Taste Is the New Bottleneck: What Designers Contribute When AI Makes Production Cheap

A designer can now describe an interface, attach a reference, and receive a working direction in minutes. A marketer can produce dozens of campaign variations. A founder can prompt an interactive prototype without waiting for a conventional design-to-development cycle. Anthropic’s analysis of 500,000 coding interactions found that user-facing work was already a major use case: UI/UX component development represented 12% of the coding conversations it studied, while web and mobile app development accounted for another 8%. Figma’s launch of Make and Sites put prompt-to-code prototyping and direct web publishing inside a mainstream design platform.

This is a real expansion of creative capacity. It is also producing a predictable side effect: an abundance of work that looks finished before it has been properly judged.

When almost anyone can generate a polished dashboard, landing page or brand concept, surface-level competence stops being a meaningful advantage. The new bottleneck is taste, not taste as personal preference, but taste as the ability to make coherent, defensible decisions under constraints.

That is where designers become more important, not less.

Production abundance creates an “average” problem

Generative tools are excellent at producing plausible combinations of familiar patterns. Ask for a fintech dashboard and you will probably get cards, charts, a sidebar, reassuring blues and a clean sans serif. Ask for a premium wellness brand and you may get muted neutrals, elegant type and spacious photography. None of this is necessarily bad. The problem is that plausibility is not the same as relevance.

AI tends to begin near the centre of the distribution: the patterns that appear often, combine easily and read instantly as “professional.” This is useful for getting off a blank page. It is dangerous when teams mistake that starting point for a product decision.

Average-looking software is locally polished but globally weak: unclear priorities, interchangeable copy, borrowed category styling, missing edge cases and no distinct position.

The cost of generating another option has collapsed. The cost of choosing well has not. In fact, the review burden can increase because teams now have more outputs, produced faster, with enough polish to discourage hard questions.

Figma CEO Dylan Field captured the shift in the company’s Config 2025 recap: in a world where AI makes software easier to build, “craft, quality, and point of view” are what make a product stand out. Figma’s 2026 research reinforces the point. It reported that 91% of surveyed designers said AI tools were helping them improve their work, while 82% of hiring managers said their need for designers had remained stable or increased.

The implication is not that AI has failed. It is that speed at the production layer makes judgment at the decision layer more valuable.

Taste is not “I know what looks good”

The word taste can sound vague, elitist or purely aesthetic. Product teams need a more useful definition.

Taste is the ability to recognise and choose the most appropriate expression of a product’s intent.

It combines context, priority, coherence, standards and point of view. That makes taste defensible: a designer should be able to explain why a calm interface suits a high-stakes financial action, or why a clever interaction should be removed because it obscures the primary task.

Good taste is not the ability to decorate an arbitrary decision. It is the ability to reject the wrong decision, even when the output is beautiful.

The designer’s role is moving from maker to editor, director and system owner

Designers will still make things. But as generation becomes cheaper, their highest-leverage contribution shifts upstream and across the system.

The designer as editor

An editor separates promising material from noise. With AI, that means reviewing options without becoming impressed by quantity or fidelity. The designer asks:

• Does this solve the stated problem or merely illustrate the prompt?

• Which part is specific to our users, and which part is borrowed convention?

• What can be removed without weakening the experience?

• Is the apparent quality coming from visual polish or from a strong underlying decision?

Editing also means knowing when not to generate again. Infinite variation can become avoidance. Sometimes the right move is to choose one direction and resolve it deeply.

The designer as director

A director supplies intent. Prompts help, but direction is larger than prompt writing. It includes selecting references, defining principles, sequencing exploration, preserving constraints and giving useful critique.

“Make it modern” is weak direction. “Reduce the visual anxiety around this payment decision; keep the fee visible; make the primary action obvious without making the user feel rushed” is product direction.

The designer is not competing with the model on output volume. The designer is creating the conditions under which the output can be judged.

The designer as system owner

AI exposes weak systems quickly. If a team lacks clear tokens, component rules, content principles and accessibility requirements, generated output will drift. One screen may look acceptable while the product becomes less coherent with every iteration.

A strong design system now does more than help humans reuse components. It supplies boundaries that AI tools can follow: semantic color roles, approved patterns, component states, tone rules, responsive behaviour and examples of what not to do.

When the production engine accelerates, the system must become more explicit.

A six-part rubric for judging AI-generated design

Before accepting generated output, score it against six criteria. A simple 1–5 scale is enough, but require a written reason for every score below four.

1. Relevance

Does the solution address a verified user need and the real context of use? Test it against a concrete scenario, not a generic persona. A visually strong dashboard is irrelevant if the user needs one quick answer on a low-end phone with poor connectivity.

2. Hierarchy

Can a user tell what matters, what happens next and what is secondary? Review the interface at a glance, in grayscale and on a smaller screen. If every card, headline and button asks for equal attention, the design has no opinion.

3. Coherence

Do interaction patterns, visual language, content and behaviour feel consistent? Check repeated actions, states and terms. AI often creates local solutions that do not belong to the same product.

4. Distinctiveness

Is the work recognisably ours, or could a competitor use it unchanged? Distinctiveness does not require novelty everywhere. It can come from a clear content voice, a particular use of motion, a confident information model or one memorable visual behaviour.

5. Accessibility

Can people perceive, understand and operate the experience across different abilities and input methods? WCAG 2.2 remains the current W3C recommendation and explicitly combines testable criteria with human evaluation. Check contrast, focus order, keyboard use, target size, labels, errors, reduced motion and assistive-technology semantics. Accessibility is not a cleanup phase; it is part of quality.

6. Business fit

Does the design support the product’s promise, operating model and level of risk? A pattern that increases conversion but creates misunderstanding may be commercially harmful. A delightful flow that requires support the company cannot provide is not viable design.

A generated direction should not move forward because its average score is high. Some categories are gates. If accessibility, relevance or business fit fails, the option fails.

How teams can build taste instead of merely hiring for it

Taste grows through exposure, articulation and repeated decisions. Build annotated reference libraries that explain why an example works, not just how it looks. Write product-specific principles that resolve trade-offs: “show the cost before commitment” is more useful than “keep it simple.” Separate exploration reviews, where range is valuable, from decision reviews, where evidence is required. Finally, record why promising directions were rejected. Over time, those reasons become a practical statement of the team’s standards.

A practical AI design review checklist

Before a generated screen, flow or brand asset is approved, confirm:

• The user, task and context are specific.

• The work solves the problem rather than merely matching the prompt.

• The primary action and information hierarchy are clear.

• Copy uses the product’s language, not generic AI phrasing.

• Components and states follow one coherent system.

• Error, empty, loading and recovery states exist.

• Keyboard, contrast, focus and screen-reader basics have been tested.

• The direction has at least one meaningful point of distinction.

• Business, legal and operational constraints are visible in the solution.

• A named human owns the final judgment.

  
  

The valuable designer is not the fastest prompter

AI will keep improving at execution, making polished artefacts less useful as evidence of quality.

The designer’s durable value is not possession of a particular tool. It is the ability to connect user reality, product strategy, visual craft, system behaviour and ethical responsibility into one coherent decision.

AI can generate options. It cannot be accountable for what a company chooses to release. It does not own the promise made to the user, the accessibility standard, the support burden, the brand reputation or the consequences of a misleading interaction.

Production is becoming abundant. Taste is the discipline that turns that abundance into something worth using.