Thinking / 01 Design × AI

AI can do the making.
What should we
still own?

Working with an AI agent made me rethink the role of a senior designer — and the parts of the work I’m not willing to hand over.

Words by Chenbiao Zhang

A personal reflection · 6 min read

01 / The first attempt

This is the first time I’ve seriously worked with an AI agent, and it has made me think much more deeply about the role of AI — and the role of a senior designer.

This is the first time I’ve seriously worked with an AI agent, and it has made me think much more deeply about the role of AI — and the role of a senior designer.

AI can already do an extraordinary amount. It can explore websites, generate code, create frames in Figma, build responsive pages in Framer, and turn rough ideas into something that looks surprisingly finished. That creates an uncomfortable question: if AI can increasingly do the making, what work is actually left for designers to own?

AI can already do an extraordinary amount. It can explore websites, generate code, create frames in Figma, build responsive pages in Framer, and turn rough ideas into something that looks surprisingly finished. That creates an uncomfortable question: if AI can increasingly do the making, what work is actually left for designers to own?

At first, I approached AI in the most obvious way. I gave it a few keywords from one of my portfolio stories and asked it to generate the full story for me.

At first, I approached AI in the most obvious way. I gave it a few keywords from one of my portfolio stories and asked it to generate the full story for me.

It did.
And it was wrong.

The output was long, boring, badly ordered, and sometimes pushed the story in a direction that did not reflect what I had actually experienced. I kept giving feedback, asking for revisions, and trying to correct the structure. But the more I iterated, the more exhausting it became. It felt as if the connection between my brain and the AI had broken.

The problem was not simply that the AI produced a poor answer. The problem was that I had asked it to reconstruct an experience it never had.

Before / AI-generated story

Keywords in.
Assumptions out.

A few prompts became a complete narrative. The sequence, emphasis and direction did not reflect what happened.

Keywords → Guessed context → Revisions

After / Human-owned source story

Experience in.
A clearer story out.

I defined the events, tensions and decisions first. AI helped organise and express the story; I reviewed what survived.

My context → AI refinement → My judgement

02 / A different starting point

First, own the story.

I realised that I needed to own the source story myself.

I realised that I needed to own the source story myself.

Instead of giving AI a handful of keywords and asking it to guess what happened, I went back to the actual experience: the sequence of events, the tensions, the decisions, the dead ends, the moments that changed my thinking, and the outcome.

Instead of giving AI a handful of keywords and asking it to guess what happened, I went back to the actual experience: the sequence of events, the tensions, the decisions, the dead ends, the moments that changed my thinking, and the outcome.

I used Miro for this because it allowed me to work in the way I naturally think — with stickies, flows, milestones and connections rather than polished paragraphs.

I used Miro for this because it allowed me to work in the way I naturally think — with stickies, flows, milestones and connections rather than polished paragraphs.

I also abstracted the project into non-confidential decisions, tensions and milestones rather than sharing detailed internal content.

I also abstracted the project into non-confidential decisions, tensions and milestones rather than sharing detailed internal content.

That gave me a new collaboration model.

That gave me a new collaboration model.

01

Mind map

Own the source.

Context, tensions, decisions.

02

Refinement

Find the thread.

Challenge structure and order.

03

Visualisation

Make it tangible.

Explore, compare, express.

04

Review

Decide what survived.

Correct, question, take responsibility.

Correct, question,

take responsibility.

03 / A better collaboration

More useful.
Still mine to judge.

The important change was simple: AI stopped generating the story for me. It started helping me interpret and express a story I had already defined.

The important change was simple: AI stopped generating the story for me. It started helping me interpret and express a story I had already defined.

Once the source material became clearer, the quality of the collaboration changed dramatically. I could give AI a much richer picture of the project without asking it to invent the missing context. It could help me identify the strongest tension, reorganise the narrative, challenge the order, and suggest where the story could be explained visually rather than through another paragraph.

Once the source material became clearer, the quality of the collaboration changed dramatically. I could give AI a much richer picture of the project without asking it to invent the missing context. It could help me identify the strongest tension, reorganise the narrative, challenge the order, and suggest where the story could be explained visually rather than through another paragraph.

Then the role of AI shifted again. Instead of only helping with words, it could explore layouts, visual hierarchy, interaction ideas and implementation. Different design directions could be produced quickly, compared and refined. A page that might previously have required many rounds of manual production could become tangible much earlier.

Then the role of AI shifted again. Instead of only helping with words, it could explore layouts, visual hierarchy, interaction ideas and implementation. Different design directions could be produced quickly, compared and refined. A page that might previously have required many rounds of manual production could become tangible much earlier.

But this only worked because I still had the final say.

But this only worked because I still had the final say.

I could look at an output and say: this sounds impressive, but it is not what happened. This section is too dramatic. This visual is beautiful, but it is explaining the wrong thing. This part needs more evidence. This one needs to disappear.

I could look at an output and say: this sounds impressive, but it is not what happened. This section is too dramatic. This visual is beautiful, but it is explaining the wrong thing. This part needs more evidence. This one needs to disappear.

The speed of production increased,
but the responsibility for judgement did not.

The speed of production increased,
but the responsibility for judgement did not.

04 / The designer’s role

From owning every output
to owning the direction.

My first answer was: bridge. I still think that word is useful, but I now think the role is slightly more active than that.

A senior designer increasingly becomes the person who connects context, intent, judgement and execution. AI can participate in the thinking. It can propose, synthesise, visualise and build. But it does not automatically know which trade-off matters, which detail is politically sensitive, what the product is really trying to achieve, or whether the result is something I would put my name on.

A senior designer increasingly becomes the person who connects context, intent, judgement and execution. AI can participate in the thinking. It can propose, synthesise, visualise and build. But it does not automatically know which trade-off matters, which detail is politically sensitive, what the product is really trying to achieve, or whether the result is something I would put my name on.

So perhaps the shift is not from designer to AI. It is from owning every output to owning the direction, judgement and standard of the output.

There are things I am increasingly comfortable asking AI to help with: synthesis, alternatives, visual exploration, repetitive production and implementation support.

There are things I am increasingly comfortable asking AI to help with: synthesis, alternatives, visual exploration, repetitive production and implementation support.

There are also things I am much less willing to outsource: context, intent, trade-offs, taste and accountability.

There are also things I am much less willing to outsource: context, intent, trade-offs, taste and accountability.

The designer may not need to manually create every pixel anymore. But the designer still needs to understand why those pixels exist.

The designer may not need to manually create every pixel anymore. But the designer still needs to understand why those pixels exist.

05 / Beyond my own workflow

Using AI is one thing.
Designing AI is another.

Working with AI in my own process also made me realise that using AI effectively is only one part of the challenge.

When AI becomes part of the product itself, the design problem changes. A traditional interface is usually expected to behave predictably. AI is different. It can be helpful, but it can also be uncertain, incomplete or simply wrong.

When AI becomes part of the product itself, the design problem changes. A traditional interface is usually expected to behave predictably. AI is different. It can be helpful, but it can also be uncertain, incomplete or simply wrong.

That means designers need to think about more than the ideal output. We need to ask:

How confident is the system?

  • How does the user know when to trust it?

  • What happens when it is wrong?

  • Can the user correct it?

  • Can they override it?

  • What should happen when the system does not know enough?

These are not only technical questions. They are experience questions.

These are not only technical questions. They are experience questions.

The more I work with AI, the more I think the role of the designer is not just to make AI feel intelligent. It is to make the relationship between human judgement and machine judgement understandable.

A practice, still evolving

That is the collaboration model I am experimenting with now. I bring the lived context and the point of view. AI helps me challenge, structure, visualise and produce. Then I review, correct and decide what survives.

That is the collaboration model I am experimenting with now. I bring the lived context and the point of view. AI helps me challenge, structure, visualise and produce. Then I review, correct and decide what survives.

In fact, this portfolio — including the page around this article — is being built through that process.

AI is assisting the work.
The judgement is not being outsourced.

AI is assisting the work.
The judgement is not being outsourced.

HMI Product Owner / UX designer

Delivered 2026

Duration 1 year, 3 months