The Designer in the Age of AI: The Work Left When Execution Is Free

The quality designers recognise as "human" is, most of the time, just specification.

Two versions of the same empty state:

No items yet. Add your first item to start tracking your inventory.

Shelves are empty — for now. Add what you've got and we'll keep count from here. Takes about a minute.

The second reads like a person wrote it. It knows who is using the product and roughly when, it sets an expectation about effort, and it makes a small joke about the shelves. The first reads like it was assembled from the feature name.

Both came out of the same model, in the same session, eleven seconds apart. Shown the pair cold, designers pick the second as the human one on the same reasoning every time: warmth, tone, empathy.

The only thing that differed was the prompt. The second was given a specification — who the user is, the moment they will be reading this, what the product must never sound like. The first was given the feature name and nothing else.

So what reads as human is not authorship. It is whether anyone decided what the thing needed to be before it existed. That distinction now accounts for most of what a designer contributes, and the profession is still filing it under taste.

The tell we think we have

The instinct to spot the machine is not just weak, it is actively manipulable.

Kosch et al. (ACM TOCHI, 2022) ran two studies, N=369 and N=100, in which participants were told an adaptive AI system was helping them with a task. They expected to do better, and the expectation persisted after the work was finished. No AI was ever running.

Being told a machine was involved moved people's judgment before they had evaluated anything at all. Which makes "can you tell if this was AI?" the wrong skill to invest in. It is a question about provenance, and provenance is exactly the thing that is about to stop being legible.

The useful question is whether the work was specified — whether somebody decided what it needed to be before it existed.

What actually got automated

John Heskett wrote a line that sounds like wordplay until you take it apart: design is to design the design of a design.

Four different things hide in that sentence. There is design, the strategy — problem framing, empathy, alignment. There is to design, the labour — execution, software fluency, exporting assets. There is the design, the plan — information architecture, systems thinking, interaction logic. And there is a design, the output — wireframes, mockups, prototypes.

Historically most of a designer's working life was spent in the second and the fourth. The labour and the artifact. Those are precisely the two the machine now does, in seconds, at a quality that clears most professional bars.

What it does not touch is the first and the third. Strategy, and the architecture of the system.

That redistribution is already visible in how teams are organised. The AI in Design Report 2026 — Designer Fund with Foundation Capital, drawn from over 900 responses across 60-plus countries — puts it in one line: "Roles are blurring as designers take on PM and engineering work, and vice versa." The same report files "preserving judgment, taste, and skill development" under craft, which tells you what practitioners think is actually at risk.

Nielsen Norman Group frames the same shift as a move from designer-as-maker to designer-as-director, and names the failure mode that comes with it: "With the ability to build new functionality more quickly, teams may flood products with low-value features."

Their blunter warning is the one worth sitting with — "Design may become marginalized as UI execution is automated." That is the correct thing to worry about. It is also the wrong thing to get defensive about, because the marginalisation only happens if the profession keeps identifying with the half that got automated.

Speed without direction

Physics has the cleaner vocabulary here.

Speed is a scalar. It has magnitude and no direction. Torque is a vector — rotational force applied in a specific direction to move something heavy.

AI supplies effectively unlimited speed. A hundred screens a minute is unremarkable now. But a hundred miles an hour in the wrong direction is just a faster way to be wrong, and the machine has no opinion about direction at all.

Direction is the human contribution. One turn of judgment drives many turns of execution — an outer loop running at human speed, an inner loop running at machine speed, roughly one to N. That is the shape of the Two-Speed Engine, and design is where it shows up most plainly, because design was always the discipline that decided what was worth building before anyone built it.

Roberto Verganti, Luca Vendraminelli and Marco Iansiti reach the same conclusion from the research side, in Design in the Age of Artificial Intelligence (Harvard Business School Working Paper 20-091): "as creative problem solving is significantly conducted by algorithms, human design increasingly becomes an activity of sensemaking — to understand which problems make sense to be addressed."

Sensemaking is the word to keep. It moves the job upstream of the artifact entirely, and — as they argue — closer to leadership, which has always been sensemaking work rather than production work.

What the job becomes

If execution is no longer the contribution, four things are.

Choosing the right problem

A model has seen millions of checkout flows and knows exactly what one looks like. What it does not know is why people abandon them.

Ask for a high-converting checkout and you will get a sleek one-click layout modelled on consumer retail. Put that in front of a ten-thousand-dollar B2B software purchase and conversion collapses, because those buyers are not impatient — they are anxious. They need trust signals, security badges, an ROI calculator they can forward to whoever signs the cheque. Nothing in the training data surfaces that. It comes from having understood the person on the other side.

The costlier version of the same mistake happens one level up. A stakeholder says engagement is down, redesign the dashboard. The machine returns ten credible dashboards in a minute. You ship one and engagement keeps falling — because three user calls would have shown that people were not logging in at all, since the notification emails had been broken for a fortnight.

The dashboard was never the problem. The machine will never tell you that, because it answers the question it was given. When someone asks what we should build, the move is to reframe: what behaviour are we trying to create? AI makes it dangerously cheap to build precisely the wrong thing, beautifully.

Writing intent a machine cannot misread

We used to appraise an artifact by the labour it implied. Polish meant time, and time meant care. That heuristic is dead — polish is now three seconds old.

So the work shifts from producing artifacts to specifying them. "Generate a user profile card" leaves every real decision to chance. The specified version names what must exist:

Generate a UserProfile React component.
Must include 4 states handled via props:
1. isLoading (skeleton UI)
2. isEmpty (no avatar uploaded)
3. isPopulated (full data)
4. isError (fetch failed)

The second prompt is not longer for its own sake. It is the thinking the first one skipped, written down where it can be checked.

The stronger move is to specify what the machine may not do. Models are maximalists — given latitude they over-design and invent features nobody asked for. So set the guardrails first:

Generate a pricing table.
CONSTRAINT 1: Use strictly Tailwind spacing tokens.
CONSTRAINT 2: Must degrade gracefully to stacked cards on small viewports.
CONSTRAINT 3: Do not use external icon libraries. Rely on typography.
CONSTRAINT 4: Ensure contrast passes WCAG AA.

Constraints are the design. Vagueness returns the average.

Knowing what to check

When high-fidelity output arrives faster than anyone can review it, reviewing everything stops being possible. You have to know what to sample, what to trust, and what to wave through.

Models produce something close to the mathematical average of their training data, and the average is usually competent and slightly dead. The job is to look at ten variations and take the one that actually lands, which means rejecting the most average option rather than the least polished one.

It also means knowing what a render conceals. A translucent interface full of backdrop blur looks like a win until it meets a mid-range Android phone, where heavy CSS backdrop filters turn scrolling to treacle. The mockup cannot show that and the model does not know it, which leaves it to whoever has actually held the device.

And as synthetic media becomes indistinguishable from the real thing, "is this authentic?" stops being answerable. What replaces it is a question about intent: why this was generated, what behaviour it is built to produce, and who controls the worldview of the model that made it. I have written about where that way of looking comes from in The Algorithmic Gaze.

Designing the loop, not the artifact

Code is no longer the constraint, which means designers are absorbing the work on either side of them — and that requires knowing how systems behave.

Design a live-updating dashboard and the model will cheerfully assume data arrives instantly. If you know the legacy database behind it takes four seconds, you reject that interface and mandate optimistic updates and skeleton loaders instead. Without that knowledge you will specify something impossible and discover it in production.

The other half is human. A model can produce an adequate artifact but it cannot defend a tradeoff to the person funding it. "We need to fix onboarding, the cognitive load is too high" loses the room. "We are seeing a 30% drop-off at this step; recovering 10% of those users is about $50k in retained ARR this quarter" does not.

Above the artifact sits a stack, and the industry is preoccupied with its bottom rung — writing a better prompt. Context design supplies the memory: brand, tokens, prior decisions. Harness design sets the boundaries: guardrails, accessibility constraints, fail states. Loop design shapes the cycle where a user acts, the system reasons and the interface adapts. Architecture design wires services and actors together toward an outcome. Value compounds as you move up it, because each layer decides the behaviour of everything below.

What that work looks for are the gaps a machine cannot see — where the loop breaks, whether an automated decision is safe for the person receiving it, whether something reads as local or as homogenised global sludge.

The average is the default

Give a model a generic brief and it returns a competent, forgettable result. Ask for a landing page for a fitness app aimed at busy professionals and you get a dark interface, a photograph of someone running, and a headline about unleashing potential. It resembles every other app on the market because it is, definitionally, the average of every other app on the market.

No amount of better tooling fixes this. It takes the two pillars, used deliberately: give the machine the context and the emotional stake it has no access to, then constrain it hard enough that the cliché is unavailable.

Consider onboarding for a tool that helps restaurants manage inventory, where the people using it are exhausted chefs doing this at two in the morning after a fourteen-hour shift. Everything interesting about that product lives in the second half of that sentence, and none of it survives a generic prompt. Someone has to know it and write it down.

What is actually scarce

The four things above have one property in common: each is a decision made before the machine runs, or a judgment applied after it stops. None of them is production.

That is the uncomfortable part for a profession that has spent thirty years getting very good at production, and building its hiring, its portfolios and its status around craft in the tool. Craft in the tool looked like a moat for thirty years because it was expensive enough, for long enough, to pass as one.

AI handed the tools to everyone. What it did not hand over was knowing which problem was worth solving, what "good" means for this particular person in this particular moment, and whether the plausible thing on screen is also the right one.

---

A version of this argument was delivered as a session at Vimani, August 2026. The slides are here.