely moved. Which parts of a project fall on which side of that line is the real question, and it’s worth knowing before you commission anything.

What it’s genuinely good at

Exploring directions

It can produce fifty directions in an hour. Most will be bad, but that’s always been true of early sketching. At the start of a project, lots of options is exactly what you need, and this is the biggest change to how the first week of a project works now.

Reading

Category research, competitor audits, forty-page brand documents, years of customer reviews: it summarises material you’d otherwise only skim. It’s quick, and accurate enough to work from, as long as someone checks the claims that matter most.

First drafts of words

It’s useful for exploring tone, testing headline variations, and getting ten versions of the same idea so you can work out which one you actually mean. It gives you a good starting point, but the final line should still come from a person.

Scaffolding

This covers placeholder images, textures, backgrounds, rough layouts and anything else that exists to be replaced. It’s also good for writing code once the design decisions have been made. An interface that’s already been worked out gets built much faster than it did two years ago.

Naming longlists

It can produce two hundred name ideas in ten minutes. Getting to a shortlist is a different job. That takes judgement about trademarks, the category, how a name feels to say and what it does to the person hearing it, and the tools don’t have that.

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

Where it falls short, and why

It’s worth understanding how these failures actually happen, because “it lacks soul” tells you nothing about when to trust it.

Typography

These models generate lettering that passes for type at first sight. What they’re missing is any sense of measurement: kerning pairs, optical alignment, and how a headline relates to body text three sizes smaller. Type is a precise craft that’s judged by eye, while generation aims for something plausible. The result seems fine until it’s set at 11px in a real paragraph, and then it falls apart.

Logos

A logo has to work under a lot of constraints at once. It needs to hold up at 16 pixels, in one colour, embroidered on a jacket, engraved, animated and on the side of a building. A model is optimising for one attractive image at one size. Those are different jobs, and the difference often only shows months later, when changing the logo has become expensive.

Systems

A brand is a set of rules that has to hold across hundreds of uses, made by people who weren’t in the room when it was designed. Image generation works one piece at a time, and each output is created largely on its own. Reference images and custom-trained models help, but consistency across a whole system is still where these tools struggle most. You can generate a beautiful poster. Getting the fortieth to clearly belong with the first thirty-nine is much harder.

Editing

These tools tend to add. Ask one to improve something and you’ll usually get more: more elements, more copy, more gradients. Knowing what to take away is a judgement about the whole piece, and the model only sees your prompt. Restraint has to come from a person, deciding again and again, against the way the tool naturally pulls.

Choosing

This is the most important one. A model tends to produce something close to the middle of what it has learned from. That’s the technology doing what it was built to do, and it’s why AI-assisted work from six different companies can end up looking like it came from one. If your goal is to be unmistakable, a tool that pulls towards the average is working against you. You can push it away from the middle with effort, but it keeps drifting back.

What this means for what you pay for

The balance has shifted. Making things used to be most of the cost of a brand project, and now it’s a much smaller part. The cost of deciding hasn’t moved: working out what the brand should be, what to leave out, which two of the fifty directions matter, and whether the result will still hold up in three years.

So our honest advice is to use the tools for everything on the first list, and do it yourself if you like. Most of it is self-serve now, and anyone charging you for a naming longlist is charging for ten minutes’ work. Pay for the second list. Or skip the second list entirely, as long as you understand what you’re giving up.

What we used on this site

Some of the code was written with AI assistance, and so was some of the research and competitor analysis. Our blog posts are drafted with AI help, then rewritten, fact-checked and edited by us. The type system, the layout, the logo and every decision about what to leave out were ours.

That’s roughly the split we’d expect on most projects, and it’s the one we’d suggest you use yourself.

Frequently asked questions

Can AI design a logo?

It can generate logo-like images quickly, which is useful for exploring ideas. What it struggles with is making a mark that works at every size, in one colour and across every material, and that doesn’t look like everyone else’s.

How do design studios use AI?

Mostly at the start of a project and for production work: exploring directions, research, first drafts of copy, naming longlists, placeholder visuals and code. The decisions about what the brand should be, and what to leave out, still come from people.

Will AI replace graphic designers?

It has already taken over some of the making. The deciding is still done by people, and that’s where most of the value in a brand project sits.