You can use AI. So can your client.

Every few weeks someone tells me, kindly, that they can design now too. "Hey Muna, we can use AI and design." Good for you. Genuinely, I mean it. So can I do your job using just AI?

This is not a designer defending his turf from a chatbot. I run my entire practice through these tools, and I probably use them more than most of the people who say this to me. It is about which direction the thinking travels. Theirs goes from the tool to them. Mine goes the other way round.

The mind-reading workflow

Here is the method, as it is usually described to me. Open Lovable, or Claude, or whichever one is trending this week. Type a sentence. Wait for it to think what you were thinking. Ship whatever comes back.

And something always comes back. That is the trap. It produces the average of everything it has ever read about your request, and the average looks remarkably like what you had in mind, mostly because you had not had very much in mind.

Ask an assistant for a landing page with nothing to go on and you get a founder quote, a customer count, a percentage and three testimonials. None of it flagged, and every line a claim that nobody ever made. It is not lying. You asked it to read your mind, and your mind was empty, so it filled it.

I run it the other way round

The mobile app I am building right now is a ground-up rebuild of one its users already rely on every day. Rebuilt rather than repaired, because the old app's structure was the root of their complaints, and patching it would not have changed the order of their work.

It did not start with a prompt. It started with the product written down: personas, an information architecture, a sitemap, a document per module, and an open questions file where every unknown gets an owner and a due date. One rule sits over all of it. Product details are never inferred. They come from the product owner, or they become a question.

That rule reaches into places nobody posts about. The screenshots of the existing app carried people's names, email addresses and an API key. They stay in Figma, cited by layer and node, and the repository records only structure: titles, labels, fields, states. Paste those screenshots into a chat window and ask it to "make this better", and all of that goes with them.

Seven agents, and one of them is paid to say no

The build runs on a crew of seven AI agents, each with one job. One runs the backlog. One designs screens in Figma. One writes the code. One looks things up for it. One writes the notes and changelogs. One handles the small jobs. And one exists purely to check the designs and the code and send work back.

It does. Today it failed a build: one red, three orange, five yellow. The work went back, came back, and passed on the re-check with notes. Every successful job is its own commit, so any single step can be undone without touching the rest.

The agents are fast. None of them decides what the product is.

Decisions, written down

In under two weeks, 72 product decisions have been logged, each with a date, an owner, a reason and the alternatives it beat. One of them says every module is designed once, after a single map of the whole app, so a screen that has been approved is never redesigned because a later feature arrived. It lists nine alternatives it rejected.

The people who will use it every day review a working build every two weeks, not pictures of one. In the last review, one of them suggested reorganising the main screen, then saw the running build and withdrew the suggestion on the call. Another asked to set prices in one more place. I said no, on the call, because a price you can change in several places is a price nobody can remember changing.

The model did not make that call. Neither did the client. I did, and it is written down with the reason.

So the model never has to guess what I am thinking. The thinking went in. The typing came out.

So, about your job

Here is the part I promised was sarcastic.

"We can use AI" is not a skill. It is a subscription. And the uncomfortable thing about a subscription is that your client can buy one too, for the same price, this afternoon.

If your contribution is the sentence you type into the box, then your contribution is the one part of the job the person paying you can already do without you. You have not automated design. You have automated yourself, and politely handed over the login.

So yes, I am happy to do your job using just AI. I already have the account.

If you want a quick test, take the tool away for a moment and look at what is left that you brought. The questions you asked. The decisions you wrote down. The request you turned down, and why. The thing you refused to let it invent. If that list is empty, the tool was not helping you do the job. It was doing it.

The honest part

Writing it all down does not make it right. It makes it visible, which is not the same thing.

Of the 50 questions in that open questions file, 45 are still open. The demo stalled on search in front of the users, in the middle of the walkthrough. And the one hour I spent watching them use it produced eight new decisions that none of my documents, and none of my seven agents, had thought of.

The crew is not free either. It runs on a plan with a monthly limit, and until yesterday the agent running the backlog woke up several times an hour with nothing to report, spending as it went, until I rebuilt it to sleep until there was news.

And it is slower to the first screen than typing a sentence. That is the point, and it is still a cost.

What I would tell anyone who says it to me

Keep using it. I am not asking anyone to stop. I am asking you to point it the other way.

Do not ask the model what you think. Tell it, in writing, before the first prompt, and then give one of your agents the job of telling you no. Write the questions before the screens. Write the decisions before the code. Put a working build in front of real people before you believe any of it.

AI did not make design easy to replace. It made it very easy to see who was only ever doing the typing.