AI in a design studio: a tool, not a selling point

When people talk about AI in branding, they almost always mix up two opposite things. On one side, models that produce an image from a sentence. On the other, code a designer writes so that a shape behaves in a certain way. Both are “generative”, both put a machine to work, but they run in opposite directions. For a studio, that is the whole question.

Why doesn’t “generative” mean “AI”?

In creative coding, you write the rule and the machine produces the result. You decide that a letter bends like a soft material, that a grid rearranges itself at a given density, that a data point becomes a line. The rule is explicit, it belongs to the brand, and it behaves the same way every time it is applied.

With an image model, it works the other way round: you describe the result, and the machine guesses the rule from everything it has seen. It is very fast, but the rule stays inside the box. You can’t read it, hand it over, or be sure it will produce the same thing tomorrow.

This is not a technical nuance. A visual identity is, at heart, a set of rules: a type system, proportions, a way of framing, a behaviour in motion. A tool that makes those rules explicit strengthens the brand. A tool that hides them produces images, not a system.

Where does AI actually come into building an identity?

At Initial, generative models come in at two specific moments, and almost never in between.

Upstream, to explore. Before drawing anything, we need to look wide: test a mood, place an idea in a realistic setting, show what an object that doesn’t exist yet might look like. Here AI works as a quick sketch. Above all, it lets us drop a route before investing in it, which is often more useful than finding one.

Downstream, to produce. Once the system is set, it has to be rolled out: formats, crops, image extensions, variations for social media. That repetitive work benefits from assistance, provided the rules were laid down first.

In between sits what actually makes an identity: positioning, drawing the mark, choosing a typeface, the logic of the system. There, we draw, we code and we decide. No prompt answers the question that matters most: what does this brand refuse to look like?

Why do generated images all look alike?

An image model learns from what already exists, and it leans towards what is probable. That is its strength for a sketch and its limit for a brand. An identity has to be improbable without being arbitrary: different enough to be recognised, consistent enough to be repeated.

The second problem is consistency. Ask a model for fifty visuals and you will get fifty slightly different readings of the same brief. The light drifts, proportions shift, type warps for no reason. For a one-off image, you fix it by hand. For a system that has to hold across hundreds of touchpoints, it is a design flaw.

The third is provenance. An appealing result can borrow, without anyone knowing, the visual language of an existing brand. Every image you keep has to be checked as you would check a reference found online, which takes back part of what the tool saved.

Is an AI-generated logo enough?

For a test or a side project, perhaps. For a brand meant to last, rarely, and for very concrete reasons.

A logo is not an image, it is the keystone of a system. It has to be drawn as vectors, stay legible at a few pixels and on a shopfront, work in a single colour, move, and live alongside a typeface. A generator delivers an image that looks like a logo. It still has to be redrawn, which amounts to doing the work.

Then there is ownership. In most legal systems, including European law, copyright protection assumes human authorship, and protection for a mark generated with almost no human input remains uncertain. A mark produced from existing work is also more likely to resemble one that is already registered. Before investing in signage, packaging or a campaign, that is a risk worth avoiding.

Finally, a logo alone does not make an identity. What makes a brand recognisable is the consistency of everything around it, and that is the job of a brand identity project, not a generator.

What does a generative tool delivered to a brand look like?

The Salomon project started from a question: what happens when type behaves like a three-dimensional object? We built an interactive tool that bends, stretches and rolls letters in real time, like a soft material under the forces of a shoe in motion. It generates unique compositions, posters, campaign visuals or 3D renders, that extend the brand’s technical energy.

People often take it for AI. It contains none: it is code. Every setting is a design decision made adjustable, whether distortion, size, number of rows, the gap between them or the stagger of the motion. You can pull a thousand variations from it, and they all belong to the same family, because the rule is the same. The tool is live on the project page, and anyone can play with it.

The same logic applies to House of Backbone, a Dubai-based marketing agency that runs campaign after campaign across many formats. Rather than only delivering fixed visuals, we delivered tools, a random grid and metaballs, that let the team generate its own visuals without leaving the system. For Trace, which designs sensors that measure athletes’ performance, part of the visuals is produced from the data: every session and every athlete leaves a different trace, in the same language.

In all three cases, the deliverable is no longer just an image. It is the machine that makes the images, with the brand’s rules built in.

What does it change for the client?

First, you see things earlier. A route can be shown in context, on an object, a poster or a screen, before it is fully drawn. The conversation is then about something concrete rather than an intention.

Second, the number of variations stops being a constraint. When producing an option costs almost nothing, the scarce resource becomes judgement: knowing which one to keep, which one to defend, and why. The studio’s work shifts towards editing. That is the paradox of AI in design: it makes creative direction more necessary, not less.

Finally, the brand can become more autonomous. A tool like the ones built for House of Backbone lets a team produce day to day without relying on the studio for every format, and without drifting from the identity. For brands that want to hold that course over time, it is also the role of ongoing art direction.

What should you ask a studio that says it uses AI?

“We use AI” says nothing. Four questions say a lot more:

• What is generated, and what is drawn? The mark, the typography and the rules of the system should belong in the second category.
• Who owns the result? The source files, and the tools when there are any, should be handed over to you.
• How is consistency guaranteed? Through written rules and tools, not case-by-case retouching.
• What happens when the studio is gone? A well-built identity keeps working without it.

A tool, then

AI is one production tool among others, as vector drawing and the web were before it. It deserves to be mastered, not shown off. What really interests us lies elsewhere: writing a brand’s rules clearly enough for a machine to apply them, and accurately enough for people to recognise it every time.

If you want to build an identity designed as a system, with the tools that come with it, let’s talk about your project.