How a studio can use AI in design without handing over judgment

AI in design process can widen the set of options and speed the chores that do not require taste. The studio still decides what is clear, what is on brand, and what ships. That split is the whole practice. The tools change. The split does not.

Judgment is what a client hired. A model can lay out ten headlines before lunch. It cannot know which sentence the company is willing to say in public, or which screen a stranger will understand in a few seconds. Those decisions stay with the people whose names are on the work.

Use it to widen options, then close them

Early in a project, a narrow first idea is a risk. Everyone falls in love with the version that was drawn on Tuesday. A model is useful here as a way to see more directions quickly: three layouts, a handful of headlines, a darker and a lighter pass at the same page. The point of the pile is the contrast, not the volume.

Treat the pile as private sketching. Label it as generated so nobody confuses it with a recommendation. Then do the actual design move, which is throwing most of it away. If everything survives, you did not use the tool to think. You used it to avoid a decision.

Bring the surviving option into the tools where the work becomes real. Hanabi sketches structure in Figma and builds in production code, with Next.js, React, and TypeScript on the sites and product interfaces the studio ships. A generated picture of a homepage is not a homepage. Type, spacing, and real sentences still have to be set by someone who will look at them on a phone.

Ask the model for options that disagree with each other. Ten variations of the same centered headline are not options. A quiet page, a page that leads with the product, and a page that leads with proof are options. You can then choose. Choice requires difference.

Speed the chores that do not need taste

Some work is real and still not judgment. Resizing a set of exports. Drafting alt text from a described frame so a person can correct it. Turning a decided palette into the tokens a codebase will use. Summarizing a long kickoff note into the decisions, with the recording still there to check.

These chores are safe to speed up when a person can tell, quickly, whether the output matches a decision already made. Alt text is a good example. The person knows what is in the picture. The draft either says that or it invents a mood. The correction is obvious. Choosing the picture was the judgment. Describing it is the chore.

Unsafe chores wear a costume of productivity. "Generate the brand voice." "Design the system." "Write the homepage." Those are the decisions. A first draft can exist, and a person has to be willing to discard it. If the team feels obliged to ship the draft because it arrived fast, the tool is setting the quality bar. That is backwards.

Keep generated code in the same review as any other code. Types, names, and the way the page fails are still someone's responsibility. Speed in the editor is not a reason to skip the pass you would have made on a hand-written component.

Three decisions that stay with the studio

Clarity is the first. Nielsen Norman Group's 2011 article on how long people stay on web pages found that users often leave within 10–20 seconds, and that you have about 10 seconds to communicate a value proposition if you want them to stay longer. A generated headline does not get a pass on that clock. Someone still has to read the first screen the way a stranger will, and cut it until the offer is visible.

Brand is the second. The sentences the company will say, the sentences it will not, the type, the color, the photographs you refuse. A model has seen every brand and will happily average them. Averaging is how a young company starts to look like a template. The studio's job is to notice the average and decline it.

Craft is the third, and it includes access. The W3C's WCAG 2 overview is the reference for web content, and it explicitly covers AI web interfaces. The principles are perceivable, operable, understandable, and robust. A generated component still needs contrast, a visible focus state, and a name on the control. Using the guidelines as a quality bar is not the same as claiming a formal certification. It means "done" includes those checks, whether a person or a model drew the first rectangle.

Write the three decisions into the critique, every time. What is unclear. What drifted off the brand. What would fail a keyboard or a rushed read. If the critique is only "make it pop," the judgment has already been handed over, and the model will pop on command.

How critique changes when the draft was generated

Say out loud that the draft is disposable. People get polite around a thing that looks finished. A generated screen often looks finished too early, which makes the room hesitant to cut it. Start the critique by naming the job of the screen, then ask whether this draft does the job. Politeness toward a machine is not a virtue.

Separate "the model did something odd" from "the direction is wrong." The first is a cleanup. The second means you pick a different option or you draw the thing yourself. Cleaning a wrong direction is how teams waste the time the tool was supposed to save.

Keep client review on the work you are willing to defend. Showing a raw pile of generations and asking the client which they like turns the client into the art director of a slot machine. You can show two directions and the reason you prefer one. That is a studio doing its job.

If the client asks whether a model made it, answer plainly. Then talk about the decisions. The origin of the sketch matters less than who checked the offer, the brand, and the build.

What never goes out under the client's name

Do not ship a claim the studio did not check. Generated copy invents numbers, customers, and awards with a straight face. Those are not drafts to polish. They are sentences to delete. The public site can talk about real projects, such as the work already on Hanabi's own pages. It cannot talk about a metric a model wished into a paragraph.

Do not ship a face, a logo, or a photograph the client does not have rights to use. A generated image can look like a real person. That is a legal and a human problem, not a style problem. Use pictures you directed or licensed. Use marks you drew.

Do not let the model talk to the client unsupervised. A chatbot on the studio's behalf will promise scope. Scope is a judgment. It belongs in a conversation a person will remember.

The hours you save on chores should show up as more time on those decisions, not as a thinner review. If the review is getting shorter because the draft "looked done," stop and spend the saved hour on the first screen.

Hanabi's services are still design and production code, with a person accountable for what ships. AI can sit inside that practice. It does not replace the critique.

How to run that critique in a small team, including a shape a founder can host, is the next piece. How to run a design critique in a small team keeps the conversation on whether the work does the job.