Why AI Automates Instances, Not Designers

Most discussions about AI treat a job as a single block of work. You're either a "product designer" or a "design engineer", and either the model replaces you or it doesn't.

That view misses how design work actually happens. In their paper Economic Scenarios for Transformative AI, researchers look at production through an economic framework from Joseph Zeira and Daron Acemoglu, but with a sharper lens. Instead of looking at jobs or broad disciplines, the model breaks work down into task instances.

Tasks versus instances in design

The distinction is simple, but it changes how you look at your backlog.

A task is a category of work. Designing a settings page, building a data table, or running an accessibility audit.

An instance is one specific performance of that work. Writing the Tailwind classes for a secondary cancel button on Tuesday morning. Wiring up the loading skeleton for an analytics chart. Mapping the error states for a Stripe checkout form when a card is declined. Figuring out why a modal's focus trap leaks on iOS Safari.

AI doesn't show up and automate "the product designer". It arrives instance by instance. Today, you might let Claude Code scaffold three standard CRUD forms and write their basic CSS grid, while you manually architect a tricky drag-and-drop kanban board. In the economic model, those aren't lumped together under one title. They are separate entries in the production function, each with its own cost, speed, and quality.

The swappability dial

The paper models this with a single parameter: the elasticity of substitution. You can think of it as a swappability dial. It measures how easily an economy can trade a human instance for a machine instance without losing value.

When elasticity is high, the human and the model are near-perfect substitutes.

Think about the mechanical parts of a designer's week:

• Translating a Figma frame into flexbox utilities.
• Generating realistic customer names and amounts for table mockups instead of pasting "John Doe" ten times.
• Writing standard form validation copy.
• Checking WCAG contrast ratios across a dark mode palette.
• Turning an SVG icon into an accessible React component with focus rings.

Because a model runs these instances in seconds for pennies, substitution wins. If your entire workday is pushing pixels between specs, drawing standard forms, or writing repetitive component props, that work gets swapped out.

The low-elasticity bottleneck

The interesting math happens when elasticity is low. Here, human and AI instances are complements. They fundamentally need each other.

Low-elasticity instances are the decisions where context, taste, and user psychology dominate:

• Deciding whether a new feature belongs in a slide-out drawer, a modal, or a separate page.
• Designing an optimistic UI rollback so a failed mutation doesn't wipe out three paragraphs a user just typed.
• Figuring out the user's mental model: do they think of projects as hierarchical folders or flexible tags?
• Applying behavioral friction: knowing when an onboarding flow should slow someone down so they don't break their database setup.
• Pushing back on a founder who wants seven metric cards on the home screen, helping them pick the single number that drives the business.

When routine instances become cheap, the total cost of shipping software plummets. A startup that used to ship one major feature a month can now prototype and ship five.

That explosion in production runs straight into a wall: the remaining human instances.

You can't ship five times as many features without five times as much product strategy and user judgment. If you generate ten new dashboards in an afternoon, someone still has to sit with users, spot where they hesitate, and protect the coherence of the whole product.

Because the model can't do those things, the human instances become the scarce bottleneck. And in economics, whatever sits in the bottleneck becomes more valuable, not less.

From artisan to director

For a product designer today, this shifts where your value lives.

You stop measuring productivity by how many hours you spent drawing boxes in Figma or hand-crafting buttons from scratch. Manual pixel labor was an accident of our tools, not the point of design.

Instead, the job looks much more like being an art director and an architect. You spend your time where elasticity is low: framing the problem, establishing the visual rules, and deciding the interaction model. You delegate the high-elasticity instances to the model, review the output with a sharp eye, and stitch the pieces into something that feels intentional.

The mechanical execution is becoming free. What's left is taste, judgment, and knowing what's actually worth building.

Anthropic. (2026). Economic Scenarios for Transformative AI. Read paper.