If AI Can Make the Screens, Do You Still Need a Designer?

An engineer I work with said, half joking, that designers are about to become unnecessary. The logic holds up. Ask a model and you get screens and flows in seconds. If the role is drawing the thing and handing it over, that role is finished.

For anyone commissioning the work, the useful question is the next one. What actually got cheaper, and what did not? Get that wrong and you will pick the wrong partner.

What got cheap is the making

Prototypes collapsed in price. A screen that took a week now runs in an afternoon. Nobody can charge a premium for showing you something.

Which means comparing quotes on the production effort alone is close to meaningless now. The numbers converge, and the thing that separates two vendors is no longer in that column.

What did not get cheap is the deciding

The expensive part sits in the middle of the product. The screen before any data exists. The wording when something fails. The page shown to someone who is trying to cancel. The moment you ask for money. And in an AI product, what appears when the model is not confident.

Most of that belongs to nobody, inside companies and outside them. It is also where trust breaks. It never breaks in the screens you show the board.

Can they throw twelve of fifteen options away?

Anyone can produce fifteen versions of a screen. That is not the hard part.

Three of them will not work for a Japanese user who needs internal consensus before deciding. One will move the numbers. One will destroy trust the first time the system gets something wrong. Being able to say which is which, with a reason, is the actual difference between two people with the same title.

You cannot estimate that in hours. The only real signal is what someone has broken before.

Something that happened in my own product

One of my apps had an upgrade flow that would not move. The screen was fine. Spacing, type, hierarchy, all correct.

I opened it one evening on my own phone, in the state a real user would be in, and it was obvious. I was asking for money before the product had given anything back. Not a research finding. A four second reaction.

I changed the order of the ask and shipped it the same night, and the numbers went up. The part that matters is not the fix. It is that the person who noticed could ship it that day. Routed through a spec and a review, that change happens next month, if at all.

Questions worth asking in a first meeting

Ask what they have thrown away, not what they have made. Which things did they decide not to build on a past project, and why. If nothing concrete comes back, you are talking to someone who has executed rather than decided.

Then ask about failure. Errors, empty states, cancellations, a lost connection. How those four were handled tells you immediately whether the person has ever owned the middle of a product.

Choose the verb, not the title

On teams building AI products the functions have already merged. Instead of designer, engineer and product manager, it is more accurate to look at what someone can do right now.

Inventing and discarding prototypes. Building something that survives production. Cleaning up, simplifying and removing what should never have shipped. Taking something live and growing it to the point where people pay. Running a mature system safely for years.

Before you hire anyone, decide which of those five your company is missing. The person you need at the start of a product is not the person you need once it is running, even when the job title on their profile is identical.

In Japan, mediocre stands out

Quality here is an assumption, not a differentiator. Anything half finished is noticed immediately, and nobody tells you. They just stop using it.

Now that making things is cheap, the volume goes up. The more volume there is, the more the feeling decides the outcome.

The short version

Designers did not become unnecessary. The role of drawing screens and handing them over did. What remains is deciding what not to build, and deciding how the product behaves when it is wrong.

That is also the part worth commissioning from outside. If you are still comparing quotes by screen count, you are comparing the wrong column.