AI and product design: what changes with agents in the workflow

For years, product design worked in a stable order: the designer thinks, the developer builds and the user validates. AI agents, systems that plan and carry out multi-step tasks on their own, are changing that order.

What I mean here are agents that read a design spec, write the component code, run accessibility tests and open a pull request for review, without anyone steering each step. Code copilots and image generators are a different conversation.

What changes in a designer’s routine#

The first change is operational. Tasks that used to take hours, like adapting screens to several breakpoints, documenting a component’s spec or generating layout variations, can now be delegated with a well-written request.

The amount of work stays the same, but the kind of work changes. The effort goes into:

• Quality criteria that the agent uses as a reference
• Critical review of what was generated automatically
• Design system architecture, so automation works without ambiguity
• Research and strategy, where agents don’t reach yet

The risk: mistaking automation for thinking#

The danger I see is handing AI exactly the parts of the work that depend on people: empathy for the people who use the product, judgment about context and the willingness to question a requirement before accepting it.

A good agent optimizes the problem you gave it. Knowing whether that is the right problem is still the job of whoever designs the product.

How I reorganized my work#

Three changes that are already part of my routine:

1. I document the intent along with the spec. Beyond measurements and colors, I record the reason behind each decision. That guides the agent and also whoever joins the team later.
2. I invest in a design system with names that explain themselves. Well-named tokens, components with explicit variants and patterns described in plain language make the system readable for an agent, without ambiguity.
3. I review what AI delivers the way I review any work from the team. It is a good starting point, and it goes through the same rigorous review.

An example of the first point. Along with the token goes a line saying when to use it:

/* Space between the label and the value of a KPI card.
   Use it only inside cards: between blocks, the space is --space-6. */
--space-2: 8px;

What stays#

After months of using agents day to day, my conclusion is that AI amplifies what you already do. If you think about the product strategically, you get more time for that; if you accept the brief without questioning it, AI only speeds up the delivery of solutions to the wrong problem.

The edge will be in knowing what to ask it for, and why.