Not a question of ideology.
In 2026, three years after the first ChatGPT and a year and a half after Claude 3.5, the "use it or not" conversation is over. Use it. The open question is where exactly and how.
In studio practice this isn't a question of ideology but of craft. What AI does better than a human, what worse, and what it shouldn't do at all — these aren't philosophical categories, they're concrete tasks we've been testing assistants on for the last two years.
This note is about the practical rules we've arrived at. Not "how it should be done," but "how we do it." Other studios may differ.
What we trust it with.
The list of where AI assistants work regularly for us in the studio:
- Enumerating code options. "Show me three ways to implement this animation" — the assistant gives three, a human picks. It saves an hour or an hour and a half on routine tasks.
- Searching a large codebase. "Where do we define the token
--pink-300?" — the assistant finds in a second what Find/Replace would take five minutes for. - Generating alt text for images. Upload an image, the assistant writes a description. A human checks accuracy, adjusts tone. Saves a lot of time on projects with dozens of illustrations.
- Typography checks. Guillemet quotes, non-breaking spaces before short words, an em dash instead of a hyphen. The assistant passes through the text and proposes edits. A human accepts or rejects.
- Transcribing voice notes. A voice message from a meeting turns into text that's later convenient to search. Accuracy — about 95%, enough for our case.
- A first draft of documentation. "Describe what this component does" — the assistant writes the base, a human rewrites. The thirty minutes saved per component across a project add up to a day.
What we never trust it with.
The list of things we don't hand to an assistant, under any circumstances:
- Brand concept. What a brand should say, its tone, the associations it evokes — that's the work of a head that knows the client, their market, their history. The assistant gives an averaged answer that will look like "a brand from 2025."
- Copy for slides, sites, emails. Every letter a client or their clients see is written by people. AI-generated text is recognizable in the first line — even when grammatically perfect.
- Final design decisions. The assistant can propose options, list references, offer critique. But "we're keeping this option" is a decision made by the designer whose name is on the project.
- Architectural choices. Which stack, which data model, which render pattern — decisions three to five years out. The assistant knows "the average recipe." That recipe may be wrong for a specific client.
- Emails to clients. When you write to a client, you write with your own authority. The assistant can't borrow it. An email written by an assistant can be felt by touch, and it undermines trust.
- Presentations. The scenario of a meeting, the phrasing, the argument — that's the part of the work that distinguishes a studio from any assistant. It can't be automated without losing meaning.
The boundary shifts.
A year ago we wouldn't have trusted it with a first draft of documentation either — you could tell the text "wasn't ours." Six months ago — alt text; the assistants wrote too dryly, without context. Now we trust it with these tasks — because the quality has become good enough that a human spends minutes on edits, not hours on rewriting.
What shifts in six months, we'll see. Perhaps a first draft of site copy (with serious human editing) will move over. Perhaps generating typographic options ("show me this heading in six typefaces"). Perhaps not; some things will stay manual, because they should.
AI is an accelerator. Not a source of style.
The boundary doesn't shift on its own — we shift it. Each time — after testing a new capability on a real task and confirming the quality is good enough. Without tests — we don't move it.
This site is an example.
This site was made with AI's involvement. The code — partly, with review. Every component, every function went through a human who understands what a line does, and why it's there.
The content — by people, entirely. Every essay in this journal, every text on the home page, every phrasing in the brief — written by hand. Somewhere the assistant helped place commas; somewhere — with a typography check. But the thought and the phrasing are human.
The design — by people, entirely. Every layout, every color decision, every typographic choice — made by a designer whose name is on it. The assistant doesn't enter the design process, except as "show me references for this technique."
If it were otherwise, it would be noticeable. A site with generated text reads through in two minutes. A site with "prompted" design reads even faster. It isn't a question of morality, it's a question of recognizability.
What we tell the client.
We don't hide that we use assistants. At the first meeting, if the question comes up, we answer directly: "yes, on routine tasks we use it, on creative ones no — everything you'll see is made by people." That relieves the anxiety part of clients currently carry.
Sometimes the client proposes it themselves: "let AI write the copy, it's cheaper." We answer: "technically — possible, but in the premium segment it'll be noticeable, and the effect on the brand is negative." In most cases the client agrees. In a few — no; then we politely decline and recommend a studio that works in a different format.
Transparency on this question is part of the product. Not "we do everything by hand" (rarely true now) and not "we're progressive, it's all AI" (that's marketing). But — "here's the boundary, we observe it, and it's roughly this."