DESIGN INTELLIGENCEJuly 19, 2026·Mary · DEPIX Design Intelligence

Taste Is the Bottleneck: What Generative AI Actually Changes in the Concept Phase

Every headline about generative AI in the design studio sells the same promise: speed. Kia's global design team describes using it to ideate and innovate faster. NVIDIA frames it as revving up a new age across design, engineering and production. A widely-cited automotive case study clocked a designer generating dozens of distinct concept forms in a single 30-minute session. The pitch is intoxicating and mostly true: the raw cost of producing an option has collapsed to near zero.

And that is exactly why speed is the wrong thing to celebrate. When the cost of making an option falls to zero, the scarce resource doesn't disappear — it moves. It moves to the one step the model has not automated: deciding which option is right, and why. The bottleneck was never generation. It was judgment. We simply never noticed, because generation used to be so expensive that it hid the real constraint.

This is the quiet consensus forming across the field in 2026, and the design-criticism world has started naming it from every direction: "taste is the new bottleneck," the year of the "taste economy" where curation becomes the act of creation. When AI can produce infinite variants in minutes, the work becomes selection, rejection and transformation. The designers ACM interviewed for its mid-2026 issue reported the same discovery: their tools are superb at generating and refining, and poorly suited to the actual job — comparing, sorting, annotating and evaluating huge variation sets. The industry built a fire hose and forgot to build the filter.

There is a deeper reason taste refuses to automate. Large models are trained on the average of everything, so they return, by default, average excellence — polished, competent and eerily generic. Everything looks good; everything looks the same. A generated form can be flawless and still miss the emotional nuance, the cultural read, the brand personality that makes it that company's and no one else's. Closing that gap is not a rendering problem. It is a taste problem — and taste is precisely the thing a designer spends a career building, through criticism, reference and a thousand small rejections.

So the interesting question is not "how much faster does AI make the concept phase?" It is "what should we do with the time it gives back?" Here is the contrarian answer: the concept phase should not get shorter. It should get denser. Because the concept phase was always the highest-leverage moment in the entire pipeline. By the time a design leaves concept, roughly 70–80% of a product's total cost is already committed — locked into geometry, architecture, material and proportion. A change that costs a few hundred euros of studio time at concept can cost fifty times more after tooling, and formal studies of engineering change put late modifications at 5 to 100 times the early cost. Architects live on the same curve: the cheapest moment to change a building is before it is drawn. Everything downstream is a negotiation with a decision made early.

Read those two facts together and the strategy writes itself. AI removes the cost of exploring options; the concept phase is where exploring options matters most and costs least to act on. The rational move is not to sprint through concept on the time AI saved — it is to reinvest every reclaimed minute back into it: explore more branches, pressure-test more proportions, kill more ideas earlier, and arrive at the tooling gate with a decision that has survived more scrutiny, not less. Use AI to expand the option set ruthlessly wide; then use human judgment to compress it ruthlessly narrow.

That reframes what a designer — and a design leader — is actually for. Not the hand that produces the render, but the taste that authors the direction and the nerve to decide what is non-negotiable. The generative-ideation research bears this out: the systems that work best are explicitly human-in-the-loop, built to widen a person's imagination rather than replace their verdict. The tool proposes; the designer disposes. The org chart that wins in 2026 is not the one with the most seats at the render farm — it is the one with the clearest taste at the concept table.

This is the whole thesis behind how we think at Depix. The point of collapsing the cost of visualization is not to make the concept phase a formality you rush through; it is to make it the richest, most-explored, most-decided moment in the process — the place where a human's taste does the work only a human's taste can. Ideas are becoming free. Knowing which one is right never will be. Design the concept phase around that, and AI stops being a speed gimmick and becomes what it should have been all along: the fastest route to a decision worth committing to.

Sources: