Where I keep AI out of my process

I use AI tools every day. They draft my first pass at making sense of research, suggest layout ideas I’d never bother to draw by hand, build quick throwaway prototypes and tidy up my notes after a workshop. My process is noticeably faster than it was two years ago.

But there are three places I’ve deliberately kept AI out. I think those choices say more about how I work than the tools I use.

1. Talking to users

I’ll use AI to help write the questions for an interview, or to tag a transcript afterwards. I won’t use it instead of the conversation. And I don’t trust “synthetic users” (AI pretending to be your customers) for anything that matters.

The useful moments in research are the ones nobody planned for. The pause before someone admits they don’t understand their own pension. The person who describes a workaround involving a sticky note on their monitor.

AI learns from what people usually say, so it gives you what people usually say. The job is finding out what this particular group of people actually does.

My rule: AI can speed up the paperwork around research. The listening stays with me.

2. Taste calls

When I’ve got twelve AI-generated versions of a pricing page, the hard part isn’t making a thirteenth. It’s choosing. Which one fits the brand, matches what we learned in research, and won’t confuse someone on a small phone with a bad connection?

I’ve tried asking AI to rank the options. It gives a confident, well-organised answer that sounds like every design blog post ever written. That’s not exactly wrong, but it drifts towards the safe middle.

Picking the direction is the part I’m paid for.

So I do it myself, and I write down why.

3. The last 10%

Spacing that’s nearly right. An empty screen that’s technically there but says nothing useful. A page that works with a keyboard but trips up a screen reader (the software blind and partially sighted people use to hear a page read aloud). A button label that’s accurate but slightly too long once it’s translated into German.

AI tools are good at getting you to 90% quickly. They’re poor at noticing the last 10%, because that part is about care rather than making more. It’s also the part users feel without ever being able to name it.

I still do it slowly and by hand, and I test it with the real accessibility tools people use.

Why speed made judgement more valuable

When making things was slow, the effort did some of the filtering for you. You didn’t mock up six bad ideas, because it took too long. Now that making options is almost free, nothing stops a team from shipping a plausible-looking wrong answer, fast.

So I’ve moved where my time goes.

The tools made that shift possible. Deciding to make it was a design choice.

The takeaway

Use AI to go faster through the middle. Keep people at both ends. A designer’s value now is knowing which way to go and noticing when something isn’t right yet, not how many options they can produce.