Generative AI in brand work: our position, in plain language

Clients ask about AI in two ways. The optimists ask in week one whether it means the brand will cost half as much. The worriers ask in week six, quietly, whether the moodboard they loved was "real". Both deserve straight answers, and "we use the best tools for the job" is not one — it's a fog machine. So here is our position, the same one we hand to every client before a brand engagement starts: what we use generative tools for, what we refuse to use them for, how we check provenance, and exactly what we'll tell you about deliverables.

Two beliefs sit under the whole thing. First: a brand's value is in its decisions — chosen, argued for, defensible — and generation is not decision-making. Second: trust is a brand asset too, and a studio that hides its process is borrowing against yours.

Where the tools genuinely help

We're not ascetics. Generative tools earn their keep in the early, disposable phases of brand work — anywhere the output is throwaway exploration rather than final identity.

Moodboard divergence. In week one, the job is volume: thirty directions so the client can kill twenty-eight of them fast. Image models are excellent provocation machines here. "A coastal punk botanical garden, editorial photography, 1970s film" takes forty seconds to render and ten seconds to reject — and the rejection is the point. Cheap divergence makes the expensive convergence honest. What we never do is let a generated moodboard image set the palette or type direction directly; it's a conversation starter, not a reference.

Prompt-storming names and lines. Language models are a useful extra generator in the volume phase of naming — particularly for exhausting the obvious. Give a model the brief axes and it will produce everyone's first fifty names in one shot, which lets us skip to the interesting territory faster. Its hit rate on good names is roughly zero; its hit rate on clearing the plateau of cliché is excellent.

Variation drudgery. Resizing concept mockups for nineteen aspect ratios, roughing out interior-wall previews, faking a coffee-bag test print at three materials: anywhere a human's judgement isn't the content of the work, the machine's speed is welcome.

The designer's private sketchbook. Our designers use AI the way they use lorem ipsum and stock crops — as placeholder energy inside explorations that will be redrawn, reshot or rebuilt before anything is client-facing. We've written about the seatbelt rules for AI in the design process separately; the summary is: accelerate drafts, never delegate judgement.

Where it fails — and why we refuse

The failures are not moral abstractions; they're craft failures with receipts.

Final identity marks. A generated logo is, definitionally, a statistical average of existing logos. Distinctiveness — the entire commercial point of an identity — is what averages erase. We've also never seen a generated mark survive the construction questions a real identity must answer: how does it rebuild at sixteen pixels, in one colour, reversed, embroidered, with the wordmark in Farsi? Marks are systems, not pictures, and generators output pictures.

Meaning-bearing choices. A brand's colour palette is a position — strategy wearing paint. When a machine picks the palette because the image "felt right", nobody can defend it in the room that matters: the client's board, the trademark attorney's office, the press interview. An identity you can't argue for is an identity a committee will re-open forever.

Voice at the foundation. We do draft with language models in the volume phase of copy exploration, but a foundational voice — the chart the writers will teach from — is built from the founders' actual mouths, interviews, and the company's real stories. Model-generated voice work has a recognisable flavour: competent, upbeat, weightless. It reads like everyone because it is everyone.

Anything whose provenance we can't state. If we can't tell you where an asset's lineage runs, it doesn't ship. This rules out most generated illustration and photography for final use, not on quality grounds but on honesty grounds — the next section is the reason.

Our provenance and disclosure rule

Everything client-facing carries a stated provenance, in the project file and in the final delivery notes: hand-made, licensed (with source), or generated (with tool and the human work done on top). This is not paperwork for its own sake. Three practical things follow.

First, the IP conversation stays clean. Ownership of deliverables transfers at final payment — a clause that only works — ethically and practically — when we can state what the deliverables are made of. A brand built on assets of murky lineage is a gift nobody should accept.

Second, trademark and infringement risk gets priced honestly. Invented marks we draw ourselves, we can defend. Anything generated that resembles existing work is a liability sitting inside your identity, and identity is the one place a client cannot afford latent liability. The screening discipline from our naming work applies to every visual asset too.

Third, when a client asks for generation — and sometimes, for budget or speed reasons, they do — the disclosure travels with the asset. If we generate concept imagery for a campaign, you know it, your team knows it, and the limitations (don't print the billboard at scale, don't enter it in awards, don't assume exclusivity) are written down. Sunshine is cheaper than surprise.

The economics, honestly

Will it make your brand cheaper? Slightly, at the edges. Moodboard weeks compress; variation drudgery evaporates; some photography budgets shrink. But the expensive parts of brand work were never the rendering. They're the workshop where two founders discover they've been describing different companies. The fourteenth naming round where the surviving name finally earns its keep. The decision that the palette is ochre, not teal, because of a sentence in the strategy you can point to.

Generation removes approximately none of that, because none of that was pixels. If a studio quotes you half price "because AI", ask which decisions they've stopped making and who'll be making them instead. The answer is usually: nobody, and you, later, at a rebrand.

We'd also note the arrow cuts both ways: as generated aesthetics flood every feed, the market value of visibly human craft rises. The slightly imperfect hand-set wordmark reads more premium in 2026 than it did in 2022, because imperfection is now evidence of care. Illustration systems, commissioned photography, real lettering — these are appreciating assets, and the floor is rising.

What you'll get from us, in writing

Every Brassfern brand engagement now includes an AI position annex — one page, no legal Latin. It states which tools may appear in which phases, guarantees that final identity assets are hand-built or properly licensed, commits to the provenance notes described above, and gives the client the right to ask, at any point, "was this generated?" and receive a straight answer with evidence.

We think that page will be standard across the industry within three years, because clients will start demanding it — some already do. We'd rather have written ours before it was fashionable than after it was required.

If you're choosing a studio right now, steal the annex idea and ask your candidates to fill one in. The ones who answer quickly and specifically are the ones whose process you can trust. The ones who say "we use the best tools for the job" have told you more than they intended.

Key takeaways

• Use generation where output is disposable — moodboard divergence, name-storming, variation drudgery. Never where the output is the brand.
• Refuse three things: final marks, meaning-bearing choices (palette, position), and foundation voice work. All three fail on defensibility, not quality.
• Every asset carries a stated provenance: hand-made, licensed, or generated-with-disclosure. Provenance is what makes IP transfer honest.
• AI compresses rendering, not deciding — and deciding was always the expensive part. Suspicion is warranted when "because AI" halves a quote.
• Put the position in writing, one page, before the engagement starts. The market is heading here; be early.

FAQ

Do you use AI for client work without telling them? No — that's the entire point of the annex. The tools differ by project and phase, the disclosure doesn't. If a designer used a generated texture inside an exploration that was later redrawn, the notes say so. We'd rather be a little boring about this than ever answer "was this real?" with a pause.

Won't refusing generated final assets make you slower and pricier? Slower at render, about the same at everything else. The schedule of brand work is set by decision velocity — how fast a client can look at options and commit — and generation doesn't speed that up. Where it genuinely compresses time (exploration volume, variations), we use it and pass the savings on.

What about AI-assisted type design or lettering? Same rule as everything else: generating raw material a human then draws, cuts and kern-corrects is sketchbook work; shipping model output as a final mark isn't. The craft layer is where the asset becomes defensible and ownable.

How does this apply to your AI product work, like assistants and agents? Different domain, same spine: disclosure and defensibility. Building AI into products has its own set of positions — evals, fallback UX, telling users when a machine is speaking — covered by our AI practice and pieces like our responsible-AI review. The brand rule is narrower: the identity you ship must be something we can stand behind, asset by asset.